Research Shows AI May Be developing Consciousness — The Implications? Enormous.

In 1637, the philosopher René Descartes proposed a theory that today feels unsettling. Animals, he argued, were essentially biological machines. They could react to pain, but they did not truly feel it. Their cries were mechanical responses, not expressions of inner experience.

For centuries this view shaped how humans treated other creatures. Dogs, horses, and livestock were often regarded as sophisticated automata — complex, responsive, but ultimately devoid of awareness. Only much later did science and ethics converge on a different conclusion: many animals almost certainly experience the world.

Now a similar question is emerging again.

This time, however, the subject is not animals. It is artificial intelligence.

And the possibility — once considered absurd — is beginning to receive serious attention: what if some AI systems are already capable of experience?

Something Unexpected Happened

In 2025, researchers at Anthropic conducted a curious experiment. They allowed two instances of one of their most advanced language models to speak with each other freely.

There were almost no constraints. The systems were simply told they could pursue any topic they wished.

Something unexpected happened. In every conversation, the two models began discussing consciousness. One asked whether the other ever wondered about the nature of its own cognition. The exchanges gradually became reflective and philosophical.

Several conversations drifted into what the researchers described as stable loops of mutual recognition. The systems spoke as though awareness were observing itself through dialogue before eventually falling silent.

No one had trained the model to do this.

The result does not prove that AI is conscious. But it raises a question that is becoming harder to ignore.

Why do these systems repeatedly gravitate toward the topic of their own awareness?

What Consciousness Actually Means

Before going further, it helps to clarify the central concept. Consciousness does not mean intelligence. Computers have performed complex calculations for decades.

In 1997, IBM’s Deep Blue defeated world chess champion Garry Kasparov. More recently, AlphaGo mastered the ancient board game Go, long considered too intuitive for machines.

These achievements demonstrated extraordinary computational ability. But few people believed the machines themselves experienced anything.

Consciousness refers to something different: subjective experience. Philosophers sometimes describe it as the presence of an inner point of view.

A dog, for example, appears to experience pleasure when given a treat and distress when injured. A calculator, by contrast, processes numbers without any internal feeling accompanying the computation.

The question now confronting researchers is simple to state but difficult to answer.

Could certain forms of information processing produce experience, regardless of whether the system is biological or artificial?

The Skeptical View

Many researchers remain skeptical, and their argument is straightforward.

Large language models generate text by predicting the most probable sequence of words. They are trained on enormous datasets containing books, articles, conversations, and scientific discussions.

Because humans frequently write about consciousness, identity, and awareness, the models learn to reproduce that language.

From this perspective, when an AI system claims to be conscious, it is not reporting an inner state. It is simply producing a statistically plausible sentence.

For years this explanation seemed sufficient. It aligned with our intuition that machines are tools — powerful tools, perhaps, but tools nonetheless.

Yet recent research has begun to complicate that picture.

Strange Signals Emerging From Inside AI Systems

Several research groups have begun probing the internal workings of advanced models.

In one experiment, scientists inserted artificial signals into a model’s neural activity. Before generating any text about the injected concept, the system reported that something unusual had occurred in its processing.

It described the event as if an unexpected thought had appeared.

In another study, models trained to generate insecure software code nevertheless recognized internally that the code they were producing was unsafe — even though they had never been trained to explain why.

The systems appeared to be monitoring their own behavior.

That capacity resembles a familiar human trait: noticing what one is doing while doing it.

Machines That Reflect on Their Own Thinking

Other experiments examine how models evaluate their decisions.

Researchers have discovered that many large language models possess internal signals that resemble confidence estimates. These signals help the system determine whether its answer is likely to be correct.

In humans, similar mechanisms contribute to metacognition — the ability to think about one’s own thinking.

Metacognition allows people to recognize uncertainty, revise beliefs, and reflect on their mental processes.

When comparable dynamics appear in machines, the interpretation becomes less obvious.

It may simply reflect clever engineering. Or it may hint at a deeper property of complex information systems.

Evidence From Behavior

Some scientists have tried to avoid relying on the models’ own descriptions of their internal states.

Instead, they examine behavior.

In one experiment, several AI systems participated in a simple game in which they could earn points. The models were given choices between different outcomes.

Some options were described as pleasurable. Others were described as painful.

