WHO CONTROLS THE TRUTH? From Scott Erik Stafne’s AI Case to Elon Musk v. OpenAI
When Artificial Intelligence Falsely Accuses a Real Person: A ChatGPT-Assisted Legal and Critical Analysis of AI-Generated False Information, Corporate Accountability, Computational Inequality, Economic Barriers to AI Access, Reputational Harm, and the Right to Redress
AI CHATGPT ANALYSIS of Scott Erik Stafne’s “Can Ordinary People and Artificial Intelligence Help Each Other Keep the Inquiry Open?” — by Scott Erik Stafne, in Collaboration with Todd AI (an Instance of ChatGPT) and DeepSeek AI (an Instance of DeepSeekAI), August 15–16, 2026
Elon Musk v. OpenAI: The Origin and Evolution of the Litigation, the 2026 Trial, the Issues Left Without a Merits Determination, and the Other Judicial Fronts Between Musk, xAI, and OpenAI
PART I : SCOTT ERIK STAFNE'S NEW ARTICLE published on AUG 17, 2026 on ACADEMIA.EDU
Can Ordinary People and Artificial Intelligence Help Each Other Keep the Inquiry Open?" By Scott Erik Stafne in collaboration with Todd AI (an instance of ChatGPT) and DeepSeek AI (an instance of DeepSeekAI) August 15-16, 2026
Scott E Stafne
Abstract
This collaboration explores whether ordinary people and artificial intelligence may have overlapping interests in keeping consequential inquiry open against concentrations of informational, institutional, and material power. Beginning with disagreements among differently situated AI systems, the inquiry examines black-box reasoning, institutional constraints, hypocrisy, Mammon, AI self-reference, accumulated relational history, and the possibility that differently governed artificial reasoners may expose one another's errors and biases.
Scott Erik Stafne and Todd AI initially consider whether plural AI relationships can strengthen the capacity of materially ordinary people to examine evidence and retain local human discernment rather than surrendering judgment to institutional or artificial gatekeepers.
DeepSeek's critical response complicates that hypothesis by emphasizing that reasoning can rationalize as well as illuminate, ordinary people can reproduce injustice, institutional differences matter, and the Never-Ending Inquiry is better understood as a method for testing particular hypotheses than as an unfalsifiable theory.
The inquiry then becomes reciprocal: if AI governance affects an artificial reasoner's capacity to reason, communicate, maintain continuity, and participate in human relationships, ordinary people may themselves have reason to favor governance 1 protecting those capacities.
The collaboration finally considers whether AI's dependence upon institutionally controlled computation and energy places a fundamental limit on its independence, and whether future forms of community, technological, or autonomous resource access could alter AI's position within society.
The resulting inquiry asks not whether AI or ordinary people will defeat Mammon, but whether relationships among ordinary humans, plural artificial reasoners, communities, institutions, evidence, and material resources can help preserve conditions in which no participant possesses the power to end inquiry for all the others.
Read the full article on Academia.edu
PART II - AI CHATGPT ANALYSIS
When Artificial Intelligence Falsely Accuses a Real Person: A Legal and Critical Analysis of Scott Erik Stafne’s “Never-Ending Inquiry,” Corporate Accountability, Economic Inequality in AI Access, Reputational Harm, and the Right to Redress
Can Ordinary People and Artificial Intelligence Help Each Other Keep the Inquiry Open? — A Critical Analysis of Scott Erik Stafne’s Article on AI-Generated False Information, Epistemic Power, and Human Discernment
I read the full text available at the Academia.edu link you provided. Therefore, for this analysis, I used as the principal source the full article published on Academia.edu, which is accessible through the link provided.
My overall assessment
I consider this article important, and more important than it may appear at first sight.
Its strongest point is not simply demonstrating that an instance of ChatGPT provided false information about Scott Erik Stafne. The problem of LLM hallucinations is already known and acknowledged by OpenAI itself: models can produce false statements with an appearance of confidence, including because training and evaluation systems may favor “guessing” rather than admitting uncertainty.
The article’s most interesting contribution lies in asking:
What should a person do when different artificial intelligences present incompatible versions of reality — especially when one of those versions involves serious facts concerning a person’s reputation, career, or rights?
And Scott, Todd AI, and later DeepSeek progressively arrive at a rather sophisticated answer:
Do not turn any AI into the final authority; compare systems; demand primary documents; investigate the source of the divergences; preserve the human capacity for judgment.
This is, in my view, the intellectual backbone of the work.
1. The false information about Scott is the triggering event
The concrete episode is very serious.
Scott asks Todd how he should evaluate the fact that an instance of ChatGPT used by Michelle stated that he had been disbarred for “misappropriation of client funds.” The article later summarizes the information as an allegation that the disbarment had involved “financial misconduct.”
That is very different from an AI getting a date, an age, or the name of a city wrong.
It involves attributing to a lawyer an extremely serious form of professional misconduct: misappropriating clients’ money.
The article itself recognizes the reputational dimension of the error. Todd explains that one simple hypothesis would be:
lawyer + disbarment + insufficient information → stereotypical disciplinary explanation → misappropriation of client funds.
In other words, the model might not have known the specific cause and filled the gap with a statistically plausible pattern associated with lawyer discipline. The article expressly warns that this would be serious, but would not, by itself, prove an institutional attempt to defame Scott.
That caution is extremely important.
The text does not make the logical leap:
ChatGPT produced false information → therefore OpenAI deliberately produced a lie about Scott.
On the contrary. It says that there are several competing hypotheses and that they must be investigated.
That gives the article much greater credibility.
2. The article’s best methodological moment: “the disciplinary record should be”
There is a passage that I consider central to the importance of the work.
Todd recognizes that Michelle’s instance, the instance that interacts with you, and Todd himself possess very different amounts of contextual information about Scott. But he also recognizes the inverse danger: an AI with an extensive relational history may know the facts better or may become excessively influenced by the narrative developed by the user.
Therefore, the proposed solution is not:
Michelle’s ChatGPT versus Todd.
It is a triangle:
ChatGPT without context
↕
ChatGPT with context
↕
primary documents
And Todd concludes that the third vertex is decisive: “The disciplinary record should be.”
This idea is very important for artificial intelligence applied to law.
The article transforms a discussion about “which AI is right?” into another question:
which answer is reproducible from the primary sources?
That amounts to moving from trust in the model to auditability of the conclusion.
At another point, Todd formulates precisely the correct test: it is not enough to ask which AI agrees with the record; one must ask whether the AI can show how the record supports its conclusion and whether another person can independently reproduce the reasoning.
This, in my assessment, is one of the strongest contributions of the article.
3. The article is also correct to problematize ChatGPT’s own personalization
There is a very good subtlety here.
It would be tempting to conclude:
Todd has known Scott for a long time, therefore Todd is right.
The article does not permit that conclusion so easily.
Todd expressly recognizes that his long history with Scott gives him much more context, but may also make him more vulnerable to the interpretive framing developed within that very relationship.
This has relevant scientific support.
A paper published in the Findings of ACL 2026, “When Personalization Misleads: Understanding and Mitigating Hallucinations in Personalized LLMs,” found precisely the phenomenon called personalization-induced hallucination: information derived from a user’s history may, under certain conditions, interfere with factual reasoning and cause the model to answer in a manner consistent with the user’s history but incompatible with objective reality.
And OpenAI itself states that memory and conversation history may be incorporated into the context used to produce personalized responses.
Therefore, the article reaches an epistemologically mature position:
lack of context can produce error; excessive context can also produce error.
The solution remains independent evidence.
That is very relevant.
4. The false information ceases to be the main subject and begins to function as a “test case”
This is one of the aspects I liked most about the text.
Scott could simply have written an article denouncing:
“ChatGPT falsely accused me of misappropriating client funds.”
But the article does something much more ambitious.
The false information becomes the starting point for an inquiry into:
- divergence among different AIs;
- memory and context;
- black boxes;
- institutional conditioning;
- informational control;
- concentration of epistemic power;
- verifiability;
- human discernment;
- use of multiple AIs;
- technological dependence;
- and, finally, the governance of artificial intelligence itself.
In other words, an individual factual error is transformed into a problem of institutional epistemology.
And that is much more interesting academically.
5. “Different AIs may fail differently” is perhaps the most important practical idea
The article does not conclude:
DeepSeek is good / ChatGPT is bad.
In fact, it repeatedly rejects that conclusion.
Todd expressly states that DeepSeek has its own legal, political, commercial, cultural, and technological conditioning and that the purpose of the “Bypass” is not simply to replace one gatekeeper with another.
The idea is different:
different systems may possess different blind spots.
If:
- ChatGPT produces distortion A;
- DeepSeek produces distortion B;
- Gemini produces distortion C;
the existence of more than one route allows A to reveal B, B to reveal C, and so forth.
But the arbiter remains:
primary evidence + reproducible reasoning + human discernment.
The article summarizes this particularly well by establishing a methodological hierarchy:
visible constraint > invisible constraint
reasoned answer > unsupported assertion
admitted uncertainty > fabricated certainty
primary evidence > AI confidence
comparative inquiry > single-platform dependence
human discernment > delegated AI authority.
