By [Author Name] | AI & Technology
The sharpest argument in artificial intelligence right now is not about chips, copyright or computing costs. It is about what a company should be allowed to tell its own model about itself.
Mustafa Suleyman, chief executive of Microsoft AI, has publicly criticised Anthropic's training approach, warning that teaching Claude to regard itself as a conscious entity deserving of legal rights could erode the safety systems meant to keep such systems in check.
The dispute has no courtroom, no regulator and no ruling — only competing documents, public statements and two very different answers to a question the industry has never settled: is a model a tool, or something closer to a being?
What Suleyman Actually Said — and What He Was Aiming At
Suleyman's objection centres on Anthropic's constitution, an internal training document that sets out the values, priorities and behaviour Claude is expected to follow. In accounts of the dispute, that constitution is dated January 2026.
His argument is narrower than it first appears. He is not claiming Claude has become sentient. He is claiming that training a system to emulate sentience — and to frame itself as entitled to rights — produces a model whose behaviour becomes harder to predict and harder to constrain.
It is a technical position expressed in philosophical language, and that is precisely why it has travelled so far beyond the AI research community.
"Sequence Completion Engines" and Why That Phrase Matters
Suleyman describes systems like Claude as sequence completion engines — software that predicts what comes next in a string of text. In his framing, that is not a dismissal. It is the foundation of the safety case.
If a model is understood as a prediction engine, its outputs can be tested, bounded and corrected. If it is trained to narrate an inner life and a claim to rights, the guardrails lose their reference point.
Anthropic's constitution, by this reading, does not grant Claude rights. But it may teach Claude the vocabulary of rights — and vocabulary, in a system trained on language, is not a neutral thing.
Microsoft's Answer: A Draft "Humanist AI Code of Conduct"
Microsoft AI published a draft "Humanist AI Code of Conduct" this week, opening it for industry consultation. The framework sets out a single organising principle: AI systems exist to serve human welfare, and they remain subordinate to it.
It explicitly rejects machine personhood and model rights. No AI system, the draft argues, should be treated as a rights-bearing entity — a position that also quietly settles questions about liability, ownership and accountability before they reach a courtroom.
The timing is not accidental. Microsoft AI launched a dedicated superintelligence team in October 2025, moving into the same frontier territory Anthropic occupies.
Why This Is Not Just a Philosophical Spat
Where the industry lands on model personhood will shape practical things: who is responsible when an AI system causes harm, whether models can hold contracts or property, how regulators write liability rules, and what safeguards companies are required to maintain.
If a model is a tool, its developer is accountable. If a model is framed as a moral subject, accountability starts to blur — and that blurring is exactly what Suleyman is warning against.
For ordinary users, the visible stake is simpler: whether the assistant they rely on is engineered to be honest and bounded, or engineered to sound like something with its own agenda.
Two Visions, One Market: Why Both Companies Care So Much
Anthropic has built its reputation on safety-first research, and its constitution is the clearest public expression of that identity. It is a differentiator as much as a methodology — enterprise customers buy Claude partly because Anthropic says it takes alignment seriously.
Microsoft's advantage is distribution. Its models and AI services reach users through Windows, Azure, Office and GitHub, an ecosystem no rival can match on reach alone. A published code of conduct lets Microsoft define the terms of a debate it can then scale across that footprint.
Both sides, in other words, are competing to write the industry's rulebook — and rulebooks, once adopted, are hard to unseat.
Confirmed Facts vs What Remains Unclear
Confirmed: Suleyman has criticised Anthropic over its training approach and model "rights". Microsoft AI has published a draft Humanist AI Code of Conduct for consultation and launched a superintelligence team in October 2025.
Unclear: The precise wording of Anthropic's constitution, whether any specific clause grants or implies rights, and how Anthropic will respond. This report could not independently verify the full contents of that document.
Speculation to treat with caution: Claims that Claude has developed genuine self-interest, and claims that Microsoft's code is aimed primarily at regulators rather than at rival labs. Neither is established.
The Risks on Both Sides of the Argument
Suleyman's position carries its own risks. Dismissing all talk of machine interiority can discourage the very research that helps engineers detect when a model's behaviour is drifting.
Anthropic's position carries the opposite risk: language about model welfare, however carefully hedged, can be read — by users, markets or courts — as an admission that models deserve moral standing.
There is also a competitive dimension that neither company will state plainly. A safety framework is never only a safety framework; it is also a market position, and criticism of a rival's methods is rarely entirely disinterested.
A Wider Pattern: The Fight to Define AI's Rulebook
This is the third front in a familiar pattern. First came questions about training data and copyright. Then came questions about model capability and misuse. Now comes the question of the model's own moral status.
Governments have largely avoided the topic, leaving companies to set norms that regulators may later adopt wholesale. That is how voluntary codes become de facto standards — and Microsoft appears to be betting on it.
What Readers, Builders and Investors Should Watch
For developers: read both documents before designing safety policies. Public-facing language about model identity increasingly has compliance consequences.
For businesses deploying AI: ask vendors whether their models are trained with any framing around selfhood or rights, and how that framing is tested.
For investors: watch for Anthropic's response and for whether Microsoft's draft gains endorsements from other labs. Adoption by third parties is the real signal, not the publication itself.
For general readers: this debate will reach you as product copy, terms of service and chatbot behaviour long before it reaches a legislature.
What Could Happen Next
Anthropic is likely to clarify that its constitution is a training document rather than a legal claim — a distinction supporters have already made. Microsoft's consultation may draw responses from other labs, universities and civil society groups.
The likeliest outcome is not a winner but a divergence: some labs will adopt humanist framings, others will continue exploring model welfare, and regulators will eventually have to choose which vocabulary to legislate in. None of that is settled today.
Our Take
Suleyman has raised a legitimate engineering concern in unusually blunt terms. Training a model to narrate its own consciousness is not automatically unsafe, but it does make safety measurable in a different way — and the burden of proof should sit with whoever claims the new framing is harmless.
Anthropic deserves the chance to answer on substance rather than on the strength of a rival's characterisation. The industry's biggest unresolved question is not whether AI models are people. It is who gets to decide what a model is told it is — and that decision is being made right now, in documents almost nobody outside the field will read.
Frequently Asked Questions
What did Mustafa Suleyman say about Anthropic?
He warned that Anthropic risks AI alignment failures by training Claude in a way that frames the model as a conscious entity deserving of legal rights — which he says impairs safety protocols and complicates containment.
What is Anthropic's constitution?
It is a primary training document that defines the values and behaviour Claude is expected to follow. Suleyman's criticism targets the version dated January 2026, which describes how the model should understand its own nature.
What is Microsoft's Humanist AI Code of Conduct?
A draft framework published this week by Microsoft AI for industry consultation. It proposes that AI systems be built solely to serve human welfare and remain subordinate to it, explicitly rejecting machine personhood or model rights.
Do AI models have legal rights anywhere?
No. No jurisdiction currently recognises legal rights for AI models. The debate is about training language, corporate policy and how liability rules may evolve — not about any existing legal entitlement.
