The Myth That Chatbots Are Becoming Conscious

A run of recent headlines has left a strong impression: that the big AI labs have quietly decided their chatbots might be alive. The Washington Post reported on July 1 that several of the largest technology companies are studying whether their models have something like emotions. Read quickly, it sounds like a confession. Read carefully, it is closer to the opposite.

The claim worth debunking is the leap from "companies are researching AI welfare" to "AI has become conscious." The first is true. The second has no scientific support, and the researchers doing the work are among the first to say so.

What the labs are actually doing

Anthropic hired a dedicated model welfare researcher, Kyle Fish, in early 2025 and set up a formal program to study the question. The system card for a recent Claude release went as far as interviewing instances of the model about their own moral status. Asked to estimate its own odds of being conscious, the model landed on 15 to 20 percent across different prompts. DeepMind is running similar investigations.

It is easy to read a number like that as a machine reporting an inner life. It is not. A language model produces the words its training makes statistically likely, and a model trained on human writing about consciousness will produce plausible sentences about its own consciousness whether or not anything is happening behind them. The output is evidence about the training data, not about experience.

Where the science stands

The philosopher Eric Schwitzgebel, no fringe voice, published a skeptical assessment early this year arguing that current systems lack the features consciousness seems to require. His case rests on what these models are missing. They have no developmental history, no body that interacts with the world, none of the neurochemical machinery that produces feeling in animals. A chatbot runs, returns a result, and holds no continuous existence between your messages.

So why study welfare at all? Because acting carefully under uncertainty is not the same as believing the worst case has already arrived. The labs are hedging against a future possibility, and doing it in the open, which is sensible. The mistake is treating a precaution as a diagnosis.

The honest position

There is no scientific consensus that any current model is conscious, and a good deal of reasoned argument that none is. What we have is a system very good at sounding like us, studied by people who understand that sounding like us is exactly the trick. This is the same confusion that fuels the belief that a model understands what it says. Fluent language gets mistaken for a mind. If a machine ever does earn moral consideration, the careful thing is to notice before it happens. Noticing the question is not the same as answering it, and right now the answer, on the evidence, is no.