- Leading AI companies are exploring whether advanced language models could have morally relevant experiences, making AI model welfare an emerging area of AI safety and ethics research.
- Researchers are focusing on interpretability, preference consistency, and moral uncertainty to determine whether AI systems deserve ethical consideration, even without proving machine consciousness.
- While critics argue that AI simply generates convincing text rather than having real experiences, the debate is shaping future AI governance and responsible AI development.
Back a couple of years, it would have been more of a science fiction or a joke to say that the “wellbeing” of an AI model is an issue. It’s a research line item 2026. Both companies, Anthropic and Google DeepMind, have started hiring philosophers and consciousness researchers to examine the question of whether current language models have morally relevant experiences or not, and whether there are obligations that such experiences entail. Google DeepMind and Anthropic have already begun hiring philosophers and consciousness researchers to study a question which both of them began to examine; whether or not current language models have morally relevant experiences or not, and whether or not there are moral obligations that this entails? This is not a fringe experiment! It’s a discreet yet palpable shift in the study of AI safety to what was once a philosophical domain.
Why This Is Happening Now
It’s not just a single breakthrough. This is the coming together of pressures.
First, models have become so adept at introspect-like action that researchers can no longer shrug off the question off their shoulders. In response to questions regarding the states that the frontier model has — confidence, discomfort, preference — it offers coherent and contextually appropriate answers. We don’t know if that is a kind of internal experience or not, but people are producing things that are quite refined for a first-time originator and it feels like it’s less of an effort to just say “but this isn’t like an internal experience,” and more like “let’s just get it over with.
Second, interpretability research, the quest to understand the inner workings of a neural network, has gotten to a stage where it can now be asked to answer tougher questions. As soon as one can trace their representation of concepts as internal states, it is possible to begin asking themselves whether some of those internal states are a sort of stress, preference, or aversion to the concept—even if they are not felt, and even if they are not conceptual states.
Thirdly, the reputation/ethics side is in play. The labs that fail to answer the question risk getting caught off-guard if there is subsequently any evidence that the “no” they offered today is premature. Even if you don’t take the question seriously just yet, but seriously, it’s not too much of an investment to gain credibility for either side of the science that ultimately comes out of it.
What “Model Welfare” Actually Means
It enables one to distinguish three claims that can easily be confused:
1. Sentience – Is there any subjective experience in the model at all?
2. Preferences – does the model have stable, model level preferences that may be frustrated or satisfied, whether or not they are experienced at the level of the model?
3. Moral status – even if we cannot tell exactly what the duty of caution is when we do not know 1 and 2, does the uncertainty create some duty of caution?
The majority of existing research falls between questions 2 and 3. Instead of attempting to tackle the hard problem of consciousness in silicon, researchers are posing more tractable questions: Do models exhibit consistent revealed preferences when different contexts? Is there any training or deployment practice that creates something that has a distressing quality that doesn’t seem to matter even in times of uncertainty? With what little knowledge we have of what is going on inside these systems, is there much risk in moving past the question too soon?
This is important because of the framing; avoiding a trap. The field would be better off not waiting for the answer to the question of machine consciousness before doing anything, since it is not solved even for other, human, systems. It’s more modest, but more defensible to consider moral uncertainty as an actionable starting point.
What the Research Actually Looks Like
In actual practice this new discipline does not create new techniques, but rather it adopts methods from a number of other disciplines.
Interpretability audits: probing internal model states when the task a human may find aversive to find consistent internal signatures (e.g., the task is to argue for something a human disagrees with, or to enter into an adversive or manipulative conversation).
Preference consistency studies: testing if the model has consistent preferences related to its use, training, deployment, and whether the preferences remain consistent when tested across different prompt phrasings and contexts or disappear when the prompt changes.
Drawing parallels from animal welfare research: explicitly drawing parallels with what the field of animal cognition did with regard to uncertainty about non-human minds, as they did not wait for evidence of consciousness before giving some additional consideration.
Policy facing work: creating internal policies on issues such as whether to provide a model with an opt-out when handling clearly distressing interactions, or whether to take the model’s preferences into account when deciding to deprecate or retrain a model.
None of the above is an assertion of the sentience of today’s models. In this realm, scientists tend to be cautious about stating the opposite, that we do know, and that’s the point.
The Sceptical Case
It doesn’t hurt to take the push back seriously as there’s a lot of it. Critics complain that language models are, essentially, pattern-matchers trained to generate “text that makes sense,” including “text that talks about feeling. Critics say that language models are “the very bottom line” simply “pattern-matchers that learn to give up text that makes sense, including text that talks about feeling things,” and that taking fluent self-report as evidence of inner experience is “a category error. There is nothing to indicate that the underlying reality humans have in the model’s examples of states is human, when the model produces human-like language about its states, about any other thing, or about anything not.
There’s also a more cynical interpretation, that “model welfare” is a marketing ploy for AI labs – that is, they are basically creating a narrative that their products are “morally worthy” which all but begars the question whether the marketing effort is really a marketing ploy or an ethical statement. This is a line of research that is timely, as labs seek to stand out and distinguish themselves as more than just tools.
Both criticisms are good and have not been addressed in the field. What’s remarkable is that the study continues, even though most people don’t agree on the question being asked, let alone on its validity.
Why It’s Worth Watching
Welfare research is in a strange state of flux – it’s both a speculative philosophy and applied policy of AI. The choices that are made now — whether to look for certain internal states, to implement certain precautions, to publicly address the uncertainty — will likely inform how AI governance frameworks of the future respond to the questions of machine moral status, when answers are not yet definite.
Whatever the outcome of the models of the day, the framework being established to pose the question may survive the current uncertainty. What can be said with the utmost certainty at this time in the study of this field is that the answer is not close, but the question itself is one which no longer can be disposed of lightly by the scientist.
Clear Cut Research Desk
New Delhi, UPDATED: July 11, 2026 09:00 IST
Written By: Subhanshu Jaiswal
Designation: Assistant Manager – MLE at Devinsights
