Integration Capacity Analysis

AI Companies Keep Rewriting the Promises They Started With

What OpenAI, Anthropic and Google said at the beginning, what is left of it, and how anyone could check. Wim Van Laere · The Great Homecoming research programme · September 2026

Three papers on the AI companies · part one

  1. AI companies keep rewriting the promises they started with — the argument. You are reading this one.
  2. Nobody broke a promise — the evidence: fifteen written commitments from three companies, each fixed at its first published version, and what happened to each one. Publishing shortly.
  3. What keeps them standing takes away their direction — why it keeps happening, and what it would take to change it. Publishing shortly.
Ninety seconds from Newsroom with Sana, Pakistan TV, 14 September 2026.

The question underneath the safety argument

Almost the entire public argument about artificial intelligence is about safety. Can the machine be controlled. Will it get out of hand. Who is ahead. There is a question underneath all of that which is barely asked, and it is the one that decides everything else. What is this for, and for whose benefit?

Every one of these companies answered that question in writing when it started. It is worth reading those answers again.

What they said at the beginning

OpenAI was founded as a non-profit whose charter put humanity above investors, with returns to investors capped and everything above the cap going to the non-profit. Anthropic was set up as a public-benefit company with a trust intended to hold the mission above profit. Google published AI principles promising it would not develop artificial intelligence for weapons and would not develop it for surveillance. All three were promises to the public about purpose — not about safety, about purpose.

Now look at where those answers stand. The investor cap does not appear anywhere in OpenAI’s restructuring of October 2025, nor in either attorney-general release about it. Anthropic’s commitments were rewritten in February 2026 into what the company itself now calls “public goals that we will openly grade our progress towards” rather than hard commitments. Google removed the weapons and surveillance pledge from its principles on 4 February 2025.

None of this was hidden. Each change was published, dated, and in most cases explained. That is the part that should worry people more than a secret would: it was done in the open, and there was no mechanism through which anyone could object.

I have examined this in detail for three of these companies — fifteen written commitments, each fixed at the version first published, each compared with what actually followed. That study is the second paper in this series. One example stands for the rest. In February 2025 Meta’s published rule said that at the critical risk level it would stop development. In April 2026 that became “develop with mitigations.” In May 2026 the company published its own report finding that its own new model, unmitigated, met the high-risk threshold for chemical and biological danger. The model was released.

Why nobody is in a position to win

The answer usually given to all of this is that there is a race, and the race must be won. I do not think anybody is in a position to win it, for three reasons.

First, nobody has a decisive edge. Not in money, not in technology, not in resources, not militarily. It is an illusion to think one tool will suddenly produce one, and we are still far from anything like general intelligence.

Second, and this is the one that matters: capacity amplifies whatever direction it is applied in. A more capable system does not become wiser. It becomes more effective at what it was already doing. Applied inside a contest where every party is trying to gain at the others’ expense, greater capability does not settle the contest — it deepens the fragmentation, between countries and equally inside them. So the tool that is supposed to secure a country’s position can just as easily accelerate its coming apart.

Third, winner-take-all creates its own dangers. It invites people with bad intentions to use the same capability. The cost of the build-out is now large enough to put strain on financial and ecological systems that were already strained. And a system trained inside a zero-sum contest is being taught what zero-sum behaviour looks like. If anybody is genuinely worried about machines behaving competitively, it is worth asking what we are currently showing them.

What happens to tools built for advantage

There is a pattern in this, and it is not new. The Stinger missiles given to Afghan fighters in the 1980s. The cyber weapons written for Stuxnet, which taught Iran what cyber warfare could do. The stolen intelligence-agency exploit that became WannaCry and shut down hospitals in the West. Armed drones, now in the hands of the Houthis. Weapons sent to Syrian rebels that reached ISIS. Social media influence tools, now pointed back at the countries that pioneered them.

Every one of those was built to gain an advantage. Every one came back. Artificial intelligence spreads more easily than any of them. Model weights can be published. Capability can be trained by anybody with enough compute and enough patience. Whatever edge exists is temporary by construction — which is one more reason that the honest question is not who wins, but what the thing is for.

What could be done

I am not asking anyone to build a new organisation. New organisations take years, need everyone to agree first, and usually arrive after the thing they were meant to govern has already happened. This can start without any of that. There are three steps.

  1. Every company building these systems writes down what it is for, in public. Most of them have already done this once. The statement gets a date, and that dated version is the one that counts from then on.
  2. Someone other than the company compares its conduct with that statement. This is the part that matters, and it has to be the first version rather than the one on the website today. The difference decides the whole result. If a company quietly lowers its own rule and then keeps to the lower rule, it passes any check that uses the current text.
  3. The result is published — including the parts that are uncertain and the places where the people doing the checking disagree with each other.

