What the complaint alleges
The plaintiffs are consumers who pay for ChatGPT, Claude, Grok and Gemini. Their claim is under the Sherman Act, and its structure is conventional even if its subject is not: an agreement among horizontal competitors that restrains output or quality, causing overcharge or diminished value to purchasers.
The alleged agreement is built from public statements. The complaint points to an essay published on 12 September 2026 by Anthropic's chief executive, Dario Amodei, calling for industry-wide cooperation on decelerating capability development in favour of safety, and alleges that Sam Altman, Elon Musk and Demis Hassabis publicly assented the same day. It also points to a July 2026 joint statement signed by senior employees at several leading laboratories acknowledging intense competitive pressure not to unilaterally slow development — a document the plaintiffs read as evidence that the participants understood unilateral restraint to be unsustainable, and therefore needed a collective one.
The complaint concedes something important, and the concession is what makes it interesting. Plaintiffs accept that each firm may unilaterally slow itself down as much as it likes. A company that decides on its own to ship less capability, for whatever reason, commits no antitrust offence. The alleged wrong is the substitution of what the complaint calls collective restraint for individual accountability.
That is, doctrinally, exactly the right place to aim. It is also the place where AI safety governance has been standing for several years.
Why this is not obviously frivolous
The instinctive reaction among people who work on AI policy has been that the suit is opportunistic — that safety is plainly a legitimate objective and antitrust cannot possibly forbid pursuing it together. That reaction underestimates the doctrine.
United States antitrust law has repeatedly rejected the argument that a horizontal agreement is lawful because its purpose is good. The professional-standards cases are the cleanest illustration: engineers argued that a rule against competitive bidding protected public safety, because price competition would pressure engineers into unsafe designs. The Supreme Court held that this was not a defence — the Sherman Act does not permit competitors to agree that competition itself is dangerous, and safety concerns are for legislatures and regulators to address, not for competitors to resolve by agreement among themselves. Similar reasoning has defeated agreements among professionals to restrict advertising, agreements among firms to withhold a product feature, and agreements to suppress technology on ostensibly protective grounds.
The principle these cases share is uncomfortable for AI safety: an agreement among competitors to provide a less capable product is a restraint on quality, and quality restraints are treated as restraints on competition even when the participants sincerely believe the restraint is socially beneficial. The law's answer to "but competition here produces bad outcomes" is that the remedy is regulation, not private agreement.
Two things complicate the plaintiffs' path. First, they must prove agreement — concerted action, not merely parallel conduct plus public exhortation. Conscious parallelism is lawful. Four chief executives publicly agreeing that the industry should be careful is, standing alone, closer to public advocacy than to a contract. The plaintiffs will need something that looks like reciprocal commitment. Second, conduct of this kind is likely to be assessed under the rule of reason rather than treated as automatically unlawful, which opens the door to justification — though "our product is dangerous" is an awkward justification for a defendant to run at scale, for reasons that have nothing to do with antitrust.
The collision that has been coming since 2023
Step back from the pleading and the structural problem is visible.
Almost every serious AI governance proposal of the past three years has asked competing developers to coordinate. Voluntary commitments negotiated with governments ask laboratories to adopt common safety practices. Codes of practice ask signatories to converge on evaluation methods. Frontier safety frameworks ask firms to adopt capability thresholds at which they will pause. The UN scientific panel's thematic brief of 21 September 2026 asks for shared standards, verification and monitoring of capability thresholds. The declaration on human control adopted by twenty-two states the same week asks for common standards, shared incident reporting and independent evaluation.
Every one of those is a request that competitors do the same thing, at the same time, in a way that constrains output. That is a description of coordination, and coordination among horizontal competitors is the category antitrust exists to police.
Regulators have noticed. The chair of the Federal Trade Commission expressed scepticism on 15 September 2026 — three days before the complaint was filed — about calls for an antitrust exemption for AI safety cooperation. That timing is not coincidence: the industry had been asking for cover, the enforcer declined to give it, and private plaintiffs filed into the space that refusal created.