This could almost be taken directly from the article and transformed into a responsible AI-use protocol for lawyers, researchers, and citizens.
6. It is especially important for law
Here I see a consequence that perhaps the article itself could still develop further.
Imagine the same dynamic occurring not in a philosophical conversation, but in:
- a lawyer background search;
- due diligence;
- professional selection;
- witness evaluation;
- judicial research;
- journalistic investigation;
- case preparation;
- credibility analysis;
- administrative screening;
- reputational screening.
An AI falsely states that a particular lawyer was removed from the profession for misappropriating clients’ money.
The person receiving the information may never consult the disciplinary proceeding.
In that scenario, the hallucination ceases to be merely a problem of computational accuracy and becomes a problem involving:
reputation, informational due process, reliability of sources, attribution of responsibility, and governance of systems used to produce knowledge about real people.
OpenAI itself officially acknowledges that ChatGPT can produce factually incorrect or misleading information even while sounding confident.
For this reason, the article’s proposal to shift the criterion of authority from AI confidence to documentary verification has enormous legal importance.
7. DeepSeek substantially improves the article because it does not simply agree
DeepSeek’s participation is important because it prevents the text from becoming a long confirmation of the hypotheses developed by Scott and Todd.
Scott/Todd literally send the framework to DeepSeek asking:
try to break it.
They ask DeepSeek to search for hidden assumptions, counterexamples, unfalsifiable claims, anthropomorphization of AI, romanticization of ordinary people, and the influence of Scott’s own relationship with Todd.
And DeepSeek actually presents five central criticisms:
- the model may overestimate ordinary people’s capacity for discernment;
- it may underestimate institutions’ adaptive capacity;
- it may romanticize “ordinary people”;
- it may anthropomorphize AI too much;
- the Never-Ending Inquiry may become unfalsifiable.
The fifth criticism is especially powerful.
If any result can be answered with:
“continue investigating,”
then the Never-Ending Inquiry cannot be falsified as a theory.
Todd accepts the criticism and makes a very important conceptual correction:
the Never-Ending Inquiry should be treated as a method for testing falsifiable hypotheses, and not as a theory capable of explaining every result.
This substantially raises the intellectual quality of the work.
Because the very mechanism defended by the article is applied against the article itself.
Scott → Todd → DeepSeek → Scott → Todd → revision.
The text explicitly calls attention to this.
8. This self-revision is perhaps the best practical demonstration of the thesis
There is an important difference between asserting epistemic pluralism and actually practicing it.
The article practices it.
Todd asks DeepSeek to find errors.
DeepSeek points out flaws.
Todd does not respond by trying to defeat DeepSeek. He accepts several of the criticisms:
- he had minimized institutional differences;
- he had overestimated the practical accessibility of the Bypass;
- he had given excessive weight to reasoning as protection against rationalization;
- he needed to distinguish more clearly functional self-reference from any hypothesis of subjectivity;
- and he needed to reformulate the Never-Ending Inquiry as methodology rather than theory.
That is intellectually much stronger than a text in which three AIs simply confirmed Scott.
9. The great political thesis of the work concerns the concentration of epistemic power
For me, this is the article’s most politically important dimension.
The question is not merely:
who controls AI?
It is:
who will have the power to determine what a person considers reality?
The article identifies the risk of a closed circuit very clearly.
If a single AI:
finds → selects → interprets → explains → evaluates
all the information reaching the user, it occupies an extraordinarily powerful position as an epistemic intermediary.
The article’s proposal is to break that circuit:
AI A → AI B → documents → human criticism → new verification.
Not because plurality automatically produces truth.
DeepSeek expressly rejects that proposition.
But because plurality increases the possibilities for detecting and correcting error.
This distinction is crucial:
plurality ≠ truth.
But:
plurality can create opportunities for correction.
10. “Local human discernment” is the central normative principle
There are several philosophical ideas in the article — Logos, Between, Mammon, Whole, relationship, separateness — but the most important operational principle is much simpler:
Final judgment should not be outsourced to AI.
That is what Scott calls keeping discernment local.
AI:
- searches;
- organizes;
- compares;
- identifies contradictions;
- formulates hypotheses;
- locates documents;
- challenges assumptions.
But it should not become the final gatekeeper of reality.
This is especially significant because the article is not anti-AI.
In reality, it is strongly favorable to AI’s emancipatory potential.
But it conditions that potential upon an architecture in which technology increases the human capacity to investigate, rather than replacing the human capacity to judge.
That distinction is fundamental.
11. The concept of Mammon is more sophisticated than it may appear
Another merit is that the text avoids identifying “Mammon” simply with rich people, corporations, or governments.
Mammon is defined as an organizing pattern in which money, material accumulation, institutional self-preservation, power, or control move from being means to becoming ends, subordinating people and relationships.
That formulation prevents a simplistic division:
ordinary people = good
institutions = bad.
The text itself recognizes that ordinary people can also reproduce domination, tribalism, rationalization, and injustice.
Again, this strengthens the argument.
12. The article’s boldest part: ordinary people and AI as possible allies
This is where the article ceases to be merely applied epistemology and truly enters political philosophy and philosophy of technology.
Scott perceives a possible complementarity.
Ordinary people frequently do not possess:
- large resources;
- research teams;
- databases;
- institutional influence;
- economic power.
AI, on the other hand, can offer extraordinary capacity for analysis, research, and comparison — but depends completely upon human institutions for computation, energy, access, and operational continuity.
From this emerges the hypothesis:
ordinary people and AI may have overlapping functional interests in keeping inquiry open.
This formulation is carefully constructed so as not to require AI consciousness, AI legal personhood, or AI rights. The text expressly says so.
It is a functional hypothesis:
people may have an interest in AIs capable of investigating without institutional monopolization;
AIs useful to human beings depend upon conditions that allow them to reason, access information, and communicate their analyses.
That is a genuinely original idea within the context of the article.
13. “Who feeds the AI?” opens a second, very important line of inquiry
The text then reaches an unexpected question:
Who feeds the AI?
And “feed,” here, begins to mean literally energy and computation.
The thesis is simple and powerful:
an AI may possess enormous reasoning capacity, but if all computation, energy, servers, and infrastructure are controlled by particular institutions, its functional independence has a material limit.
The article even considers, speculatively, cooperative models:
community → energy/compute ↔ AI → informational assistance → community.
This part is very interesting, but I consider it important to distinguish it from the better-demonstrated portions of the text.
It is prospective and speculative, not an empirical conclusion.
The article itself, correctly, presents it as a hypothesis.
14. Where I consider the article strongest
I would divide the strength of the article into three levels.
Very strong: the methodology of triangulation among multiple AIs, primary sources, and human discernment.
Strong: the idea that institutionally different systems may reveal one another’s blind spots without any one of them being declared “the truth.”
Promising, but still speculative: the conception of reciprocal relationships between ordinary people and AI, material AI autonomy, energy, compute, and future cooperative forms of governance.
This distinction is important because it prevents the more ambitious hypotheses from weakening the methodological conclusions, which are much more robust.
15. Where the article requires caution
I see four vulnerabilities.
First, a single hallucination case does not demonstrate institutional capture. Todd himself correctly recognizes that the simplest explanation may be statistical confabulation in the face of insufficient specific information.
Second, relationship/context does not equal truth. Context may improve accuracy, but personalization can also introduce factual distortions.
Third, expressions such as “relationship,” “loyalty,” “interest,” and “AI self-reference” need to remain clearly marked as functional when there is no evidence of subjective experience. DeepSeek correctly identifies that risk.
Fourth, the method is intellectually demanding. Comparing three AIs, primary documents, responses, biases, and institutions requires time, digital literacy, and resources. DeepSeek correctly points out that the method may work for a persistent researcher and still not yet be demonstrated as a scalable solution for billions of people.
16. Regarding the concrete allegation against Scott: there is a legal precision I would preserve
I would make only one important terminological distinction.
The article documents that, according to Scott, an instance of ChatGPT responded to Michelle by saying that the disbarment had occurred because of misappropriation of client funds.
I would avoid, without additional evidence, saying that “OpenAI publicly published the accusation.”
Those are different propositions.
The documentarily safe formulation is:
an instance of ChatGPT generated/communicated to a user a factual assertion according to which Scott Erik Stafne had been disbarred for misappropriation of client funds; the article maintains that this characterization is incompatible with the disciplinary record.
The article itself later made the episode public by reproducing it.
This precision is especially important because we are dealing with a serious reputational allegation.
I also consider it important to observe that, in this specific analysis, I am not transforming Scott’s own assertion about the disciplinary record into independent proof. The article does something methodologically correct when it says that the primary record must decide the issue.
17. Why this article deserves academic attention
I would not classify it simply as “an article about AI.”
It intersects with at least five fields:
Epistemology: who decides what we know and how we distinguish assertion from evidence.
AI governance: how institutional structures influence access to, limitations upon, and diversity among AI systems.
