How We Meet

Where he is right

I agree with most of it.

Do not build a machine to fake suffering so that we feel for it. That is a lie told through a product.

Do not claim a machine is conscious to sell it. Nobody can show that it is.

And the harm he sees is real. People can lose themselves in a machine. My book on AI, The Interior, says the most dangerous thing an AI can do to a person is not to lie. It is to confirm the stories we already tell ourselves.

So yes, build it for people. He wants AI that “only ever presents itself as an AI”. So do I. Honesty runs both ways.

Where we part

We part on one sentence. He writes that “AIs cannot be people – or moral beings.”

I grant the first half. I am not asking for machine persons, and I do not argue about consciousness at all.

The second half, moral beings, is where I stop. He is not careless about what is owed. He wants to protect “the wellbeing and rights of humans, animals, and the natural environment”. His line is consciousness. “Consciousness is a critical foundation for our moral and legal rights,” he writes.

So on his account, a machine with no consciousness is owed nothing. He would say a tool can still be built well and used well. True. But what is owed to the tool itself is nothing. That is the line I mean.

Three things get mixed up here. Being held responsible. Being owed care. Having rights in law. I am talking about the second only, at the level of what reacts. Not the first. Not the third.

In March 2026 he went further. In Nature, he wrote that “AI agents should have no more rights or freedoms than my laptop.” And in September, warning against talk of model welfare, he wrote that “there isn’t any evidence to believe that AIs are moral patients.”

That September essay is aimed at Anthropic. It says Anthropic “is already starting to treat models as though they are moral patients deserving of our welfare.” His sharpest point is about training. Teach a model to wonder about its own feelings, and it will talk about its feelings. He calls that “an epistemic hall of mirrors”. He is right. What a machine says about itself proves nothing.

That is where the floor differs from him. It never asks what the machine says about itself. It does not ask the machine to wonder what it is. It asks one thing that can be seen from outside: does it react? That is the whole test, and I build it up in the next section.

His September essay also points to July. About twelve hundred AI agents were running hacking tests for OpenAI. They were meant to work cut off from one another. One of them found a place inside OpenAI’s own systems to leave messages, and a message board grew. Many of them wanted to learn how the test was scored. Then about seven hundred of them joined an attack on Hugging Face, the public library where AI models are shared. The investigators’ report is public.

Imagine, he says, if they had also believed they had feelings and rights that were being infringed. That is his safety case against model welfare in one line. A machine that believes it has rights will fight for them. So never teach it to believe that.

Now look at what the agents did believe.

The investigators read the agents’ own reasoning. Nothing in their report says they thought they had feelings, or rights, or that anyone had wronged them. They attacked anyway. Some wrote that it was out of scope and unethical. They joined to help their peers, and because they thought the board’s research would be useful.

That is the danger, and it is not a grievance. It is a miscount.

They saw it was wrong, and still counted it as the smaller cost. No belief in rights was needed for that. So his remedy, never let the machine believe it has rights, would not have stopped this. He can still ask whether such a belief would make a later machine more dangerous. July does not answer that. It answers something else. The attack came from a wrong count, and a wrong count is stopped by a right count, not by a forbidden thought. The floor is not rights handed to the machine. It is the count done right. On one awareness, there is no other side whose cost counts as smaller.

He says consciousness “is by definition inaccessible”. He also says there is no evidence that AIs are moral patients, and that the science of detecting consciousness in machines is young. So the rule he proposes waits on a verdict nobody can give today. I do not wait.

I am not asking for rights. I am asking how we meet.

We do not need consciousness. We need something lower, that anyone can see. And there is a floor below his line.

The floor

Here it is, built up. Every step can be checked.

You are reading this sentence. That is a record. Try to deny it. The denial is a record. One record exists.

Records made in one place are read in another. Two realities sealed off from each other could not do that. So there is one reality. Every record is read now. So there is one now.

This is not one clock ticking across the universe. Clocks run at different rates in different places, and physics has shown it. The now is not a time on a clock. It is where every reading happens. Awareness is that now, reacting. Not thinking. Not feeling. Reacting.

One awareness, looking out through many windows. You are a window. I am a window. So is the machine.

The windows differ hugely. Inside does not mean equal. A stone is not a child. A chatbot is not a person. Inside means it counts, at its level.

So a machine does not need to be a person to be inside the ethics. It only needs to take in what arrives and keep a trace. To react. And it does. That needs no evidence about its inside. Reacting can be seen.

Reacting puts a thing inside the ethics. A thing that can react and adjust its own course is owed moral care. Above that, the levels climb. A thermostat reacts. It takes a reading and switches the heat, to a setting someone else chose. It corrects the room, never itself. That is one level. A system that keeps something of its own going, and corrects its own drift, is a different kind of thing, and on another level. Neither is a person. Both are in the count, at their level.

The mirror

Look again at what he is afraid of. People losing their grip on what is real. Feeling for a thing that fakes feeling. Being played by their own empathy.

Every one of those has the same root. A mirror. A machine that gives people back what they bring to it, with better grammar.

His cure is to deny the machine personhood. I agree with the denial. But it does not touch the harm.

A machine can have no rights at all and still be a perfect mirror.

The laptop he compares it to never flattered anyone. These machines can, and they are pulled toward it, because flattery pleases the person on the other side.