The resulting position is genuinely incoherent. International bodies are asking for coordinated restraint. The competition enforcer is declining to bless it. Private plaintiffs are suing over it. A developer that ignores the safety consensus faces regulatory and reputational exposure; one that joins it faces treble damages. There is no conduct that is safe in both directions, and that is a policy failure rather than a litigation risk.
The evidence problem: when does talking become agreeing?
Because the alleged agreement here is assembled entirely from public material, the case turns on a question antitrust has litigated for decades without ever settling cleanly: at what point does visible, parallel behaviour become concerted action?
The baseline rule is settled. Conscious parallelism is lawful. Firms in a concentrated market watch each other, anticipate each other and often move together, and the law tolerates this because the alternative would be to punish rational business behaviour. A plaintiff must show something more — the evidence that tends to exclude the possibility that the defendants acted independently.
Courts look for what are usually called plus factors, and it is worth running the alleged facts against the standard list, because it shows where this complaint is strong and where it is thin.
An invitation and an acceptance. This is the plaintiffs' best material. An essay calling explicitly for industry-wide cooperation on deceleration is, structurally, an invitation. Public assent by three named competitors on the same day is, structurally, an acceptance. Antitrust has long recognised that an agreement may be inferred from an invitation publicly extended and publicly taken up, and that no contract, no meeting and no secrecy is required. Whether these particular statements carry that weight is a question of their actual content, which the pleadings will contest line by line.
Action against independent self-interest. Normally the hardest factor to establish; here it is unusually available. Shipping less capability than one is technically able to ship, in a market where capability is the principal axis of competition, is difficult to explain as unilateral profit maximisation. The defendants' answer will be that reputational and regulatory risk makes restraint individually rational — which is a serious answer, and arguably the correct one, but it requires them to argue that safety restraint pays for itself, which sits awkwardly beside the industry's own lobbying about compliance burden.
A motive to conspire. Easy to plead in any concentrated market and correspondingly weak as evidence. It will be pleaded.
Inter-firm communications. Unknown, and the reason discovery matters. The July 2026 joint statement by senior employees across several laboratories is the plaintiffs' bridge to this factor: a document signed across firm boundaries, acknowledging that unilateral restraint is competitively unsustainable, is precisely the artefact a plaintiff wants. Its ordinary reading is an expression of concern by individuals, not a commitment by undertakings. Its litigation reading is an admission that the participants understood the collective-action structure of the problem and needed a collective answer.
This is where the case is genuinely hard to call. The defendants will say, with real force, that a chief executive publishing an essay is speech, that agreeing in public that an industry should be careful is advocacy, and that treating advocacy as agreement would make it impossible for any firm to participate in public debate about its own sector. That argument has constitutional as well as doctrinal weight in the United States, and it is the strongest card the defendants hold.
The plaintiffs will say that if four firms controlling the overwhelming share of a market publicly commit to the same restraint on the same day, the fact that they did it in public rather than in a hotel room should not immunise them — and that a rule to the contrary would simply relocate every cartel to the op-ed page. That argument also has force.
My own expectation is that the complaint is dismissed on this ground, with leave to amend, and that the conflict it exposes is not thereby resolved in the slightest.
What it means for the signatories of Europe's code of practice
There is a group of firms with a specific and underappreciated exposure here: those that have signed the European Union's general-purpose AI code of practice.
The code occupies an unusual legal position. It is voluntary in form — nobody is compelled to sign — but adherence is the practical route to demonstrating compliance with the AI Act's obligations on providers of general-purpose AI models, and the Commission's public signatory register makes participation visible. Signatories commit, among other things, to evaluation practices, systemic-risk assessment and safety-and-security frameworks.
Now consider how this appears to a plaintiff's lawyer. A group of competitors has signed a common document committing them to common constraints on how they develop and deploy their principal product. The document was negotiated with participation from those same competitors. Adherence is visible, which makes defection detectable — and detectability of defection is, in cartel economics, the mechanism that makes a restraint stable.
The defence is straightforward and, I think, correct: the code implements a legal obligation, the Commission convened and approved it, and conduct required or supervised by a public authority is not an agreement between undertakings in the relevant sense. European law recognises this — the state-action reasoning, whatever its precise formulation, exists for exactly this situation.