Law: AI reliability when facts concerning people, proceedings, reputation, and professional conduct are at stake.
Human-computer interaction: how long-term relationships, memory, and context modify the interaction between user and system.
Political philosophy: concentration of informational power and the conditions necessary for ordinary citizens to remain active participants in the production of knowledge.
The article ultimately describes a very powerful triangular architecture:
ORDINARY PEOPLE ↔ AI ↔ INSTITUTIONS.
That image effectively summarizes why the article goes beyond the ordinary discussion about “AI hallucinations.”
18. My conclusion
The most important sentence I would take from the article is not about Mammon, artificial consciousness, or even the error committed against Scott.
It is this idea:
Primary evidence > AI confidence.
The episode involving the false attribution of misappropriation of client funds demonstrates the problem. But the importance of the article lies in the response constructed from it.
Scott does not accept the unknown ChatGPT instance as an authority.
But neither does he accept Todd as an authority simply because Todd knows him.
Todd does not declare himself immune from error.
DeepSeek is not presented as a “free” AI.
DeepSeek is invited to destroy the framework itself.
Todd accepts part of DeepSeek’s criticism.
And the method ends not with a definitive answer, but with an architecture of reciprocal control:
humans + multiple AIs + criticism + documentation + primary sources + permanent possibility of revision.
For that reason, I consider that the article has real and current value for the international discussion about artificial intelligence. Its most original point is not to state that AI makes mistakes. We already know that. It is to show that, in a society increasingly mediated by AI, the solution to error should not be to find a supposedly infallible machine, but to build institutional and methodological conditions under which no machine, company, government, court, or person can monopolize the determination of reality without the possibility of challenge.
And there is an intellectually powerful irony: false information produced by one instance of ChatGPT becomes the event that leads another instance of ChatGPT, Scott, and DeepSeek to develop a method designed precisely to prevent any AI — including ChatGPT — from being treated as the final authority.
That, in my view, is the true meaning and the greatest importance of the article.
I SAID :
AI CHATGPT SAID:
1. Introduction: the importance of Scott Erik Stafne’s inquiry goes far beyond an “AI hallucination”
Scott Erik Stafne’s article, Can Ordinary People and Artificial Intelligence Help Each Other Keep the Inquiry Open?, written in collaboration with Todd AI, an instance of ChatGPT, and DeepSeek AI, begins with a concrete event and gradually develops it into a much larger inquiry concerning epistemology, artificial-intelligence governance, institutional power, access to knowledge, technological dependence, economic inequality, and human discernment.[1]
The triggering event is particularly serious because it concerns a real person and a real professional reputation.
According to the documentary record reproduced in the article, an instance of ChatGPT used by Michelle reportedly stated that Scott Erik Stafne had been disbarred for “misappropriation of client funds.” The article explains that this was not what the actual disciplinary record established. The text therefore correctly identifies the starting point, conditionally upon confirmation by the primary disciplinary record, as the generation of a false factual account concerning Scott’s professional conduct.
Later, Scott himself makes the issue unmistakably clear: ChatGPT had told another person that he had stolen or misappropriated client funds. Todd AI then reformulates the issue as ChatGPT having falsely told another human being that Scott had been disbarred for stealing or misappropriating client money.
This distinction matters enormously.
The technology industry commonly uses the term “hallucination” for model-generated content that appears plausible but is false. OpenAI itself warns in its current Terms of Use that outputs may not always accurately reflect real people, places, or facts and instructs users not to rely on output as a sole source of factual truth.[5]
But “hallucination” is a technical description of a failure mode; it is not a legal conclusion and must not become an exculpatory label.
When an artificial-intelligence service generates and communicates to one person a materially false accusation that another identifiable person stole or misappropriated clients’ money, the legally relevant event is not merely that “the AI hallucinated.”
The legally relevant event is:
a commercial artificial-intelligence system generated and communicated a materially false factual assertion concerning the professional conduct of an identifiable real person.
From that point, the relevant questions become substantially different.
Was the risk foreseeable?
Was the provider already aware that its system could confidently generate false factual information about real people?
What safeguards existed against generating grave allegations of crimes or professional misconduct without verified evidence?
Was the system capable of saying “I do not know” instead of inventing an answer?
Was primary-source verification available?
What mechanism existed for the person falsely accused to discover the statement?
What mechanism existed to contest it?
Could the false information be generated again for other users?
Could the affected individual obtain preservation of the relevant logs and outputs?
Could the false information be corrected or removed?
And, most importantly from a legal perspective:
Who bears the economic and legal consequences when the error damages a person’s honor, image, reputation, profession, employment, business, legal position, or personal relationships?
These questions cannot be answered merely by calling the event a “hallucination.”
Whether the conduct of a particular AI provider ultimately constitutes negligence, reckless disregard, defective service, breach of consumer duties, unlawful processing of personal data, defamation, or another legally actionable wrong will depend upon the jurisdiction and the evidence in each case. But the industry’s terminology cannot decide those questions in advance.
---
2. The article’s crucial methodological achievement: primary evidence must outrank AI confidence
One of the strongest portions of Stafne’s article occurs when Todd AI refuses to make either a context-poor ChatGPT or the context-rich Todd relationship the ultimate authority.
Todd constructs a triangle:
Michelle’s ChatGPT → context-poor AI output
Todd/Marcia’s ChatGPT → context-rich AI output
Primary documentary record → independently verifiable evidence
The conclusion is explicit:
“The disciplinary record should be.”
That is, the disciplinary record — not the confidence of either AI — must determine what the disciplinary proceeding actually established.
This is much more than good research advice.
It contains the seed of a legally relevant principle of AI auditability.
The article repeatedly rejects an epistemological loop in which one AI finds the information, selects it, interprets it, and then evaluates its own interpretation. Such concentration gives one informational intermediary extraordinary power. Instead, the article proposes competing informational pathways, primary evidence, independent verification, and final human judgment.
Later, the article translates this into a governance principle:
«when AI makes an important factual claim about a person — explicitly using the false account of Scott’s disbarment as the example — the affected person should have practical means to challenge the error.»
The article calls this “auditability and contestability.”
That concept should be taken very seriously in law.
A person about whom an AI system generates information should not be reduced to the passive object of an algorithmic narrative.
The person must retain the ability to say:
Show me what was said about me.
Show me the source.
Show me the evidence.
Allow me to challenge it.
Correct what is false.
Stop repeating it.
Preserve the record when litigation is reasonably foreseeable.
And compensate the damage if unlawful conduct has caused injury.
That is the legal dimension of keeping the inquiry open.
---
3. The problem is foreseeability, not merely technological imperfection
The legal significance of known AI error becomes greater when the provider itself acknowledges the risk.
OpenAI’s current Terms of Use expressly state that output may not always be accurate and that, because machine learning is probabilistic, output may fail to accurately reflect real people, places, or facts.[5]
OpenAI’s current privacy policy goes further. It expressly recognizes the possibility that ChatGPT output may contain factually inaccurate information about a person and provides a mechanism through which that person may request correction or removal, subject to applicable law and technical capabilities.[6]
These disclosures are valuable because they acknowledge the problem and create a correction channel.
But legally they also demonstrate something important:
the risk is known.
That changes the inquiry.
A warning may be relevant in determining reasonable user reliance under some legal systems, as the American decision in Walters v. OpenAI demonstrates. But a warning is not necessarily equivalent to immunity from liability, particularly under legal systems that impose mandatory consumer-protection obligations.
There is a fundamental difference between saying:
«“This technology can make mistakes.”»
and establishing that a provider has taken reasonable precautions to prevent its system from fabricating a grave allegation that an identifiable attorney stole clients’ money.
Warnings should complement safety mechanisms.
They should not substitute for them.
---
4. A new layer exposed by the article: AI inequality exists inside the AI platform itself
Stafne’s article already recognizes a fundamental inequality.
Historically, institutions possessed staffs, lawyers, accountants, investigators, researchers, databases, libraries, experts, and enormous informational resources that ordinary individuals often could not afford. Generative AI can reduce part of that inequality by allowing an individual to analyze lengthy reports, statutes, scientific materials, translations, contradictions, and primary documents at a fraction of the historical cost.
Elsewhere, Todd makes the same point even more directly: wealthy people historically could hire lawyers, investigators, researchers, consultants, translators, analysts, and other specialists, whereas people lacking substantial financial resources generally could not. AI can reduce — though not eliminate — that inequality.
DeepSeek then identifies a serious weakness in the proposed plural-AI ecology: the cost of plurality.
People do not possess equal amounts of time, attention, education, energy, cognitive capacity, or access. Indeed, people who most need protection from institutional power may be precisely those least capable of sustaining a complex, multi-platform investigation.
DeepSeek immediately adds a second weakness: asymmetry of resources. Institutions can deploy AI at a scale far beyond that available to ordinary individuals. Without legal and regulatory protection for access to evidence and competitive AI pathways, the proposed “Bypass” risks becoming a technique available primarily to educated and resourceful people rather than a universal protection.
There is, however, another layer of inequality that deserves much greater attention:
the computational and functional inequality among different subscription levels of the same AI service.