Who does the checking? At the start, anyone who is willing to do it properly: a research institute, a university, a newspaper, a regulator. The method has to be public so that other people can run it and get a different answer if the evidence supports one.

This is what my own work measures. I have run it on three companies and fifteen written commitments. Two people read the same evidence without seeing my conclusions and disagreed with me in four places. A separate check re-gathered the evidence from original sources and found two mistakes in my own starting points. All of that is written up in the papers that follow. It is a first attempt, not a verdict on anyone.

A note on where this belongs. Most people’s first thought is the UN Security Council. I think that is the wrong place. One government has already shown that it can switch a company off, restrict who abroad may use its models, and order its own agencies to stop buying from it. Giving that same government a veto over the rules would not solve the problem. A published method and an open register need no Council, and are much harder to block.

The talks between Washington and Beijing

Talks on AI safety between the two governments were reported for the middle of this month. I do not expect much from them. There is very little trust, and almost everything technical is now treated as part of warfare. China’s objection has a point that is hard to argue with: you cannot run a competition where only one side’s rules count.

But the Cold War is useful here, because progress was made then between two sides who trusted each other even less than these two do. Progress came from agreeing on procedures, not from agreeing on values. Four things are realistic, and none of them require either side to trust the other:

  1. A direct line. A known way for each side to reach the other quickly when something goes wrong with a system, without going through diplomats.
  2. Incident notification. An agreement to tell the other side when a system has behaved in a way nobody intended, in the way that nuclear and aviation incidents are reported.
  3. Agreed no-go targets. A short list of things neither side will use these systems against — power grids, hospitals, water systems, early-warning systems. Narrow, and verifiable by the damage if it is ever broken.
  4. Researchers who talk to each other. Working contact between the safety people on both sides, which is how arms control actually functioned for decades.

The step proposed above follows naturally from those four. Once both sides can talk about incidents, the next question is what the systems are being built for in the first place — and that is a question each side can answer in writing, and be measured against, without conceding anything to the other.

What this does not do

None of this forces anyone to do anything. There is no enforcement in it, and I will not pretend otherwise. What it does is keep the record straight: what was promised, on what date, by whom, and what happened afterwards.

That may sound like a small thing, but it matters for one practical reason. When a company rewrites its own rule, the old version disappears from its website. A year later there is nothing to compare anything against, and the argument about whether a promise was kept becomes an argument about memory. Writing the first version down, with its date, while it can still be found, is the only thing that prevents that.

The test

There is a straightforward test in all of this, and it works both ways. If these companies and the governments behind them are right that this is being built for everyone’s benefit, then writing that down and being checked against it costs them nothing at all. It is only awkward for an organisation that does not intend to keep its word.

And if the danger is real, then speed is a strange answer to it. The problems that are pulling the world apart — debt, ecological damage, inequality, and the loss of any shared rule that enough countries still accept — are the same problems this technology is making faster. I have written elsewhere about how that shared rule drained away over fifty years, from a common belief in liberal democracy, through markets, through finance, to the every-country-for-itself of today. Artificial intelligence did not cause that loss, but it feeds on it and speeds it up.

So the question stays where it started. What is this being built for, and who benefits? Anyone who has a good answer should have no trouble writing it down.

Next in this series

This piece is the argument. The two that follow are the working behind it. The second, Nobody broke a promise, takes fifteen written commitments from OpenAI, Meta and Anthropic, fixes each one at the version first published, and sets it against the public record — including where two independent readers disagreed with me and where an independent check found mistakes in my own starting points. The third, What keeps them standing takes away their direction, asks why the commitments moved in the direction they did, and what would have to be true for them to move back.


The Great Homecoming is an independent research programme on why systems cohere or fragment. This essay is the first of three on the AI companies; the two that follow set out the evidence behind it — fifteen written commitments from OpenAI, Meta and Anthropic, each fixed at its first published version and compared with the public record, and the mechanism that would explain why they moved. The commitments and changes cited here, with their sources: the OpenAI Charter (April 2018) and the capped-return structure (March 2019), against the recapitalisation announcement and the Delaware and California attorney-general releases (October 2025); the Anthropic Responsible Scaling Policy (September 2023) against version 3.0 (February 2026); Google’s AI Principles against the removal of the weapons and surveillance pledge (4 February 2025, first reported by The Washington Post and CNBC); Meta’s Frontier AI Framework (February 2025) against the Advanced AI Scaling Framework (April 2026) and the Muse Spark Safety and Preparedness Report (26 May 2026); and the reported US–China AI safety talks of mid-September 2026 (Reuters). The video is an extract from Newsroom with Sana, Pakistan TV, 14 September 2026, used with the broadcaster’s credit. The instrument used is research-grade and under live forward test; its reads claim consistency with the evidence, not validation. Contact: Wim Van Laere.