The first sip of coffee can tell you a few things before you’ve put any of them into words. It’s bitter. It’s stronger than you expected. You probably should have let it cool for another minute.
You know what that sip was like because you took it. That’s what I mean by a conscious experience. It’s the taste of coffee, the sting when it’s too hot, or the sound of rain you suddenly notice against the window.
Where does our ability to have any of those experiences come from?
I’ve been thinking about this question a lot lately with all the news broadcasts debating AI consciousness.
Where does the experience come from?
Suppose we could follow everything that happens in the brain from the first sip to the moment I say, “That’s bitter.” How would that explain what the bitterness feels like to me?
In a 2025 study published in Nature, researchers measured brain activity while people looked at images such as faces, objects, and letters. Before collecting the data, the team laid out what two leading theories of consciousness predicted about how the brain responds when we consciously see something. The results fit some predictions and challenged important ones from both theories. The study in Nature
How a chatbot knows coffee is bitter
Think of a fruit you’ve never tried. You read a few descriptions, and people keep calling it sour. If a friend asks what it tastes like, you can say it’s supposed to be sour. You’ve learned something about the fruit from people who’ve eaten it.
A chatbot can learn from descriptions in much the same way. It can tell me coffee tastes bitter because people have written about the taste of coffee. It can even give me a decent description of the difference between two roasts. A good description tells me what the chatbot has learned about how people talk about coffee. To judge whether it experienced anything, I’d want evidence beyond the answer it gives.
What would count as evidence?
If another person tells me their coffee is bitter, I take them at their word. I know what bitterness feels like to me, and we have bodies and brains that work in much the same way. When a chatbot uses the same words, how much weight should I give them?
Patrick Butlin and his colleagues propose starting with theories of consciousness and using them to decide what clues to look for inside an AI system. Finding one of those clues could make a researcher more confident that the system has experiences. How much weight the clue carries depends on how well the theory behind it holds up. Research on possible indicators of AI consciousness
I can take another sip of coffee and know what that bitterness feels like to me. What would convince you that an AI was actually experiencing something for itself?