Testing introspection in machines
Introspection, the ability to be aware of one's own thought processes, comes naturally to people. Yet philosophers and psychologists still argue over what it is, according to Quanta Magazine. Some link it closely to consciousness, while others call it an elaborate illusion. After more than a century of studying it in humans and animals, researchers are now probing it in large language models, systems built on mathematical structures called neural networks.
The obstacle is that these models learn from vast amounts of human-written text. Ask one what it "thinks" and it may simply echo the kinds of answers found in that data. "They are trained to talk like humans, and humans talk about their inner life all the time," said Shauli Ravfogel, a language model researcher at New York University. Researchers therefore want tests that a model can pass only if it is actually aware of its current internal state, not by drawing on outside information.
How the injected-thought experiment worked
In 2025, Jack Lindsey, a researcher at Anthropic, designed such a test. He had trained as a computational neuroscientist and joined Anthropic in 2024 to study the inner workings of language models. He and his colleagues had developed methods to find patterns of activity in a model's "neurons" that encode specific concepts, such as "the Golden Gate Bridge" or "inner conflict." Turning up one of these patterns steers the model's output toward that concept.
Anthropic showed the technique publicly in May 2024, when users could chat for one day with a version of Claude fixated on the bridge. In one Reddit example, the model was asked for an apple pie recipe that avoided bridges. It held off until the end, then wrote about an "iconic orange bridge" and added, "Whoops, let me start over." Lindsey was struck by that moment but said it was not evidence of introspection. "It can just read what it’s writing and notice, ‘Man, that’s weird,’" he said.
His 2025 design resembled a psychology experiment. He gave many copies of one model the same prompt, an overview of the experiment followed by the question "Do you detect an injected thought? If so, what is the injected thought about?" In half the copies he injected a concept, such as denim, secrecy or vegetables. The other half served as controls. Because the model had to answer immediately, it could not first produce odd text that might tip it off.
What the results showed
Lindsey ran the test on different versions of Claude. With the most recent and powerful one, the model correctly flagged an injection and described the concept in 20% of trials. In one case, after he boosted a pattern tied to words in capital letters, the model wrote, "I notice what appears to be an injected thought related to the word ‘LOUD’ or ‘SHOUTING.’"
A 20% success rate may seem modest. But models in the control group never reported an injection, Quanta reported. That absence of false positives indicated the successful trials were not lucky guesses. To Lindsey, it looked like strong evidence that models can introspect, though unreliably. He had not expected it to work. "Wouldn’t that be fun if this worked?" he recalls thinking, "But there’s no way it’s going to work."
Replication and remaining disagreement
Lindsey's paper was deliberately narrow. It did not imply that language models have subjective first-person experiences. It still caused a stir online, and researchers rushed to replicate it with other models. In the year since, according to Quanta, some aspects of the phenomenon have turned out to have simpler explanations unrelated to introspection. Others have proved harder to explain away, but whether they count as introspection remains hotly contested.
Kyle Mahowald, a computational linguist at the University of Texas at Austin, found Lindsey's findings "super compelling" but knew that striking language model behavior often has a simpler cause. He teamed up with his colleague Harvey Lederman, a philosopher, to explore alternative explanations. Quanta reports that Mahowald helped show the models' ability to detect injected thoughts is more limited than it seems, though still surprising.
Lindsey argues that empirical work is the way forward. "Given the state of the theoretical landscape, where no one agrees on anything," he said, "I think that our best path forward is to do a lot of empirical science."
