Consciousness in Artificial Intelligence

The problem the report sets itself

Asking "is this AI conscious?" is the wrong question to start with. The science of consciousness has no canonical test even for biological systems; pretending one exists for AI would import all the ambiguity of that scientific landscape and add some.

The 2023 report Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, co-authored by nineteen philosophers, neuroscientists, and AI researchers — including Yoshua Bengio and David Chalmers — sets itself a different question:

What would a system have to look like for any of the major contemporary theories of consciousness to consider it a candidate?

This is a deliberately theory-pluralistic move. Rather than commit to one theory and ask whether AI satisfies it, the authors extract indicator properties from each leading theory and ask whether any current AI satisfies enough of them, across enough theories, to be treated as a serious candidate.

The theories surveyed

The report draws on six families of consciousness theory:

• Recurrent processing theory — consciousness requires recurrent feedback loops in perceptual processing.
• Global workspace theory — consciousness requires broadcast of information to a shared workspace.
• Higher-order theories — consciousness requires representation of one's own mental states.
• Predictive processing — consciousness is associated with generative models predicting incoming signals.
• Attention schema theory — consciousness involves a model of one's own attention.
• Agency and embodiment theories — consciousness is tied to having goals and acting in a world.

From each, the authors extract specific architectural or functional indicator properties that a system would need to exhibit. They are careful to mark these as necessary conditions on the relevant theory, not sufficient conditions for consciousness.

Applying the indicators

The report applies the resulting indicator list to several real systems, including large language models, multimodal models, and reinforcement- learning agents.

The verdict, briefly stated: current systems satisfy some of the indicators on some of the theories. None comes close to satisfying enough indicators on enough theories to be treated as a clear candidate for consciousness. But also — and this is the part that has caused most discussion — none of the indicators are obviously unsatisfiable by near-future AI architectures.

"Our analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators."

That second clause is what gives the report its operational weight. Consciousness in AI, on this analysis, is not blocked by a missing principle. It is blocked by engineering choices that could change.

Why this matters

Three things matter about the report, beyond the verdict itself.

First, it makes the welfare debate tractable. Before the report, "is this AI conscious?" was an unstructured argument. After the report, there is a checklist — not authoritative, but at least specific. Future disagreement can be about whether a system satisfies indicator X, not about whether the question is even well-posed.

Second, it brings the science of consciousness and AI research into the same conversation. Both fields had been operating in mutual neglect. Nineteen co-authors across both fields, signing the same document, is itself an event.

Third, it shifts the burden of design. If indicator-satisfying AI is technically buildable, then the question of whether to build it is no longer "wait for nature to decide." It is a choice. The report stops short of recommending against building such systems, but it makes clear that doing so would be a choice with moral weight.

What the report does not do

It does not resolve any of the underlying theoretical disputes about what consciousness actually is. It does not endorse one theory over others. It does not provide a number for the probability of AI consciousness — and is careful to say it cannot. What it provides is the most concrete framework to date for arguing about AI consciousness without descending into pure intuition exchange.

For Anthropic's Model Welfare research and the Taking AI Welfare Seriously policy proposal, this indicator framework is the methodological backbone. Without it, those efforts would be claims without scaffolding. With it, they have something to point at.