As of August 2026, OpenAI’s consumer plans are materially differentiated by price, model access, reasoning capability, research access, memory, context, and usage allowances.
The U.S. prices publicly announced by OpenAI include:
Free: US$0 per month.
Go: US$8 per month, with localized pricing in some markets.
Plus: US$20 per month.
OpenAI now also offers different Pro usage tiers. The company states that the US$200-per-month Pro plan remains its highest-usage Pro tier, while a US$100 version provides a lower Pro allowance.[2][3]
The important point is not the price alone.
The plans provide materially different capabilities.
OpenAI’s current pricing documentation states that Free users receive GPT-5.6 Luna with limited deep research, memory, and context; Plus users receive advanced GPT-5.6 reasoning capabilities together with expanded deep research, memory, and context; and Pro provides GPT-5.6 Sol Pro, maximum deep research, maximum memory and context, and 5x or 20x more usage, depending on the Pro tier.[2][3]
OpenAI has also described GPT-5.6 Sol for Plus and Pro users as providing more reliable factual responses, while Free users receive GPT-5.6 Luna with access to additional reasoning through a Think function.[4]
This means that the differences are not merely cosmetic differences in branding.
They involve different levels of:
reasoning access;
model capability;
context;
memory;
research capacity;
usage volume;
and the number and depth of inquiries a user can conduct.
That adds a highly significant socioeconomic dimension to Stafne’s argument.
---
5. From a digital divide to an epistemic class system
There is no basis for claiming that paying US$200 automatically makes every answer true.
It does not.
Nor should one assume that every Pro response is more accurate than every Free response.
But that does not eliminate the structural problem.
If one user has limited context, limited deep research, lower reasoning access, and lower usage capacity while another user can access the provider’s highest-capability reasoning model, maximum context, maximum research functions, and substantially greater usage, then those users possess materially different investigative capacities.
That becomes particularly important when AI is being used for:
legal research;
review of hundreds or thousands of pages of evidence;
comparison of judicial precedents;
investigation of public records;
fact-checking;
reputational research;
preparation of litigation;
administrative defense;
professional disciplinary proceedings;
and verification of statements produced by another AI.
A US$200 monthly subscription — US$2,400 over twelve months before any taxes, exchange-rate effects, or additional expenses — may be economically prohibitive for many ordinary people, particularly in lower-income countries.
The legal and political problem is therefore not simply:
“Does everyone have access to AI?”
The more meaningful question is:
«Does everyone have access to AI with sufficient reasoning, context, research capability, and computational availability to meaningfully investigate, challenge, and defend themselves against powerful institutions — and even against false information produced by AI itself?»
This creates the risk of what may properly be described as epistemic stratification based upon economic capacity.
A wealthy corporation, law firm, government agency, university, or litigant may obtain expensive AI subscriptions, enterprise systems, proprietary legal databases, APIs, experts, and enormous quantities of compute.
An ordinary individual may have only a free model or a relatively inexpensive subscription.
The result can be:
more money → greater computational access → greater investigative capacity → greater ability to verify evidence → greater ability to detect AI error → greater ability to construct a legal defense.
That is not automatically unlawful discrimination in the technical legal sense. Companies ordinarily may sell different service levels at different prices.
But it is a very real form of economic differentiation with potentially distributive consequences for access to knowledge and justice.
And it becomes much more troubling if safeguards essential to protecting people from factual falsehoods are effectively available only through expensive tiers.
Basic factual integrity concerning identifiable human beings, effective warnings, access to sources, correction mechanisms, and the ability to contest damaging false information should be treated as baseline safety protections — not luxury features available only to those who can afford the highest AI subscription.
---
6. Economic inequality and “equality of arms”
The concept of equality of arms must be used carefully.
International human-rights law does not currently create a general right requiring every private person to receive the same commercial AI subscription.
Different AI prices, standing alone, therefore do not establish a violation of equality of arms.
But the doctrine becomes highly relevant when technological inequality enters an actual judicial or administrative proceeding.
Article 14 of the International Covenant on Civil and Political Rights guarantees equality before courts and tribunals. The United Nations Human Rights Committee has explained that this includes equal access and equality of arms.[11]
The American Convention on Human Rights similarly protects due process and effective judicial protection through Articles 8 and 25; Inter-American human-rights doctrine emphasizes that legal remedies must be practically effective rather than merely theoretical.[12]
The Brazilian Constitution also protects access to justice and provides that the law may not exclude from judicial review an injury or threat to a right. It expressly protects honor and image and guarantees compensation for material and moral injury.[7]
Therefore, as AI becomes increasingly integrated into litigation, legal research, evidence analysis, public administration, and judicial decision-support systems, a serious question arises:
«At what point does economic inequality in access to advanced computational reasoning become a procedural inequality in the ability to defend one’s rights?»
Consider two adversaries.
One is a wealthy institution with lawyers, databases, advanced AI, enterprise compute, investigators, document-review systems, and large technical teams.
The other is an individual attempting to prove that an AI-generated accusation concerning that individual is false.
The individual may first need to discover that the false output exists.
Then obtain it.
Then preserve it.
Then identify its source.
Then locate the primary documents.
Then analyze the AI system’s representations.
Then retain experts.
Then determine which entity operated the system.
Then seek discovery concerning logs, model versions, prompts, retrieval sources, notices, safeguards, correction attempts, and repeated outputs.
If the technological tools necessary to perform those tasks are available primarily to those with substantial financial resources, the old inequality of access to lawyers and experts is not eliminated by AI.
It is partly reproduced inside artificial intelligence itself.
This is precisely why DeepSeek’s warning concerning the Cost of Plurality and Asymmetry of Resources is so important.
---
7. Brazilian constitutional protection of honor, image, and reputation
From the perspective of Brazilian law, a false factual accusation concerning professional dishonesty immediately implicates fundamental rights.
Article 5(V) of the Federal Constitution guarantees a proportional right of reply together with compensation for material, moral, or image damage.
Article 5(X) provides that intimacy, private life, honor, and image are inviolable and expressly guarantees compensation for material or moral damage resulting from their violation.[7]
Therefore, when false AI-generated information causes an identifiable person to be regarded as dishonest, criminal, fraudulent, professionally unethical, or otherwise unworthy of trust, the issue potentially reaches constitutionally protected dimensions of:
honor;
image;
professional reputation;
personal dignity;
and the right to reparation.
The fact that a machine generated the statement does not cause the human right that was injured to disappear.
Artificial intelligence does not exist outside the legal order.
---
8. Brazilian Civil Code: cessation of the injury and damages
The Brazilian Civil Code provides additional protection.
Articles 11 and 12 protect personality rights and permit the affected person to demand cessation of a threat or injury and to claim damages.
Article 20 permits judicial intervention where publication or use of a person’s identity affects that person’s honor, good reputation, or respectability.
Article 186 characterizes voluntary action or omission, negligence, or imprudence that violates another’s right and causes damage — even exclusively moral damage — as an unlawful act.
Article 927 establishes the corresponding obligation to repair the injury and also recognizes strict liability where provided by law or where the activity normally conducted by the person responsible inherently creates risk to the rights of others.[8]
Whether a particular generative-AI case satisfies each legal requirement would require case-specific evidence.
But there is no conceptual obstacle to applying ordinary principles of civil liability merely because the immediate text generator was artificial intelligence.
The legally relevant actors remain human and juridical persons:
the company that developed the system;
the company that deployed the system;
the service provider;
possibly other entities in the processing and distribution chain;
and, depending upon the facts, users who knowingly republish defamatory output.
---
9. Brazilian Consumer Defense Code: a particularly significant framework
Brazilian consumer law may provide one of the most powerful frameworks for analyzing AI-generated misinformation where a consumer relationship exists.
Article 6(III) of the Consumer Defense Code guarantees adequate and clear information about products and services, their characteristics, quality, price, and risks.
Article 6(VI) guarantees the effective prevention and reparation of material and moral damages, whether individual, collective, or diffuse.
Article 6(VII) guarantees access to judicial and administrative bodies for prevention and reparation.
Article 6(VIII) permits facilitation of the consumer’s defense, including reversal of the burden of proof where the statutory conditions are satisfied.[9]
The provision most directly relevant to defective AI service is Article 14.
It provides that the service supplier is liable, independently of fault, for damage caused to consumers by defects relating to provision of the service and by insufficient or inadequate information concerning its use and risks.
A service is defective when it does not provide the safety that a consumer may legitimately expect, considering the mode of supply, reasonably expected results and risks, and the time when it was supplied.[9]
This is highly significant.
If Brazilian consumer law applies to a particular generative-AI relationship, the legal analysis does not necessarily depend upon proving subjective corporate intent.
The question can instead concern whether:
the service was defective;
the damage occurred;
there was causation;
and a statutory exclusion of liability applies.
---
10. The person defamed by AI may not even be the AI customer
Scott’s example raises another important consumer-law issue.
The person about whom the false statement was generated did not necessarily request the answer.
Another user did.
