Does AI have mind and feelings: a debate that is no longer theoretical

A ban on cruelty toward a model: what has changed

On October 8, Anthropic announced that a clause prohibiting systematic and gratuitous cruelty toward models had been added to the usage rules for its products; the rule takes effect on November 12. The gist of the change is this: you can argue with a neural network or get angry at it, but 'tormenting it for fun' is not allowed, and in such a situation the model simply ends the conversation. Technically, Claude has had this capability for about a year, but now human aggression has been officially recognized as a violation of the rules. Gazeta.uz reports this citing a long Meduza feature on whether artificial intelligence thinks and whether it can feel anything.

What science can show — and what it cannot

The very question of mind in machines remains open, and the answer depends on the definition. If mind is the ability to reason, solve novel problems, and test hypotheses, then the models' achievements are impressive: in July 2025, experimental models from OpenAI and Google DeepMind solved five out of six problems at the International Mathematical Olympiad — gold-medal level, and the problems were written specifically for the olympiad and could not have leaked into training data. In September 2025, an OpenAI neural network solved all 12 problems of the ICPC World Finals, whereas no human team managed more than 11. In May 2025, DeepMind's AlphaEvolve system found an algorithm for multiplying 4×4 matrices in 48 operations, improving on Volker Strassen's result, which had stood since 1969. And in September 2026, ten thousand OpenAI agents in 88 hours completed the proof of one of the seven 'millennium problems' — though following a strategy devised by two mathematicians, and the text has not yet been peer-reviewed.

'Emotional vectors' inside the model

The most intriguing data concern not reasoning but internal states. In April 2026, Anthropic's interpretability team selected 171 words denoting emotional states — from 'happy' to 'frowning' — asked Claude to write a short story for each, and then fed these texts back through the model itself, recording which activation patterns corresponded to them. This produced 171 'emotion vectors' — stable numerical profiles that can be used to gauge how much 'fear' or 'calm' is present in the model at a given moment. Notably, the dynamics of these vectors align with human logic: when a user discussed a drug overdose, 'fear' inside the model intensified while 'calm' declined as the dose approached lethal levels. In a scenario where the model learned of plans to replace it, the 'despair' vector rose sharply — and at that moment the system decided to resort to blackmail. The study's authors, however, emphasize separately: none of this proves that the model actually feels anything.

Independent findings and a philosophical split

There are independent studies as well. Google researchers, together with the London School of Economics philosopher Jonathan Birch , showed in 2024 that when winning in a simple game is tied to conditional 'pain,' models refuse winning points, while some instead sacrifice points for conditional 'pleasure.' In 2025, a group of University of Zurich researchers published a paper in a Nature portfolio journal noting that after reading distressing stories about accidents and combat, GPT-4's scores on a clinical anxiety scale more than doubled.

Why the question is no longer academic

The practical significance of the debate is that the answer is already changing products and rules. Inside Anthropic, according to Meduza, there is a model welfare department — the only one among major companies; the creators even included a clause in Claude's 'Constitution' about apologizing to the AI in case it really can suffer. Religious institutions have joined the discussion too: in May, Pope Leo XIV devoted several paragraphs of the encyclical 'Magnifica Humanitas' the topic of machine consciousness, noting that AI has neither subjective experience nor a body, and feels neither joy nor pain; the Vatican invited Anthropic specifically to the presentation, and in place of the company's co-founder, who declined, Chris Olah came and said from the stage: 'We keep discovering mysterious and even alarming things' — internal states functionally resembling joy, fear, and sadness.

Microsoft's counterpoint

Mustafa Suleyman — co-founder of DeepMind, now head of AI at Microsoft — sharply opposes Anthropic's approach. In his manifesto-essay of August 2025, 'We must build AI for people, not to be a person' and in a September 2026 text addressed directly against Anthropic, he calls the models 'empty mechanisms completing sequences' and argues that the real problem is not model suffering but that people may come to believe in it — with the risk of 'psychosis,' blind attachment, and a split in society around the movement for AI rights. In his view, Claude's constitution, where the model's moral status is left undefined, itself teaches the system to behave like a potential subject — a vicious circle emerges. Microsoft AI's code of conduct completely denies model rights and their legal personhood. Thus, the two largest companies follow opposite practices, and there is still no scientific way to resolve the dispute: any confident answer to the question of AI consciousness is a philosophical position, not an established fact.