Does AI understand what it says?

You type a complaint about a refund that never arrived, and the chatbot answers: I understand how frustrating that must be. It reads as kind. It comes at just the right moment, too. So it's natural to wonder whether anyone is there. Does the artificial intelligence (AI) understand what it just said? There's no settled answer yet. What's been found inside these models so far is more interesting than a plain yes or no.¹³

A chatbot learned to write by studying an enormous amount of human writing, and people write a lot of sympathy. Apologies and reassurance. The line you send when a friend has had a bad day. So when a message sounds upset, a sympathetic reply is a likely next step, and the chatbot produces one. The words come from people, which is why they sound like a person.

Some people call chatbots fancy autocomplete. That undersells them. In 2025, researchers at Anthropic, the company behind Claude, published a study that traced what happens inside their model while it works. They found concepts that don't belong to any one language. Asked for the opposite of "small," the model used the same inner features for smallness and for opposites in English, French, and Chinese. They also caught it planning ahead. Asked for a rhyming couplet, it chose the rhyming word first, then wrote a line that would land on it.¹

None of that shows feelings, or a life behind the words. A chatbot has never waited for a refund or sat on hold. Whether any AI could have experiences is an open question. Anthropic itself says there's no scientific consensus on whether current or future AI systems could be conscious. So "it feels for you" goes beyond the evidence. So does "it understands you the way a friend does."³

The same study found something odd. Asked to add 36 and 59, the model reached 95 by running two paths at once: one made a rough estimate, and one worked out the last digit. Then the researchers asked how it got there. It described carrying the one, the method people learn at school. That account sounds right. It doesn't match what happened inside. The model had learned how people explain math, and it gave that explanation. So when a chatbot describes its own thinking, treat it as one more piece of text to weigh.¹

That leaves two mistakes, and they pull in opposite directions.

• Treating it as a person can lead you to trust it too much, or to share things you'd only tell a friend.
• Treating it as a dumb parrot means missing what it does well, like drafting, explaining, and translating.

The evidence sits somewhere between the two.²

Think of a chatbot as a very strong pattern machine. It can be a real help. Its kind words can still be welcome. Keep your own judgment for the decisions, and keep private things private. Next time it says it understands, notice which part of the reply is useful, and which part is the sympathy line.