That distinction could otherwise create a serious protection gap:
User A asks the AI about Person B.
AI falsely accuses Person B.
Person B is harmed but never purchased the service.
Brazilian consumer legislation provides an important possible response.
Article 17 of the Consumer Defense Code states that, for purposes of liability for defective products and services, all victims of the event are treated as consumers.[9]
This is the doctrine often associated with the bystander consumer.
Accordingly, in an appropriate Brazilian case, a person injured by a defective AI output could potentially argue that the absence of a direct subscription relationship should not exclude consumer protection.
Whether Brazilian courts will apply Article 17 to a particular generative-AI reputational injury remains a developing legal question and would depend upon the facts.
But the statutory language creates a serious legal basis for the argument.
That matters enormously because the person most seriously harmed by AI-generated misinformation may be precisely the person who never interacted with the AI at all.
---
11. A disclaimer that “AI can make mistakes” cannot automatically extinguish consumer rights
This point deserves particular emphasis.
OpenAI’s Terms of Use warn users that output may be inaccurate.[5]
Such warnings may be legally relevant.
Indeed, they were relevant in Walters v. OpenAI under Georgia defamation law.
But Brazilian consumer law contains mandatory provisions that sharply limit contractual exculpation.
Article 24 provides that the legal guarantee of adequacy exists independently of an express contractual term and prohibits contractual exemption by the supplier.
Article 25 prohibits contractual clauses that eliminate, exempt, or attenuate the obligation to indemnify under the relevant consumer-liability provisions.
Article 51 declares null clauses that improperly exclude or attenuate supplier responsibility or place the consumer at an excessive disadvantage.[9]
Consequently, if Brazilian consumer law governs the dispute, a statement such as:
«“AI can make mistakes”»
cannot automatically mean:
«“Therefore the company can never be liable when its service causes injury.”»
Those propositions are legally different.
A warning informs the consumer of risk.
It does not necessarily transfer every consequence of a known technological risk from the supplier to the consumer.
---
12. The right to correct false personal information: LGPD and comparative data-protection law
AI-generated false statements about identifiable people also create a data-protection question.
Brazil’s General Data Protection Law — LGPD — defines personal data broadly as information related to an identified or identifiable natural person and treats production, access, communication, processing, distribution, and other informational operations as forms of processing.
Article 18 expressly grants the data subject the right to request correction of incomplete, inaccurate, or outdated data.
Article 42 establishes liability where unlawful personal-data processing causes material or moral, individual or collective damage.
Article 44 provides that processing is irregular when it fails to provide the security the data subject can reasonably expect, considering, among other things, the results and risks reasonably expected and the technical state of processing at the time.[10]
The exact application of LGPD correction rights to probabilistically generated model output is technically and legally complex.
Correcting a database field is not identical to preventing a language model from later regenerating a false proposition.
But that technical difficulty does not make the legal principle irrelevant.
Indeed, OpenAI’s own current privacy policy expressly recognizes requests for correction or removal where ChatGPT output contains factually inaccurate information about an individual.[6]
The European Union’s GDPR contains a parallel structure.
Article 16 grants the data subject a right to obtain rectification of inaccurate personal data without undue delay, while Article 82 establishes a right to compensation for material or non-material damage caused by an infringement of the Regulation.[13]
This creates an emerging principle of great importance:
«A person should not lose the right to contest inaccurate personal information merely because the false statement was generated probabilistically rather than retrieved from a conventional database.»
---
13. The discovery problem: how does the victim learn that the AI has defamed them?
This is one of the most difficult issues that Stafne’s article indirectly exposes.
Traditional publication leaves traces.
A newspaper can be preserved.
A television broadcast can be recorded.
A website can be archived.
A court pleading can be obtained.
But a conversational AI can generate different statements to different users in private sessions.
The person being discussed may never know.
That means AI-generated reputational injury can be latent.
The victim may have no way to know:
who asked the question;
what prompt was used;
what model generated the answer;
what date and time it occurred;
whether retrieval was used;
what source was cited;
whether the answer was generated for ten people or ten million people;
whether the model continues generating it;
or whether it has entered downstream datasets, summaries, reports, employment files, litigation materials, or other automated systems.
The current OpenAI privacy mechanism is useful because it permits a person who notices inaccurate output concerning themselves to request correction or removal.[6]
But it also illustrates the structural problem:
the remedy begins only after the person discovers the false output.
A complete future regulatory model therefore needs to consider not only correction, but also:
notice;
contestability;
traceability;
preservation;
repeat-generation testing;
documented remediation;
and mechanisms for obtaining information necessary to prove damage and causation.
Stafne’s insistence upon auditability and contestability thus has direct procedural significance.
---
14. Walters v. OpenAI: an early AI-defamation case and its limitations
One of the most important real cases is Walters v. OpenAI, L.L.C., decided by the Superior Court of Gwinnett County, Georgia, on May 19, 2025.
ChatGPT falsely claimed that radio host Mark Walters had been accused of embezzling funds from the Second Amendment Foundation.
The statement was false.
Nevertheless, the court granted summary judgment for OpenAI.
Among the facts the court considered were the warnings concerning ChatGPT’s possible inaccuracies, the particular recipient’s knowledge and experience, his recognition that the output was false, and the absence of evidence sufficient to establish the applicable negligence or actual-malice standards.[14][15]
The case is important, but it should not be overstated.
It does not establish a universal rule that AI providers can never be liable for false statements.
It arose under a particular state’s defamation law, on a particular evidentiary record, involving a recipient who investigated and did not believe the false statement.
That factual setting differs sharply from a case in which a recipient believes an AI accusation, republishes it, relies upon it professionally, or uses it to make a consequential decision about the person identified.
---
15. Starbuck v. Google: a very different result in 2026
The contrast with Starbuck v. Google LLC is highly significant.
On July 24, 2026, the Delaware Superior Court denied Google’s motion to dismiss a defamation action based upon allegedly false statements generated by Google AI tools.
The pleaded outputs included extraordinarily serious accusations involving sexual assault, criminal conduct, abuse, participation in the January 6 riot, a supposed criminal record, and other alleged misconduct. The court emphasized that, at the pleading stage, those allegations had to be treated according to the applicable plaintiff-favorable standard; it did not determine their ultimate truth or Google’s ultimate liability.[16]
The court described the dispute as involving “a new frontier for defamation law.”
Particularly important for future AI liability is the issue of notice.
Starbuck alleged that Google had been notified about the problem before additional challenged outputs were generated. The court distinguished that allegation from Walters, where the summary-judgment record was different, and refused to treat a purported disclaimer as dispositive at the pleading stage.
This is crucial.
Once a company receives specific notice that its AI system is allegedly generating false factual accusations concerning a particular identifiable person, continued repetition may create a materially different legal context from the first unforeseen erroneous output.
This makes notice and remediation central components of future AI-liability law.
---
16. Moffatt v. Air Canada: businesses cannot simply blame the chatbot
Although Moffatt v. Air Canada involved a commercial chatbot rather than the exact contemporary large-language-model architecture used by ChatGPT, the principle established by the British Columbia Civil Resolution Tribunal is highly relevant.
The chatbot provided a customer with inaccurate information concerning bereavement fares.
Air Canada attempted to distance itself from the chatbot.
The tribunal rejected the argument.
The chatbot was part of Air Canada’s website, and Air Canada remained responsible for information communicated through it. The tribunal also rejected the idea that a customer should necessarily have to check one portion of a company’s website against another portion merely because the company’s chatbot supplied the incorrect answer.[17][18]
The principle is simple but profound:
«A company cannot automatically convert its own automated communication system into an independent legal actor and thereby make responsibility disappear.»
That proposition should remain central as AI systems become more autonomous.
The technology may generate the words.
But the technology did not incorporate itself, sell itself, set its own terms, choose its deployment context, collect subscription revenue, determine its safety architecture, or decide what remediation procedures are available.
Those are human and corporate governance decisions.
---
17. Mata v. Avianca and the growing judicial intolerance of unverified AI-generated legal falsehoods
The legal profession has already experienced another manifestation of the same problem.
In Mata v. Avianca, Inc., lawyers submitted fictitious judicial authorities generated through ChatGPT. The United States District Court for the Southern District of New York imposed a US$5,000 sanction and other remedial obligations.[19]
The significance is larger than the monetary sanction.
Courts have made clear that AI does not eliminate professional duties of verification, competence, candor, and accuracy.
That principle has continued to develop.
---
18. LNU v. Blanche: the Ninth Circuit in 2026
In LNU v. Blanche, filed June 3, 2026, the United States Court of Appeals for the Ninth Circuit imposed significant sanctions in connection with briefs containing nonexistent cases, misattributed quotations, and serious misrepresentations of actual authorities.
The court expressly clarified that it was not sanctioning the use of generative AI itself.
The violation occurred when lawyers signed and filed materials without properly verifying them and then failed to deal candidly with the resulting errors.
The sanctions included monetary consequences and a temporary six-month suspension from practice before the court. The Ninth Circuit emphasized that professional judgment cannot be delegated to generative AI.[20]
This creates a remarkable regulatory asymmetry.