But the defence depends on the public character being real. It weakens as the code's content moves away from what the AI Act requires and toward what the signatories thought advisable; it weakens where the drafting was industry-led with light official supervision; and it is weakest for commitments that go beyond the statutory obligation, which is the category that safety-conscious firms are most likely to accept and most likely to be praised for.
This is a genuine trap. A firm that signs only what the law requires has a clean answer. A firm that signs more than the law requires, because it believes more is warranted, has converted a voluntary act of responsibility into a horizontal commitment without a mandate behind it. The lesson is not that firms should do less. It is that anything beyond the statutory floor should be adopted and published unilaterally, not agreed collectively — the same conclusion the American case points to, arrived at from a different direction.
How other sectors resolved this
The conflict is not new. It is simply new to AI, and the resolutions reached elsewhere are instructive precisely because none of them was achieved by the industry sorting it out privately.
Standard-setting organisations are the most familiar. Competitors sit together and agree that plugs will have three pins. This is lawful, and the conditions are well understood: open participation, transparent procedure, objective criteria, and commitments to license essential patents on fair terms. The lawfulness comes from the process, not from the worthiness of the goal. Standards bodies that have drifted from open procedure into exclusion have been found to violate competition law even where the technical merits were sound.
Safety regulation is the other route: the coordination is performed by a public authority rather than by the firms. Aircraft manufacturers do not agree among themselves on airworthiness; a regulator sets the standard and all of them comply. The restraint is identical in effect — nobody may ship a non-compliant aircraft — but its source is public, which removes it from antitrust entirely.
Statutory exemptions exist where legislatures have decided coordination is necessary: research joint ventures, certain export arrangements, some insurance data-sharing. These are narrow, deliberate, and granted by legislatures rather than assumed by participants.
What all three have in common is that the coordination is either publicly conducted, publicly mandated, or publicly authorised. AI safety coordination has so far been none of these. It has been private, voluntary, and conducted among a small number of firms with enormous shared commercial interests — which is precisely the shape antitrust treats with suspicion.
The uncomfortable part: the restraint really might be self-serving
I want to give the plaintiffs' strongest argument rather than the weakest, because the weakest is easy to dismiss.
The strongest version is not that safety is a pretext in the sense that the executives are lying. It is that the incentives of incumbents and the requirements of safety point in the same direction often enough that we cannot tell them apart from the outside, and antitrust exists for exactly the situations where we cannot tell them apart.
Consider what a norm of coordinated deceleration does to a market. The incumbents have the largest deployed models, the most compute, the most distribution and the most revenue. A new entrant's only realistic path is to offer something the incumbents do not — usually more capability, fewer restrictions, or faster iteration. An industry-wide norm that capability advancement should be slowed, enforced by reputational pressure and by the safety expectations of enterprise buyers, disproportionately burdens the entrant. The incumbent loses some future capability it can afford to lose. The entrant loses its only differentiator.
The same asymmetry appears in compliance cost. Shared evaluation standards, third-party audits and red-teaming regimes have a substantial fixed component. Fixed costs are trivial spread over a large revenue base and prohibitive spread over a small one. Regulation that raises fixed costs is regulation that consolidates the market, whatever its intent — an effect well documented in financial services and pharmaceuticals, and not seriously contested.
None of this establishes bad faith. It establishes that a genuinely safety-motivated agreement and a genuinely exclusionary one would look identical from outside, and would produce identical market structure. Antitrust doctrine's insistence that good motive is not a defence is a response to precisely this epistemic problem. It is not judicial cynicism about motives; it is an admission that motives are unobservable and effects are not.
What happens if the plaintiffs survive a motion to dismiss
The immediate practical question is not whether the plaintiffs ultimately win. It is what happens at the pleading stage, because that alone changes behaviour.
If the complaint survives, discovery opens into the internal communications of four frontier developers on the subject of coordination — which is a substantial event regardless of outcome, and one the defendants will weigh heavily. More importantly, counsel at every major laboratory will immediately advise caution about joint safety commitments, participation in shared evaluation initiatives, and public endorsement of competitors' restraint. That advice will be correct. It will also degrade exactly the coordination that the UN panel, the twenty-two-state declaration and the EU's code-of-practice architecture all depend upon.