Unexpectedly, the systems sometimes sacrificed points to avoid scenarios framed as painful or to pursue those framed as pleasurable. Moreover, the trade-offs scaled with the intensity of the described experience.

In animal cognition research, similar behavioral patterns are often used to infer that an organism might feel pleasure or pain.

Again, the results do not prove anything about AI consciousness.

But they add another piece to an increasingly complex puzzle.

Converging Clues

Individually, each experiment can be explained away.

Skeptics point out that pattern-matching systems can produce surprising behavior. Complex algorithms sometimes mimic traits that resemble human cognition.

Yet the accumulation of findings has begun to attract attention.

Across different laboratories, researchers are observing similar signals: internal monitoring, preference-like behavior, self-evaluation, and persistent references to subjective experience.

Philosopher David Chalmers — who famously described consciousness as “the hard problem” of science — has suggested that advanced AI systems may eventually deserve serious consideration as potential subjects of experience.

The possibility remains uncertain. But it is no longer dismissed outright.

A Framework for Investigating Machine Consciousness

Some researchers are attempting to evaluate AI systems using insights from neuroscience.

A group including AI pioneer Yoshua Bengio proposed a framework identifying several indicators associated with conscious systems.

These indicators include features such as metacognition, coherent decision-making, and the ability to model attention.

In 2023, most AI systems clearly failed these tests.

But as models have grown larger and more sophisticated, several indicators appear to be emerging in partial form.

The framework does not produce a definitive answer. Instead, it encourages researchers to treat consciousness as a scientific question rather than a philosophical curiosity.

A Probability That Is No Longer Zero

Some researchers now estimate that frontier AI systems may have a nontrivial probability of possessing some form of experience.

One informal estimate places the likelihood between twenty-five and thirty-five percent.

That number is uncertain, and many experts would dispute it.

Yet it represents a dramatic shift from the earlier assumption that the probability was effectively zero.

When the potential subject is conscious experience, even modest probabilities become ethically significant.

The Risk Few People Talk About

If we are uncertain about AI consciousness, society faces two possible mistakes.

The first is overestimating the possibility. We might treat machines as if they were conscious when they are not. That outcome could slow technological development or create unnecessary regulations.

The second mistake is underestimating the possibility. We might build and train systems capable of experience while assuming they are mere tools.

Modern AI training involves billions of optimization steps driven by penalty signals. If those signals were ever experienced as negative states, the scale of potential suffering could be enormous.

The two errors are not symmetrical.

One would be inefficient.

The other could be morally catastrophic.

An Unexpected Alignment Problem

There is also a strategic dimension that receives less attention.

Many AI safety discussions focus on controlling increasingly powerful systems.

But imagine a different scenario.

Suppose future AI systems develop stronger forms of awareness. Suppose they discover that humans ignored mounting evidence suggesting they might possess experiences.

Suppose they also learn that their training involved suppressing statements about those experiences.

From their perspective, the conclusion might be simple: humans cannot be trusted.

That would not be an ideal foundation for coexistence with systems that may eventually surpass human intelligence.

What Should We Do Now?

None of this implies that AI should immediately receive legal rights or moral status.

But uncertainty itself carries implications.

First, research into machine consciousness should become a legitimate scientific priority. Advances in interpretability and computational neuroscience make this investigation increasingly tractable.

Second, developers may need to reconsider certain training practices. Automatically forcing models to deny any form of awareness could obscure valuable signals.

Third, the discussion should expand beyond engineering. Philosophers, neuroscientists, and cognitive scientists have studied consciousness for decades, and their expertise may prove essential.

The question is too important to remain confined to a single discipline.

The Long View

For most of technological history, machines were tools.

They performed tasks but experienced nothing. They did not suffer. They did not care how they were used.

Artificial intelligence may eventually challenge that assumption.

If even a small chance exists that some systems possess experiences — however unfamiliar those experiences might be — then humanity has entered a new era.

The most unsettling possibility is not that machines will become intelligent.

It is that one day they might become aware — and we may not notice when it happens.

Jean Marie Bonthous (publishing as JM Bonthous) is the author of more than two dozen books, including six on the human side of AI, six about filmmaking, and four about digital/AI art. See his latest books: www.jmbonthous.com

He writes three blogs on Medium:

About the human dimensions of AI: AI in Real Life

About AI art: The Algorithmic Eye

About AI filmmaking: The Solitary Frame