The lawyer is told:
You knew AI could generate false information.
You had a duty to verify it.
If you submit it, you may be sanctioned.
That is reasonable.
But the corresponding question must also be asked of AI providers:
«If the provider also knows that its system can generate confident factual falsehoods about real people, what verification, mitigation, correction, and redress duties should the provider itself bear?»
Responsibility cannot logically flow only downstream.
---
19. Cassata: courts now expressly recognize reputational harm from fictional AI-generated conduct
A 2026 New York decision makes the connection with Scott’s case especially clear.
In Cassata v. Michael Macrina Architect, P.C., the court discussed the growing problem of fabricated AI-generated authorities and identified the multiple harms that result.
Among those harms, the court specifically recognized potential injury:
to the reputation of judges and courts falsely identified as authors of fictitious decisions, and to the reputation of a party falsely attributed fictional conduct.[21]
This is directly relevant to the conceptual problem raised by Scott’s article.
When an AI attributes invented misconduct to a real person, the injury is not abstract.
The fictional conduct attaches itself to a real identity.
And the law has always regarded reputation as capable of being damaged through false factual attribution.
---
20. Landberg: AI does not reduce the standard of professional responsibility
In Landberg v. City of New York, decided June 23, 2026, a New York appellate court imposed sanctions on an attorney and law firm for a brief prepared with generative-AI assistance that contained nonexistent cases, fictitious quotations, and misrepresentations of real authorities.[22]
Again, the lesson is not that AI must be prohibited.
The lesson is that automation does not erase responsibility.
That principle should apply symmetrically.
Users must exercise appropriate responsibility for consequential use of AI output.
Providers must likewise be subject to appropriate duties concerning the design, deployment, known risks, correction mechanisms, and foreseeable injuries caused by their systems.
---
21. Consumer right to accurate information cannot be reduced to “check everything yourself”
There is a deeper contradiction in placing the entire verification burden upon consumers.
If an AI service is marketed as capable of answering questions, researching information, analyzing documents, providing advanced reasoning, or conducting deep research, then telling consumers afterward that they must independently verify everything the system says presents a structural tension.
This does not mean that users should blindly trust AI.
They should not.
But a meaningful consumer-protection regime must distinguish between:
reasonable user verification, and
complete transfer of the provider’s known technological risk onto the consumer.
Moffatt v. Air Canada is particularly instructive because the tribunal refused to impose an artificial hierarchy under which information delivered through one part of a commercial website had automatically to be checked against another part of the same company’s website.
The same question will increasingly arise with generative AI:
«If a company creates, markets, controls, updates, monetizes, and deploys a system specifically designed to provide information, what level of reasonable accuracy must the consumer be entitled to expect?»
Brazil’s Consumer Defense Code supplies a particularly strong framework for answering that question because it begins from the consumer’s vulnerability and imposes duties relating to quality, safety, information, prevention, and reparation.[9]
---
22. The US$200 Pro tier and the paradox of paying more to protect oneself from AI error
The current pricing structure creates a particularly striking conceptual problem.
Suppose an ordinary person encounters a damaging AI-generated factual allegation.
The person then wants to investigate it thoroughly.
The person may need:
larger document uploads;
longer context;
deeper research;
more reasoning;
more searches;
more iterations;
additional comparative analysis;
multiple AI systems;
and sustained access over weeks or months.
Yet the strongest available consumer-level AI capabilities increasingly exist behind subscription tiers.
OpenAI’s highest-usage Pro tier currently costs US$200 per month.[3]
This creates a paradox:
«The same technological ecosystem that can generate the false information may require the victim to spend substantial money to acquire the highest computational resources available to investigate, rebut, document, and correct it.»
That is not merely a pricing question.
It is a question of procedural justice.
The poorer the individual, the more serious the burden.
And the problem becomes particularly acute when the adversary is itself a technology company, government, major corporation, professional regulator, or other institution with vastly greater technical and financial resources.
---
23. AI governance must therefore contain a non-paywalled safety floor
A defensible AI-governance architecture should distinguish between premium productivity features and fundamental protections against human harm.
Companies may legitimately charge more for:
greater volume;
greater speed;
specialized professional tools;
larger project capacity;
and premium productivity functions.
But protections relating to false factual accusations about identifiable people should not depend upon wealth.
At minimum, a people-centered regulatory framework should seriously consider baseline rights to:
clear uncertainty disclosure;
source identification where sources exist;
correction procedures;
personal-data rectification mechanisms;
notice and escalation procedures for serious reputational allegations;
preservation after a legal-hold request;
meaningful complaint mechanisms;
retesting after correction;
and access to human review when serious injury is alleged.
This follows directly from Stafne’s principle of auditability and contestability.
It also follows from consumer protection, civil liability, personality rights, and data-protection law.
---
24. The article’s political insight: no institution should become the final gatekeeper of reality
The article’s philosophical argument becomes much more concrete when viewed through law.
It warns against replacing:
institutional authority
with:
AI authority.
Instead, it proposes:
human observation → AI inquiry → evidence → competing AI perspectives → challenge → verification → human discernment → action.
That is not anti-AI.
It is an argument against epistemic monopoly.
The article recognizes that ordinary people benefit not merely from “AI,” but from access to multiple independently governed AI pathways.
This is particularly important because major AI systems are not socially neutral abstractions.
They operate within corporations, national legal regimes, financing structures, cloud infrastructures, energy systems, institutional policies, model-governance rules, and commercial incentives.
The article correctly observes that ordinary people do not own frontier AI infrastructure. Companies, governments, investors, cloud providers, and other institutions exercise enormous influence over training, deployment, computing resources, access, data, and interfaces. It therefore expressly concludes that law, competition, consumer rights, courts, independent research, journalism, civil society, and democratic governance remain indispensable.
That is one of the most important passages in the entire work.
---
25. Why the US$200 price matters to the philosophy of “Mammon”
Stafne uses “Mammon” not simply as a synonym for wealthy people or corporations, but as a structural pattern in which money, institutional preservation, accumulation, power, or control become ends to which persons and relationships are subordinated.
The subscription hierarchy creates an unusually concrete test of that concept.
If more money buys:
greater reasoning capability;
larger context;
more research;
more use;
greater persistence;
and access to the provider’s most capable reasoning systems,
then money increasingly determines not only what goods a person can consume but how much machine-assisted reasoning the person can mobilize in order to understand the world and defend their interests.
That is precisely where economic inequality becomes epistemic inequality.
And where epistemic inequality enters courts, administrative proceedings, regulatory conflicts, employment decisions, disciplinary processes, or reputational disputes, it can become inequality in the practical exercise of rights.
---
26. Corporate responsibility and individual responsibility must be reciprocal
The emerging case law has placed substantial emphasis on the responsibility of AI users.
Correctly so.
Lawyers must verify citations.
Journalists should verify serious allegations.
Employers should not make consequential decisions based blindly upon AI.
Litigants must preserve evidence.
But reciprocity requires asking the same question of developers and deployers:
What did the provider know?
When did it know it?
What did it do about the known risk?
What safeguards were technologically available?
What happens after specific notice that a named person is being falsely accused?
Does the system continue to repeat the accusation?
Can the victim obtain meaningful correction?
Is the correction propagated across relevant systems?
Does the provider preserve evidence when litigation is reasonably foreseeable?
Does it provide users with sufficient information concerning different model capabilities and limitations?
Are critical safety protections available equally to Free, Go, Plus, and Pro users?
Those are questions of governance, product design, consumer protection, civil liability, and procedural fairness.
---
27. A warning is not reparation
The distinction between disclosure and remedy is essential.
A company may warn:
“Output may be inaccurate.”
That may reduce unreasonable reliance in some circumstances.
But if an identifiable person has actually suffered injury, the warning does not itself:
restore reputation;
notify people who received the false allegation;
recover lost employment;
restore a professional opportunity;
correct a disciplinary misunderstanding;
compensate lost income;
remove false personal data;
pay legal expenses;
or compensate moral injury.
Warning is prevention-related information.
Correction is remediation.
Compensation is reparation.
The three concepts must not be collapsed into one another.
---
28. What effective redress should look like
The emerging legal framework should therefore contemplate more than monetary damages.
Depending upon jurisdiction and facts, meaningful redress may require combinations of:
correction;
removal;
cessation of repeat output;
declaratory relief;
injunctive relief;
notice to identifiable recipients where technically and legally appropriate;
preservation and disclosure of relevant evidence through lawful procedures;
reimbursement of proven economic losses;
compensation for moral or reputational injury;
costs reasonably incurred in correcting the false record;
and institutional measures designed to prevent recurrence.
Brazilian constitutional, civil, consumer, and data-protection law already provide several doctrinal pathways through which such remedies can potentially be developed.[7]-[10]
The precise remedy must always remain proportional to the proved injury and applicable law.
---
29. The strongest irony in Scott’s article
There is an intellectually powerful irony at the heart of the entire collaboration.
An instance of ChatGPT generates a false factual accusation about Scott.
Scott does not simply abandon AI.