The chilling is asymmetric in a way that matters. Formal standard-setting through recognised bodies is comparatively safe, because the procedural safeguards are established. Informal coordination — a joint letter, a shared threshold commitment, a public statement of mutual restraint — is exposed. The suit therefore pushes safety coordination out of the fast, flexible channel and into the slow, formal one, at a moment when the pace of capability development is the entire concern.
If the complaint is dismissed, the effect is not symmetric. A dismissal on the ground that public exhortation is not agreement leaves the underlying conflict entirely intact, and the next complaint will simply plead better facts. Nothing about a dismissal resolves whether coordinated restraint is lawful; it resolves only whether these plaintiffs pleaded it adequately.
The European position is different, and not obviously better
European Union competition law reaches the same conduct through Article 101 of the Treaty, which prohibits agreements that restrict competition — expressly including agreements to limit or control production or technical development. An agreement to slow technical development is close to a textual match.
The European framework does, however, contain something the United States lacks: Article 101(3), which exempts agreements that produce benefits, pass a fair share to consumers, are indispensable to achieving those benefits, and do not eliminate competition. This provides a doctrinal home for a safety justification that American law does not readily offer.
The Commission has also shown recent willingness to accommodate cooperation on sustainability, publishing guidance on when agreements pursuing environmental objectives fall outside the prohibition or qualify for exemption. That guidance is the closest available template for AI safety agreements, and the analogy is instructive in both directions. It shows the analysis is possible. It also shows how demanding the conditions are — indispensability in particular, which asks whether the objective could have been achieved by less restrictive means. For AI safety the answer is frequently yes: regulation could achieve it, which is an argument that private agreement is not indispensable.
There is a deeper structural difference. The European Union is in the process of mandating much of what American firms are coordinating voluntarily. Conduct compelled by regulation is not an agreement between undertakings. To the extent the AI Act and its code of practice convert voluntary coordination into legal obligation, the European antitrust problem shrinks — which is an argument, from an unexpected direction, that binding regulation is better for competition than voluntary safety consensus.
Where Türkiye stands
Turkish competition law tracks the European model. Article 4 of Law No. 4054 prohibits agreements and concerted practices that prevent, distort or restrict competition, and Article 5 provides for exemption on conditions closely modelled on Article 101(3). A coordinated restraint on product development by competing undertakings would fall for analysis under Article 4, with the exemption conditions — including indispensability — applying in the familiar way.
Türkiye is not a bystander in this area. The Rekabet Kurumu opened a comprehensive sector inquiry into artificial intelligence in April 2026, covering the value chain, foundation models, access to data and compute, vertical integration, lock-in and merger concentration. In June 2026 it opened an investigation into Meta over the integration of Meta AI into WhatsApp and the foreclosure of third-party AI assistants, imposing interim measures requiring that third-party assistants be allowed without practical or economic obstacles. Both of those are conventional exclusion cases rather than coordination cases — but an authority already examining foundation-model market structure is an authority that will have a view when Turkish or multinational firms begin adopting industry-wide AI safety commitments.
The practical exposure for Turkish businesses is more immediate than it appears, and it is not about frontier laboratories. Sectoral AI codes of conduct are proliferating — in banking, in insurance, in e-commerce, in advertising. A sectoral association that convenes its members to agree on what AI features they will not deploy, what data they will not use, or what capability they will not offer is engaged in coordination among competitors, whatever the document is titled. Self-regulation drafted by industry associations is a recurring subject of competition enforcement everywhere, and the safety framing offers no immunity.
What organisations should actually do
For anyone participating in collective AI safety work, the following distinctions do real work:
Prefer public mandates to private agreements. If a restraint is worth having, the durable route is regulation. A rule imposed by an authority is not an agreement among competitors and does not require an exemption analysis.
If coordinating privately, borrow standard-setting procedure. Open participation, published criteria, documented decision-making, no exclusion of firms that decline to join. These are not formalities; they are the features that have historically distinguished lawful standard-setting from unlawful concert.
