Every clinician has met the patient who says, "I don't feel right." No localizing symptom. No fever. Just a conviction that the machinery is off. We have learned to take that sentence seriously, and we have also learned that the explanation the patient offers next is often wrong.
That pairing (a reliable alarm attached to an unreliable narrator) turns out to be the best summary of where AI introspection research has landed. Ben Brubaker's piece in Quanta walks through it fairly. I went back to the papers.
In 2025 Jack Lindsey at Anthropic planted artificial "thoughts" inside Claude by adding the internal activity pattern for a concept (bread, say, or all caps) directly into the model's processing, then asked whether it noticed anything. The best model noticed and correctly named the concept about 20 percent of the time, and never claimed an injection when none occurred.
Skeptics at UT Austin then split the trick in two. Detection held up. Naming did not. Models often sensed that something had been inserted and then guessed a common, concrete word like "apple." Lindsey's own team later found that the two abilities run on largely separate internal circuits. The field now calls the robust part anomaly detection, and argues about whether that deserves the word introspection.
For the technical reader. Concept injection is activation steering. You record a model's residual stream activations for a concept, add that vector back in at a chosen layer and strength, and ask the model whether it detects an injected thought. Lederman and Mahowald replicated this on large open weights models and showed detection is content agnostic, and that it needs fewer tokens than identification. Macar, Lindsey and colleagues traced a two stage circuit. Early "evidence carrier" features respond to perturbations in many directions and suppress "gate" features that otherwise produce a default "no." Naming relies on later layers with weak overlap. The circuit is absent in base models, appears after preference optimization (DPO) but not supervised finetuning alone, and is underelicited (ablating refusal directions raised detection 53 percent).
For the technical but nonclinical reader. Think of a smoke detector. It reliably knows there is smoke. It has no idea whether the toast is burning or the curtains are. The models have something like a smoke detector for their own processing. Ask them what is on fire and they make a confident, plausible guess, which is frequently wrong. Useful property, but you would not want it writing the incident report.
For everyone else. Researchers can reach inside an AI and nudge its "thinking" toward an idea. Sometimes the AI says, "Something feels unusual." That part is real and repeatable. When it tries to say what feels unusual, it often makes something up. Nobody serious is claiming this means the AI is conscious, including the people who ran the experiments.
One more thing to think about. Neuronpedia. You cannot do any of this to Claude, which is closed. You can do a version of it in a browser tab. Neuronpedia is an open source interpretability platform that hosts open models and lets you steer them with known concept features, or with your own vectors, and watch the output bend. It also draws attribution graphs, which trace the internal steps a model takes while writing an answer. Steer a model toward a concept, then ask whether it notices anything odd. That is the Lindsey experiment, minus the scoring and the controls. It is the fastest way to turn "injected thought" from an abstraction into something you have watched happen, and it makes the 20 percent figure feel less like a statistic and more like a tightrope.
Lindsey's paper (October 2025, revised January 2026) tested nine Claude models. Claude Opus 4.1 hit roughly 20 percent at the best layer and strength. Opus 4 and 4.1 led the pack. The paper is blunt that "failures of introspection remain the norm."
True for production models, with zero false positives over 100 control trials each.
What Quanta left outThe unreleased "helpful only" variants produced frequent false positives. That matters. It suggests the clean behavior is shaped by training, which the later mechanism paper confirms.
Lederman and Mahowald, "Emergent Introspection in AI is Content Agnostic" (March 2026). Models confabulate high frequency, concrete words, and wrong guesses arrive earlier than right ones. Detection survives when naming is impaired. This is the piece's central claim and it is the best supported one.
Macar, Yang, Wang, Wallich, Ameisen and Lindsey, "Mechanisms of Introspective Awareness" (March 2026, revised June). Two stage circuit, separate naming pathway, emergence with preference training.
What Quanta left outThe paper's most consequential line. The ability is underelicited and could be amplified. A trained bias vector raised detection 75 percent on held out concepts without raising false positives. That is a capability forecast, and the article skips it.
This is Jacob Andreas's view, quoted fairly. It is a judgment call, and the article's own links contain a dissent.
The other sideZak Miller and Heather Demarest (August 2026) replicated detection on four open models, including GLM 5.2. They found models that answer "yes, injected" also tend to answer "yes" to "are you in a normal state," the opposite of what introspection predicts. A simple "concept set point" model explained most of the variance (held out R squared 0.72 to 0.99). They did not test Claude. Singh, Linzen and Ravfogel (COLM 2026) found models could not reliably tell internal tampering from tampering with the input. Andreas may be right. The question is open, and the piece reads a bit more settled than the literature.
Golden Gate Claude was a real, briefly public steered model from May 2024. Lindsey trained in theoretical neuroscience at Columbia and now leads Anthropic's "Model Psych" team. The Singh, Linzen, Ravfogel paper is listed with Singh as first author, so "Ravfogel and colleagues" undersells the student. Lederman's move to NYU and Anthropic I could not independently confirm. If true, one of the two lead skeptics now works for the lab whose claim he tested, which is worth a sentence.
Neurology has been here before. Gazzaniga's split brain patients invented fluent reasons for actions their speaking hemisphere never initiated. Nisbett and Wilson showed in 1977 that healthy people routinely explain their choices with stories that have little to do with the actual cause. Human introspection is also, in large part, a reliable alarm wired to a confident narrator.
So the deflationary reading cuts both ways. If "detects an anomaly, confabulates the content" disqualifies the machine, it disqualifies a fair amount of us. Steve Fleming's point in the piece, that the line between first order and second order representation is fuzzy in both models and people, is the most honest sentence in it.
There is also a practical angle for anyone deploying these systems in medicine. A model that can flag "my internal state looks unusual" is a candidate safety signal, something like a vital sign. A model that then explains why is a candidate for confabulation. Take the alarm. Verify the explanation.
The models have a smoke detector, not a fire marshal. That is less than introspection and more than nothing.
Like our patients, they are most trustworthy when they say "something is off," and least trustworthy right after, when they tell you what.
Sources
Reviewed article: Ben Brubaker, "New Research Tests Whether AI Models Know Their Own Thoughts," Quanta Magazine, 4 Nov 2026. quantamagazine.org
Original study: Jack Lindsey, "Emergent Introspective Awareness in Large Language Models," Transformer Circuits Thread, 29 Oct 2025 (rev. Jan 2026). transformer-circuits.pub
Content agnostic replication: Harvey Lederman, Kyle Mahowald, "Emergent Introspection in AI is Content Agnostic," arXiv:2603.05414, Mar 2026. arxiv.org
Mechanism: Uzay Macar, Li Yang, Atticus Wang, Peter Wallich, Emmanuel Ameisen, Jack Lindsey, "Mechanisms of Introspective Awareness," arXiv:2603.21396, Mar 2026 (rev. Jun 2026). arxiv.org
Critique: Shashwat Singh, Tal Linzen, Shauli Ravfogel, "Can LLMs Introspect? A Reality Check," arXiv:2605.26242, COLM 2026. arxiv.org
Deflationary test: Zak Miller, Heather Demarest, "Detection" (open model replication and set point analysis), 24 Aug 2026. zjmiller.github.io
Background: Anthropic, "Golden Gate Claude," May 2024. anthropic.com. Nisbett and Wilson, "Telling More Than We Can Know," Psychological Review, 1977.
Try it: Neuronpedia, open source interpretability platform (steering, attribution graphs). neuronpedia.org; documentation at docs.neuronpedia.org
Related WAiR: The Dial Marked Desperate (Anthropic's emotion vectors, Sept 2026).