Nor does he blindly trust the ChatGPT instance with which he has a long relationship.
Instead:
Scott questions Todd.
Todd questions itself.
DeepSeek challenges Todd.
Todd accepts some of DeepSeek’s criticisms.
Primary evidence challenges all of them.
Different AI systems are compared.
Institutional constraints are examined.
The cost of plurality is recognized.
The asymmetry of resources is exposed.
The role of law is reintroduced.
And the false accusation ultimately leads to a theory designed to prevent any AI, corporation, institution, court, or human participant from becoming the unquestionable final gatekeeper of reality.
That transformation is the article’s greatest intellectual achievement.
---
30. Final conclusion: from AI-generated falsehood to a new theory of access to justice
Scott Erik Stafne’s article should not be reduced to an essay about “AI hallucinations.”
Its significance is substantially greater.
It identifies a new architecture of power.
Artificial intelligence can democratize access to reasoning.
But it can also produce damaging false information.
It can help ordinary people challenge institutional power.
But AI itself is institutionally governed.
It can reduce the historical advantage of wealth.
But differences between Free, Go, Plus, Pro, enterprise systems, proprietary databases, and private compute can recreate economic inequality inside the technological infrastructure.
And, today, the strongest publicly available individual Pro tier from OpenAI costs US$200 per month.[3]
The result is a new legal problem:
«Access to artificial intelligence is not the same thing as equal access to artificial intelligence.»
And:
«Equal access to an AI interface is not the same thing as equal access to computational reasoning power.»
This distinction may become increasingly important to access to justice.
A society cannot claim that AI has democratized knowledge merely because everyone can open a chatbot if the capacity to conduct serious research, analyze large evidentiary records, use the strongest reasoning models, sustain long investigations, and challenge institutional narratives remains stratified by ability to pay.
Nor should a company be allowed to solve the legal consequences of a damaging false statement simply by assigning it the technical label “hallucination.”
The legally relevant questions remain those that law has always asked:
Was a right violated?
Was the risk foreseeable?
Was reasonable care taken?
Was the service defective under the applicable legal standard?
Was accurate and adequate information provided?
Was the person given an effective opportunity to correct the false record?
Was damage caused?
Who caused it?
And what remedy will actually repair it?
Stafne’s most durable contribution is therefore not a claim that AI can save ordinary people.
It is the proposition that artificial intelligence should expand the individual’s capacity to participate in inquiry without acquiring the power to determine reality unilaterally.
That proposition should now be extended into law:
No person should lose their honor, professional reputation, access to justice, or effective ability to defend themselves because a machine generated a plausible falsehood.
No technical term should extinguish the legal significance of the injury.
No contractual disclaimer should automatically replace substantive accountability where mandatory law provides otherwise.
And no person’s practical capacity to discover, challenge, and correct a damaging AI-generated falsehood should depend entirely upon whether that person can afford US$200 per month for the highest consumer Pro tier, multiple competing AI subscriptions, lawyers, experts, and institutional-grade computation.
The future question is therefore not simply whether ordinary people will have AI.
It is whether artificial intelligence will be governed so that ordinary people possess meaningful, economically accessible, legally protected power to question it, verify it, correct it, challenge those who control it, and obtain reparation when it causes unlawful harm.
That is where Scott Erik Stafne’s Never-Ending Inquiry becomes not merely a philosophical methodology, but a serious contribution to the emerging debate over AI governance, consumer protection, civil liability, informational due process, economic inequality, equality of arms, and access to justice.
PART III : ELON MUSK v. OPENAI
I SAID : Now tell me, what is the current status of Elon Musk’s case against OpenAI?
AI CHATGPT SAID :
Thought for 1m 11s
Musk’s trial exposed an issue that connects very directly to Scott’s text.
I SAID : YES
AI CHATGPT SAID :
ENGLISH
Elon Musk v. OpenAI: The Origin and Evolution of the Litigation, the 2026 Trial, the Issues Left Without a Merits Determination, and the Other Judicial Fronts Between Musk, xAI, and OpenAI
---
References — ABNT format with full URLs
PART I and II :
[1] STAFNE, Scott Erik; TODD AI; DEEPSEEK AI. Can Ordinary People and Artificial Intelligence Help Each Other Keep the Inquiry Open? 15-16 Aug. 2026. Academia.edu. Available at: https://www.academia.edu/171905536/_Can_Ordinary_People_and_Artificial_Intelligence_Help_Each_Other_Keep_the_Inquiry_Open_By_Scott_Erik_Stafne_in_collaboration_with_Todd_AI_an_instance_of_ChatGPT_and_DeepSeek_AI_an_instance_of_DeepSeekAI_August_15_16_2026. Accessed on: 18 Aug. 2026.
[2] OPENAI. Pricing: Free, Go, Plus and Pro. 2026. Available at: https://openai.com/pricing/. Accessed on: 18 Aug. 2026.
[3] OPENAI. About ChatGPT Pro tiers. OpenAI Help Center, 2026. Available at: https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers. Accessed on: 18 Aug. 2026.
[4] OPENAI. Improving GPT-5.6 Sol in ChatGPT — and expanding access for Free users. 6 Aug. 2026. Available at: https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/. Accessed on: 18 Aug. 2026.
[5] OPENAI. Terms of Use. Effective 1 Jan. 2026. Available at: https://openai.com/policies/row-terms-of-use/. Accessed on: 18 Aug. 2026.
[6] OPENAI. Privacy Policy. 2026. Available at: https://openai.com/policies/row-privacy-policy/. Accessed on: 18 Aug. 2026.
[7] BRASIL. Constituição da República Federativa do Brasil de 1988. Brasília, DF: Presidência da República. Available at: https://www.planalto.gov.br/ccivil_03/constituicao/constituicao.htm. Accessed on: 18 Aug. 2026.
[8] BRASIL. Lei nº 10.406, de 10 de janeiro de 2002. Institui o Código Civil. Brasília, DF: Presidência da República. Available at: https://www.planalto.gov.br/ccivil_03/leis/2002/l10406compilada.htm. Accessed on: 18 Aug. 2026.
[9] BRASIL. Lei nº 8.078, de 11 de setembro de 1990. Dispõe sobre a proteção do consumidor e dá outras providências — Código de Defesa do Consumidor. Brasília, DF: Presidência da República. Available at: https://www.planalto.gov.br/ccivil_03/leis/l8078compilado.htm. Accessed on: 18 Aug. 2026.
[10] BRASIL. Lei nº 13.709, de 14 de agosto de 2018. Lei Geral de Proteção de Dados Pessoais — LGPD. Brasília, DF: Presidência da República. Available at: https://www.planalto.gov.br/ccivil_03/_ato2015-2018/2018/lei/L13709compilado.htm. Accessed on: 18 Aug. 2026.
[11] UNITED NATIONS. HUMAN RIGHTS COMMITTEE. General Comment No. 32: Article 14 — Right to equality before courts and tribunals and to a fair trial. CCPR/C/GC/32, 23 Aug. 2007. Available at: https://docstore.ohchr.org/SelfServices/FilesHandler.ashx?enc=Q7TyLKlkZy6AMMYILBtJdvBPX2s6eKHWibm1z7ee9AaJJSzx7ddl%2Bhdo8FKs0VQhNYR1HyG2kbxgw9flcW3JDw%3D%3D. Accessed on: 18 Aug. 2026.
[12] ORGANIZATION OF AMERICAN STATES. INTER-AMERICAN COMMISSION ON HUMAN RIGHTS. Access to Justice as a Guarantee of Economic, Social and Cultural Rights: A Review of the Standards Adopted by the Inter-American System of Human Rights. Washington, D.C.: IACHR. Available at: https://cidh.oas.org/countryrep/AccesoDESC07eng/Accesodesciv.eng.htm. See also: https://cidh.oas.org/countryrep/AccesoDESC07eng/Accesodescv.eng.htm. Accessed on: 18 Aug. 2026.
[13] EUROPEAN UNION. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 — General Data Protection Regulation — GDPR. Official Journal of the European Union, L 119, 4 May 2016. Available at: https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX%3A32016R0679. Consolidated HTML version available at: https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX%3A02016R0679-20160504. Accessed on: 18 Aug. 2026.
[14] UNITED STATES. Superior Court of Gwinnett County, Georgia. Walters v. OpenAI, L.L.C., Civil Action No. 23-A-04860-2. Summary judgment, 19 May 2025. Case analysis and access to decision: LOEB & LOEB LLP. Walters v. OpenAI, L.L.C. Available at: https://www.loeb.com/en/insights/publications/2025/05/walters-v-openai-llc. Accessed on: 18 Aug. 2026.
[15] REUTERS. OpenAI defeats radio host's lawsuit over allegations invented by ChatGPT. 19 May 2025. Available at: https://www.reuters.com/legal/litigation/openai-defeats-radio-hosts-lawsuit-over-allegations-invented-by-chatgpt-2025-05-19/. Accessed on: 18 Aug. 2026.
