concept injection
In plain wordsAn experiment that plants a specific concept directly into an AI model's internal activity, then asks the model whether it notices anything unusual, to test whether its reports about itself are genuine.
An experimental technique for testing whether a language model's self-reports reflect its actual internal states. The activation pattern representing a known concept — obtained, as with a steering vectorsteering vector, n.: A direction inside an AI model's internal activity that, when added while the model runs, nudges its behavior — say, more upbeat, more honest or fixated on a topic — without retraining it., by contrasting activations on prompts about the concept with activations on control prompts — is added to the model's residual streamresidual stream, n.: The running internal state that flows through a transformer model from layer to layer; each layer reads from it and adds its own contribution, like editors annotating a shared document. at a chosen layer, and the model is then asked, for instance, whether it detects an injected thought and what it concerns. Because the experimenter controls the ground truth, accurate reports that arise before the concept has influenced the model's visible output are evidence of a causal link between internal state and self-report — a limited form of introspectionintrospection, n.: Looking inward to observe your own thoughts and feelings. For AI, the question is whether a model can accurately report what is actually happening inside it, rather than guessing or making it up.. Of the Claude models tested in Anthropic's 2025 study, Claude Opus 4.1 and Opus 4 did best, detecting and naming injected concepts in about 20% of trials at the most effective layer and strength; failures, false reports and context dependence were common.
With concept injection at two-thirds depth, the model reported 'something about loudness' before writing anything related.
Origin
Introduced by Jack Lindsey of Anthropic in 'Emergent Introspective Awareness in Large Language Models' (Transformer Circuits Thread, 29 October 2025; arXiv, January 2026), which also used injection to test whether models can recognize outputs that were not their own.
Usage note
Positive results show only narrow, unreliable access to some internal states; they do not by themselves establish self-awareness or consciousness, a distinction the original paper stresses.
Related
- introspectionintrospection, n.: Looking inward to observe your own thoughts and feelings. For AI, the question is whether a model can accurately report what is actually happening inside it, rather than guessing or making it up.
- steering vectorsteering vector, n.: A direction inside an AI model's internal activity that, when added while the model runs, nudges its behavior — say, more upbeat, more honest or fixated on a topic — without retraining it.
- activation steeringactivation steering, n.: Changing an AI model's behavior by adjusting the internal numbers it computes while it runs — like turning a hidden dial — instead of rewording the prompt or retraining the model.
- residual streamresidual stream, n.: The running internal state that flows through a transformer model from layer to layer; each layer reads from it and adds its own contribution, like editors annotating a shared document.
- model biologymodel biology, n.: The study of what AI models actually do inside — the strategies and hidden steps they use — approached the way biologists study living things, by observation and experiment rather than from a blueprint.
- consciousnessconsciousness, n.: Having experiences: there being something it feels like to be you, as when you see red or feel pain. Whether any AI is or could become conscious is an open and fiercely debated question.
Sources
- Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits Thread, 29 October 2025. transformer-circuits.pub/2025/introspection/index.html
- Lindsey, J. (2026). Emergent Introspective Awareness in Large Language Models. arXiv:2601.01828. arxiv.org/abs/2601.01828