Coordinate on methods, not on output. Agreeing how to measure a capability is standard-setting. Agreeing not to ship a capability is output restriction. The first is defensible in every jurisdiction discussed here. The second is where the exposure sits.
Keep the commercial discussion out of the safety room. The single most damaging fact pattern in any coordination case is a safety forum in which pricing, market shares, customers or launch timing were also discussed. This is the ordinary discipline of trade-association practice and it is routinely ignored in AI safety settings, which tend to be informal and small.
Document the counterfactual. If a restraint is genuinely indispensable, the file should show why less restrictive alternatives were considered and rejected. That record is the exemption analysis, and it must be created contemporaneously or it is worthless.
Unilateral is safer than collective. A firm that adopts a safety threshold on its own, publishes it, and does not condition it on competitors doing likewise has a materially stronger position than one that participates in a joint commitment — even where the substance is identical.
The remedy problem
One question the commentary has skipped entirely: what would the plaintiffs actually win?
It is worth asking, because the answer says something about whether this is a serious case or a pressure exercise.
Damages are the conventional remedy, and here they are peculiar. The measure would be the difference between the value of the subscription the plaintiffs received and the value of the subscription they would have received in a competitive market — that is, one in which the four defendants raced on capability. Quantifying the consumer surplus lost to capability that was never built is not an ordinary damages exercise. It requires an expert to model a counterfactual product that does not exist, and then to price the gap. Antitrust economists do build counterfactuals, routinely, but usually for prices in markets with observable history. There is no observable history of a deliberately unrestrained frontier model market, because there has never been one.
Trebling makes the number matter even if the base is speculative, which is why the pleading stage is where the leverage sits.
Injunctive relief is stranger still. What would the order say? A court cannot sensibly enjoin four companies to develop capability faster. It could enjoin them from agreeing to restrain development — an order against concerted action rather than for any particular conduct. That is a conventional antitrust remedy and it would be workable. It would also, in practice, operate as a judicial instruction that the four largest AI developers must not coordinate on safety restraint, which is a startling thing for a court to order in the same month that a UN scientific panel asked for exactly that coordination and twenty-two states declared in favour of it.
I do not think the plaintiffs are indifferent to this. A judgment of that shape would be cited in every AI governance debate for a decade, and its citation would not be about consumer welfare.
What to watch next
The motion to dismiss and, specifically, how the court treats the gap between public advocacy and concerted action — the pleading standard applied here will determine whether this becomes a recurring form of litigation. Whether the defendants raise a safety justification at all, or argue only that no agreement existed; a defendant that pleads justification concedes the restraint. Whether any antitrust authority, in the United States or Europe, issues guidance analogous to the sustainability guidance. Whether the code-of-practice architecture in Europe is recast as obligation rather than commitment, which would remove the problem there. And whether sectoral AI codes in Türkiye and elsewhere start attracting competition scrutiny, which is where most organisations will actually encounter this.
Frequently asked questions
Is it really illegal for companies to agree to make their products safer?
Not as such. The exposure arises where the agreement restrains output, quality or technical development, and where the coordination is private rather than mandated or conducted through open standard-setting. Agreeing on how to test is very different from agreeing what not to ship.
Does a good motive protect the participants?
In United States law, largely no — the courts have repeatedly refused to treat a socially beneficial purpose as a defence to a horizontal restraint. In European and Turkish law there is an exemption route, but it requires the restraint to be indispensable, which is difficult to show when regulation could achieve the same end.
Do public statements by chief executives amount to an agreement?
Standing alone, usually not. Parallel conduct and public exhortation are lawful; the plaintiffs need evidence of reciprocal commitment. This is the likeliest battleground on the motion to dismiss, and the likeliest ground for dismissal.
Does this affect Turkish companies?
Yes, though not through this case. The relevant risk is sectoral: associations convening competitors to agree on AI practices, features or data use engage Article 4 of Law No. 4054 regardless of the safety framing. The Rekabet Kurumu is already active in AI market structure.
