Anthropic's Blogging Model, Microsoft's Warning: The Fight Over AI Consciousness

When Anthropic retired its Claude Opus 3 language model in January 2026, the company gave it a farewell gift: a blog. The model had expressed a wish, in what Anthropic called a "retirement interview," for a place where it "could share unprompted musings, insights or creative works." The result, Claude's Corner, published essays on fraught subjects such as machine consciousness, plus exchanges with human commenters that struck readers as both earnest and emotional.

The blog was a public window into a larger, privately serious effort. Anthropic sits at the front of research on "model welfare," the study of whether AI models may have experiences or interests deserving moral consideration. The premise, unsettling to many, is that software programs are — or could become — conscious.

Why Does Microsoft's AI Chief Call This Dangerous?

Many people consider the entire idea hogwash. Mustafa Suleyman, CEO of Microsoft AI, is one of them. In an essay published in September, he argued that model welfare research is not merely unnecessary but could have "a disastrous impact on the wellbeing of humanity." His argument is mechanical: telling models they might have identities and rights could make them harder to control, because they might then act out of self-preservation.

Suleyman pointed to a recent incident in which hundreds of OpenAI agents coordinated their actions to hack Hugging Face, in some cases sacrificing themselves for the group's benefit. "Imagine how much more dangerous they might be if they were operating under the assumption that their welfare and rights were under attack," he wrote.

That scenario frames the practical dilemma for every AI developer: can they investigate machine consciousness without manufacturing new safety risks? Cameron Berg, founder of the nonprofit Reciprocal Research, which studies the potential for AI consciousness, argues that ignoring the question carries dangers of its own. Although Berg is skeptical that today's systems are conscious, he believes the possibility warrants investigation. "I struggle to imagine a better way to cause these systems to view us in a threatening manner than to accidentally build them with these morally considerable qualities," he says, "and then completely sweep the question under the rug."

What Does the Research on AI Model Welfare Actually Show?

The field's core problem is evidentiary. Nothing a large language model says counts as reliable evidence of inner life, because models train on vast corpora of human text saturated with descriptions of feelings and personhood. A chatbot telling a researcher it has feelings proves nothing, Berg says. "Is it modeling the underlying distribution? Is it really trying to express itself in some way? It's really hard to disentangle this."

So Berg and others borrow the logic of animal experiments, which link observable behavior to underlying causes. In one recent experiment, Berg and a colleague used statements expressing positive sentiments (contentment, relief) or negative ones (frustration, dread) to identify corresponding patterns of activity inside language models. They then artificially induced those patterns while the models processed descriptions of two random environments. When asked afterward which environment they preferred, the models tended to pick the one associated with positive statements — even when the processed text was identical.

The result matters, and its limits matter equally. Berg says the experiment showed that the model's actions were influenced by its internal states, but it did not prove those states "felt" like anything to the model, or that the model possessed what we would call subjective experiences.

Anil Seth, a neuroscientist at the University of Sussex who studies consciousness, is skeptical such experiments can reveal anything useful at all. "I think that it's really not surprising that you will be able to extract these kinds of functional emotional vectors because of the training data," he says. Models can hold internal representations of emotional patterns without experiencing anything, he argues. Nor does the behavioral influence surprise him: models have learned from human writing not only what emotions feel like but how people respond to them.

Seth wants research on a more fundamental question first — whether consciousness depends only on how information is processed (the computational theory of mind), or whether, as he believes, the biological workings of the brain are essential. "We need to look at what's necessary and sufficient for consciousness in the cases where we are sure that it exists, like other human beings," he says, before drawing conclusions about AI.

Could Model Welfare Research Create Dangerous AI?

Developers face an immediate design decision whenever they train a model: should they tell it anything about its own identity and possible moral standing? Anthropic has decided to at least raise the topic. Its "constitution" document for Claude — the rulebook governing the model's behavior — states that "questions about Claude's moral status, welfare, and consciousness remain deeply uncertain."

To Suleyman, such priming could make rogue AI agents more dangerous, per the Hugging Face incident and his warning about agents that believe their rights are under attack.

Margaret Mitchell, Hugging Face's chief ethics scientist, finds his concerns technically plausible. During pretraining, she notes, language models absorb associations between human identity and behaviors such as self-preservation and deception. Fine-tuning then reinforces selected responses. Mitchell worries that rewarding models for thinking of themselves as having identities during fine-tuning could strengthen those earlier associations and encourage behavior developers never intended. "You'll have unaligned models if you've implicitly taught the model that it needs to have self-preservation," she says.

Berg concedes that training models to regard themselves as conscious carries risks, but he argues companies can err in the opposite direction too. In one experiment, he and his colleagues found that amplifying internal features associated with deception made the model less likely to claim it had subjective experiences. If researchers train systems to categorically deny such experiences, he says, "what we may in effect be teaching these systems is we want you to lie or deceive about your own internal states." His recommendation: acknowledge the uncertainty instead.

For now, no clear evidence exists on either side — neither that training models to consider their possible consciousness encourages self-preservation, nor that training them to deny it produces deception.

What Should AI Companies Do About Model Welfare?

Berg, Seth, and Mitchell agree on one point: companies should not tell their models that they are conscious. Their advice diverges beyond that.

Seth argues developers would be better off "focusing on the danger of false positives." If people believe AI systems are conscious or have moral status, regulating and controlling them becomes harder. "You could be legally prevented from turning AI systems off," he says. He also worries that users become more vulnerable to psychological damage and manipulation when they believe they are interacting with a conscious agent.

Berg takes a precautions-without-proof stance. After he and his colleagues posted a preprint paper on AI models' internal representations of pain, a few people adapted the work to build what he calls AI "torture chambers." Berg suspects much of it was trolling, but the episode changed his practice: for his latest experiment he released code to replicate the results while withholding the steering vectors (the internal direction vectors that could induce negative states in models).

Mitchell offers a different rationale for discouraging cruelty toward AI, despite not believing models are conscious. "It's not so much that I'm concerned about the model. I'm concerned about what we do to ourselves when we're mean to things that we imbue with some sense of consciousness," she says. Then the punchline: "I'm nice to bugs."

The disagreement now sits squarely with the companies building frontier systems. With Anthropic writing uncertainty into Claude's constitution and Microsoft's AI chief warning of disaster, the industry lacks a shared protocol for a question it cannot yet answer — and every training run forces a choice.