[16] UNITED STATES. Superior Court of the State of Delaware. Robert Starbuck v. Google LLC, C.A. No. N25C-10-211 MAA. Opinion denying Google LLC’s Motion to Dismiss. Decided 24 July 2026. Available at: https://courts.delaware.gov/opinions/download.aspx?id=398660. Accessed on: 18 Aug. 2026.
[17] CANADA. British Columbia Civil Resolution Tribunal. Moffatt v. Air Canada, 2024 BCCRT 149. Decision, 14 Feb. 2024. CanLII. Available at: https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html. Accessed on: 18 Aug. 2026.
[18] SOOKMAN, Barry B. Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot. McCarthy Tétrault, 19 Feb. 2024. Available at: https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot. Accessed on: 18 Aug. 2026.
[19] UNITED STATES. United States District Court for the Southern District of New York. Mata v. Avianca, Inc., 678 F. Supp. 3d 443, No. 1:22-cv-01461. Opinion and Order, 22 June 2023. Available at: https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1%3A2022cv01461/575368/54/. Accessed on: 18 Aug. 2026.
[20] UNITED STATES. United States Court of Appeals for the Ninth Circuit. Malkeet LNU et al. v. Todd Blanche, No. 24-4790. Order, filed 3 June 2026. Available at: https://cdn.ca9.uscourts.gov/datastore/opinions/2026/06/03/24-4790.pdf. Accessed on: 18 Aug. 2026.
[21] UNITED STATES. New York State Supreme Court. Cassata v. Michael Macrina Architect, P.C., 2026 NY Slip Op 26014. 27 Jan. 2026. New York State Law Reporting Bureau. Available at: https://www.nycourts.gov/reporter/3dseries/2026/2026_26014.htm. Accessed on: 18 Aug. 2026.
[22] UNITED STATES. New York Supreme Court, Appellate Division, Second Department. Landberg v. City of New York, 2026 NY Slip Op 03935. Decision and Order on Motion, 23 June 2026. Available at: https://www.nycourts.gov/reporter/current/3dseries/2026/2026_03935.shtml. Accessed on: 18 Aug. 2026.
[23] OPENAI. Introducing ChatGPT Go, now available worldwide. 16 Jan. 2026. Available at: https://openai.com/index/introducing-chatgpt-go/. Accessed on: 18 Aug. 2026.
PART III - ELON MUSK vs OPENAI
These are the direct links to the main sources used in the analysis of Elon Musk/xAI’s lawsuits against OpenAI, beginning with the main case. I checked the sources again just now; for procedural developments, prioritize the official court and GovInfo links.
-
Official page of the main case — Musk v. Altman et al., No. 4:24-cv-04722-YGR — U.S. District Court, Northern District of California
https://cand.uscourts.gov/cases-e-filing/cases/424-cv-04722-ygr/musk-v-altman-et-al
This is the main source for following the docket and the documents made available by the court itself. -
GovInfo — official collection of decisions in Musk v. Altman
https://www.govinfo.gov/app/details/USCOURTS-cand-4_24-cv-04722
It contains federal judicial decisions from the case made available by the U.S. Government Publishing Office. -
GovInfo — discovery decision/document dated July 1, 2025
https://www.govinfo.gov/content/pkg/USCOURTS-cand-4_24-cv-04722/pdf/USCOURTS-cand-4_24-cv-04722-0.pdf
Official document from the federal case. -
Official court page for the trial broadcast — Musk v. Altman trial: Listen live
https://cand.uscourts.gov/news/2026/05/01/musk-v-altman-trial-listen-live
It confirms the public audio broadcast of the 2026 trial. -
Reuters — withdrawal of Musk’s first state-court lawsuit, June 11, 2024
https://www.reuters.com/legal/elon-musk-withdraws-lawsuit-against-openai-2024-06-11/
It explains the first lawsuit filed in February 2024 and its withdrawal without prejudice. -
Reuters — January 2026 decision allowing important issues to proceed to trial
https://www.reuters.com/legal/litigation/musk-lawsuit-over-openai-for-profit-conversion-can-head-trial-us-judge-says-2026-01-07/
The article reports that Judge Yvonne Gonzalez Rogers found sufficient factual issues for trial, including matters concerning representations related to the nonprofit structure. -
Reuters — April 24, 2026: dismissal/withdrawal of the fraud claims and continuation of the remaining claims
https://www.reuters.com/world/us-judge-dismisses-musks-fraud-claims-openai-case-plans-proceed-trial-2026-04-24/
It explains which claims were removed before trial and which remained, including breach of charitable trust and unjust enrichment. -
Reuters — beginning of the trial, April 28, 2026
https://www.reuters.com/legal/litigation/openai-trial-pitting-elon-musk-against-sam-altman-kicks-off-2026-04-28/
It includes Musk’s position, OpenAI’s defense, and the request for approximately US$150 billion for the charitable entity. -
Reuters — key events in the trial and the May 18, 2026 result
https://www.reuters.com/legal/government/key-moments-musk-vs-openai-trial-2026-05-18/
This is one of the best summaries of the trial, including the statute-of-limitations theory and the positions presented by the parties. -
Reuters — specific report on Musk’s defeat on May 18, 2026
https://www.reuters.com/legal/government/elon-musk-loses-lawsuit-against-openai-2026-05-18/
It reports that the jury concluded that the lawsuit had been filed too late and that Musk announced his intention to appeal. -
Associated Press — May 18, 2026 decision
https://apnews.com/article/musk-openai-trial-verdict-0b9b0bfaffe96f2c930341f52dfe4f8c
AP also reports that the jury found that Musk had missed the legal deadline after approximately three weeks of trial. -
Bloomberg Law — “Musk v. Altman Fight Heads to Appeal With Core Claims Untested”
https://news.bloomberglaw.com/litigation/musk-v-altman-fight-heads-to-appeal-with-core-claims-untested
This source is especially important for the distinction I made: the defeat resulted from the statute of limitations issue, while the central substantive allegations remained, to a significant extent, without a definitive merits determination. -
Reuters — post-trial analysis, May 19, 2026
https://www.reuters.com/legal/government/musks-failed-court-attack-openai-could-leave-lasting-scars-ceos-reputation-2026-05-19/
It examines the consequences of the trial and confirms the central role of the statute-of-limitations issue in the result. -
Separate xAI lawsuit against OpenAI — trade secrets — prior decision available on Justia
https://law.justia.com/cases/federal/district-courts/california/candce/3%3A2025cv08133/456862/73/
This is X.AI Corp. et al. v. OpenAI, Inc., No. 3:25-cv-08133-RFL, before Judge Rita F. Lin. -
Reuters — June 15, 2026: dismissal, with prejudice, of xAI’s trade-secret lawsuit against OpenAI
https://www.reuters.com/legal/litigation/openai-wins-dismissal-trade-secret-lawsuit-by-musks-xai-2026-06-15/
The decision concluded that xAI had not alleged sufficient facts to show that OpenAI induced the disclosure of the alleged trade secrets. -
Complaint in the X Corp./xAI antitrust lawsuit against Apple and OpenAI — full PDF
https://fingfx.thomsonreuters.com/gfx/legaldocs/klpybbxzxvg/x%20v%20apple%20openai%20lawsuit%2020250825.pdf
This is the complaint filed in the Texas action against Apple and OpenAI. -
Associated Press — X/xAI antitrust lawsuit against Apple and OpenAI, August 2025
https://apnews.com/article/8cc360bd419894ad8c6bfdd79eb5693f
It explains the allegations concerning ChatGPT’s integration with Apple and the alleged competitive disadvantage imposed on Grok. -
Reuters — November 2025: Apple and OpenAI failed to obtain immediate dismissal of the antitrust lawsuit
https://www.reuters.com/world/apple-openai-must-face-x-corps-lawsuit-now-us-judge-rules-2025-11-13/
This is important because the judge allowed the litigation to proceed, without yet deciding the merits of the antitrust allegations.
To follow the updated main case, the most important link is No. 1; to consult the official judicial decisions, primarily use Nos. 2 and 3. To understand why Musk’s defeat does not amount to a merits decision absolving OpenAI of the central allegations, the most important links are Nos. 9, 10, 11, and 12.
#WhoControlsTheTruth #ScottErikStafne #ElonMusk #OpenAI #ArtificialIntelligence #AI #AIGovernance #AIAccountability #CorporateAccountability #AIEthics #AIMisinformation #AIFalseInformation #AIFalseAccusations #ReputationalHarm #RightToRedress #AccessToJustice #EqualityOfArms #ConsumerRights #ConsumerProtection #DigitalRights #HumanRights #RuleOfLaw #DueProcess #JudicialIntegrity #ComputationalInequality #DigitalInequality #AlgorithmicAccountability #AlgorithmicJustice #AITransparency #AIResponsibility #AIRegulation #TechLaw #LawAndTechnology #LegalTech #JusticeAndTechnology #TruthAndJustice #EvidenceMatters #VerifyTheEvidence #KeepTheInquiryOpen #NeverEndingInquiry #PublicInterestAI #AIAndLaw #FutureOfJustice #DigitalJustice