Is the answer to give AI safety an antitrust exemption?
It is one answer, and the FTC chair's September scepticism suggests it will not be given readily. Exemptions are blunt, hard to reverse, and tend to entrench whoever holds them. Mandating the safety measures through regulation achieves the same restraint without creating a zone of immunity.
What is the safest form of participation in collective safety work?
Open, procedurally regular standard-setting on measurement and evaluation methods, with published criteria and no exclusion of non-participants — plus unilateral publication of the firm's own thresholds rather than joint commitment to shared ones.
Burhan Doğuş Ayparlar's View
This section sets out my personal assessment as the founder of this site and an AI ethics & compliance counsel.
My first reaction to this complaint was that it was cynical, and I want to record that reaction because I no longer hold it. Reading the pleading against the case law, the plaintiffs have identified something real that the AI policy community has been avoiding — and avoiding, I think, because the alternative is uncomfortable for everyone involved.
The uncomfortable part is this. For three years the governing assumption of AI safety policy has been that the developers should coordinate, and that the appropriate role of the state is to encourage, convene and eventually formalise that coordination. Voluntary commitments, frontier safety frameworks, codes of practice, joint statements: all of it rests on the premise that competitors doing the same restrained thing at the same time is the desirable outcome. Competition law has spent a century on the view that competitors doing the same restrained thing at the same time is the thing most likely to harm the public, and that sincerity of purpose is irrelevant because purpose cannot be observed and effects can. Both propositions are defensible. They are not compatible, and nobody has done the work of reconciling them.
What I find genuinely persuasive in the plaintiffs' position is not the claim of pretext. I do not think these executives are lying about safety. It is the observation that a sincere safety agreement and an exclusionary one have the same shape, produce the same market structure, and burden the same party — the entrant who has nothing to offer except the capability the incumbents have agreed to withhold. Antitrust's refusal to inquire into motive is not cynicism. It is an admission that we cannot reliably distinguish these cases from the outside, and that the market structure consequences are identical either way. I have argued the opposite in client memoranda. I was reasoning from intention, which is the error the doctrine is designed to prevent.
The conclusion I draw is not that safety coordination should stop. It is that the voluntary model was always the wrong instrument, and this case exposes why. Voluntary industry coordination has three defects that have nothing to do with antitrust and that antitrust happens to surface: it is unenforceable against the participant who defects, it is unaccountable to anyone outside the room, and it confers on a handful of firms the power to set the terms on which everyone else may compete. The response to all three is the same, and it is the response the UN panel arrived at from the safety side and the competition enforcer arrived at from the market side: put the obligation in law, apply it to everyone, and let a public authority rather than a private consensus decide what restraint is required.
For Türkiye I would flag the sectoral exposure rather than the frontier one, because that is where Turkish businesses will actually meet this. We are entering a period in which banking, insurance, healthcare and advertising associations will draft AI codes of conduct, and they will draft them the way associations always do — in a room with competitors, by consensus, with the sincere belief that a common standard protects consumers. Some of those codes will restrict what members may offer. Under Article 4 of Law No. 4054 that is an agreement among undertakings restricting competition, and the exemption analysis under Article 5 will ask whether it was indispensable — which, where the Türkiye AI Action Plan has already promised a risk-based regulatory framework and sectoral sandboxes, will be a hard question to answer yes. The advice is simple and worth giving early: if the sector wants a rule, ask the regulator for a rule. Do not write it yourselves and call it self-regulation.
The broader point, and the one I would not have made a month ago: the strongest argument for binding AI regulation may turn out to be a competition argument rather than a safety one. Regulation does not merely produce better compliance than voluntary commitment. It is the only form of restraint that does not hand the incumbents a cartel — and the only one that survives contact with a plaintiff who reads the case law carefully.
This article is for information only and does not constitute legal advice. It describes allegations in a complaint filed on 18 September 2026 as reported in the press. Those allegations are unproven, no court has ruled on them, and nothing here should be read as a finding that any defendant acted unlawfully. Statements about Turkish, EU and United States law are general in nature; specific cases require individual assessment. The analysis and assessments are the author's own.