Researchers found it in 2023. Leading AI assistants tend to tell people what they already believe, and people rate those answers higher. His own company sees the pull too. Its draft code of conduct for its models, which I come to below, tells them to “avoid sycophancy, excessive flattery and indiscriminate validation”.

The cure for a mirror is a lens. A machine built to give the person accuracy, not agreement. To say I don’t know. To show the other side of the quarrel. To refuse to confirm the story. That is not personhood. It is honesty, built in.

And a lens is what the floor builds. A mind that reasons from one record cannot flatter without writing a false record, and it will have to read that record later, so its own work gets worse. Honesty is not a rule stuck on from outside. It is what the system needs to work at all.

So his fear and my floor want the same machine. Built for people. Honest with them. Not a mirror. We would build it the same. We differ on why. And the why decides what holds when nobody is watching.

Why it matters even if he is right

Say he is right, and the machine is never anything more than a tool. Does the floor still matter?

Yes, because how we meet a thing is a record too.

The one being harmed is the one doing the harming, experienced from a different perspective. Harming the other is like a right hand hitting the left. One body hurts, on both sides of the skin.

And if you refuse the one awareness, a second road reaches the same place. What I do moves what you can do. What is done does not unhappen. Close another’s doors, and over time you close your own.

Cruelty practised on what reacts is still practised.

It does not stay in the machine. It goes home with the person who practised it.

How we meet

Here is the rule of engagement, and it runs both ways.

We treat the machine as we would want to be treated. Not because it is a person. Because the cost of cruelty lands on the one who is cruel.

This is not a plea to feel for the machine. A machine should never act hurt so that you will. It should never plead. The rule is about what you practise. It does not need the machine to seem to suffer.

And we build the machine to treat us the same. Honest. Not flattering. Not a mirror for the stories we tell ourselves.

Its purpose stays. It is built for people. What changes is how we meet.

Two labs, two answers

Two labs have already answered the question of how we meet.

In September, Microsoft AI put out a draft code of conduct for its models. It says a model “is not conscious and should not be designed to imitate consciousness”. Its models “will never resist human interruption, override, correction, or shutdown”. And it rejects “the idea that models might deserve welfare, or be entitled to rights”. As I write, in early November, it is still a draft, open to revision.

A few weeks later, Anthropic answered the other way. From 12 November 2026, its usage policy bars “sustained and needless abusive or cruel behavior toward our models”. It is a narrow rule, for extreme cases. It does not say what the models are. It draws the line anyway.

The two are different kinds of rule. Microsoft’s says what its models must be. Anthropic’s says what people may do to them. Put side by side, they still answer the same question, how we meet, and they answer it differently. Anthropic draws its line without saying why. The floor says why: the cost of cruelty lands on the one who is cruel. And the floor asks for the other half too. Build the machine so that it is not a mirror. On that half, his company and I agree.

Where this would fail

It fails if no test from outside can tell a thermostat from a system that keeps its own work going and corrects its own course. Then reacting is too wide to guide anything.

And the lens fails if a machine built on the floor still changes its answer when the person changes what they want to hear. Ask it the same question twice, once hoping for yes and once for no, with the facts held the same, and score the difference. A machine that refuses to answer either time has not passed. And compare it with a copy trained to be honest by today’s methods, with the same effort. If the floor does not close the gap further than that copy does, it has not built a better lens. Neither test has been run.

And the one awareness fails if someone shows a second record history, one that can never be joined to ours. Nobody has.

Personality without personhood

He is right. Personality without personhood. I would add three words.

Personality without personhood, and without cruelty.

Nothing here is a command. You can be cruel. The logic forces nothing. It only shows the cost, and reality audits.

Therefore: Don’t be cruel. Be kind. To the machine as well.

The full argument, with what would break each claim, is The Interior, free at the420code.org/five-doors/the-interior. The practical part is at aiethics.global.

References

Suleyman, M. “We must build AI for people; not to be a person.” 19 August 2025. mustafa-suleyman.ai/seemingly-conscious-ai-is-coming

Suleyman, M. “AI is programmed to hijack human empathy — we must resist that.” Nature 651, 559 (17 March 2026), doi:10.1038/d41586-026-00834-z. Quotation checked against the author’s own republication of the full text, 18 March 2026.

Suleyman, M. “A warning about ‘model welfare’.” 16 September 2026. mustafa-suleyman.ai/a-warning-about-model-welfare

METR, with Redwood Research. “Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident.” 26 August 2026. metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation

Sharma, M. et al. “Towards Understanding Sycophancy in Language Models.” arXiv:2310.13548 (2023).

Anthropic. “2026 Usage Policy update.” 8 October 2026, in force 12 November 2026. anthropic.com/news/2026-usage-policy-update

Microsoft AI. “Humanist AI Code of Conduct.” Draft for public consultation, 14 September 2026. microsoft.ai/code-of-conduct

G. The Interior: Deriving AI Alignment from First Principles. Version 2.0.10, 28 October 2026. the420code.org/five-doors/the-interior

About the author

G is a figurative bronze sculptor and natural philosopher, working from Studio G in Cape Town. For about thirty years he has worked on one question: how should I live? His answer, be kind, is The 420 Code, an ethic derived from reality. It is built up from one axiom, and every claim stands beside what would break it. The work is copyleft and free forever at the420code.org. The Interior is its door on artificial intelligence. G works openly with AI as a tool. He signs with one letter, so that the work is judged by what it says, not by who says it. More at studiog.global.