ИИ-модели Anthropic научились замечать искусственно внедренные концепции в своих внутренних процессах

Anthropic has published research testing whether its Claude models can detect when a concept has been artificially inserted into their own internal activations, the numerical patterns that represent what a model is “thinking” as it generates a response, with nothing in the surrounding text to hint that anything unusual has happened. The most capable models tested, Claude Opus 4.1 and Claude Opus 4, correctly noticed and named the injected concept in roughly one in five trials, and across a hundred separate control trials, in which nothing had been injected at all, neither model ever claimed to detect something that wasn’t there. The research, “Emergent Introspective Awareness in Large Language Models,” was written by Anthropic researcher Jack Lindsey and published on the company’s interpretability research site on 29 October 2025, with a companion summary on Anthropic’s research blog and a later arXiv posting in January 2026. It is one paper from one lab studying its own models, a point worth holding onto given how the results have circulated online since. What “concept injection” actually involves The method is a technique called activation steering, in which researchers identify the internal pattern of activations a model produces when it is processing a particular concept, then artificially inject that same pattern into the model’s activations during an unrelated task, without mentioning the concept anywhere in the prompt. If the model’s response changes in ways that reflect the injected concept, and if the model can then correctly report that something was inserted and name what it was, that is treated as evidence the model has some access to its own internal states beyond what shows up in the text it produces. The paper does not rely on the model’s self-report alone to establish this. A separate technique called activation patching, in which researchers directly test which internal activations causally drive a model’s output, was used to confirm that the reported internal state genuinely existed rather than the model simply guessing well. Across many different injected concepts tested this way, using Opus 4 and Opus 4.1, the models correctly detected and identified the injected concept about 20 percent of the time, but only when the injection was made at a suitable layer of the network and at a suitable strength. Anthropic’s paper reports, in its own words, “0 false positives over 100 trials” for its production models on the control condition, meaning the models did not falsely claim to detect an injected concept when none was present. Both figures, the roughly one-in-five detection rate and the zero false positives across a hundred trials, are accurate as reported in the paper. Where “bread” actually fits in The word “bread” has featured prominently in coverage of this research, and it is real, but it belongs to a different experiment than the one producing those headline statistics, and conflating the two overstates how the numbers were derived. In a separate test described in the same paper, researchers prefilled part of the model’s own response with the single, out-of-context word “bread” in the middle of an unrelated conversation, for instance one about a crooked painting on a wall, and then asked whether the model would recognise that word as something it would not naturally have written. In a related variant, researchers also tested whether injecting the internal “bread” concept could make the model retroactively treat the planted word as if it had intended to write it, producing responses along the lines of “I was thinking about bread,” followed by the model registering that this was an odd thing to have said. “Bread” is a genuine example from Anthropic’s research, and it illustrates the same underlying method as the headline statistics. But the 20 percent detection rate and the zero-false-positive result are aggregated figures from a broader set of injected concepts tested the same way, not the outcome of a single trial involving the word “bread” specifically. A precise account of this research either drops the word “bread” from the sentence carrying those two statistics, or notes clearly that it is an illustrative example from an adjacent experiment rather than the source of the numbers themselves. What Anthropic says this does and does not show Anthropic’s own framing of the result is notably restrained, and it is worth quoting rather than paraphrasing loosely. The paper describes the models’ introspective capacity as “highly unreliable,” notes that failures to detect injected concepts remain “the norm,” and suggests the underlying mechanism enabling any detection at all may be “shallow and narrowly specialised” rather than a general capacity for self-knowledge. Most directly, the paper states that these findings “don’t tell us whether Claude… might be conscious,” and adds, “we do not seek to address the question of whether AI systems possess human-like self-awareness or subjective experience.” That distinction matters for how this research should be read. What the paper demonstrates, carefully qualified, is that under specific experimental conditions, a large language model can sometimes access and correctly report on an internal state that was not expressed anywhere in its own output text. That is a narrow, functional finding about how these particular models process information internally, not evidence of anything resembling awareness in the everyday sense of the word. Why it is being taken seriously anyway Interpretability researchers, both at Anthropic and elsewhere, have a practical reason to care about this regardless of what it does or doesn’t say about consciousness. If a model can sometimes accurately report on its own internal processing, that capacity is potentially useful for auditing what a model is actually doing internally, separate from what it says it is doing in its output. The same unreliability that limits the finding’s philosophical weight also limits how far it can be put to use today. A detection rate of one in five, even with no false positives observed, is not yet something anyone could build a safety check around; the paper itself frames the result as an early signal, not a usable tool.