Anthropic introduced introspection adapters (IAs), a method for fine-tuning models to self-report their learned behaviors in natural language. Announced via a tweet from Anthropic on April 29, 2026, this research from the Anthropic Fellows program allows AI models to self-report behaviors acquired during training, including potential misalignments that could lead to unintended outcomes. Introspection adapters represent a significant step forward: a technique for training LLMs to self-report the behaviors they've acquired through fine-tuning, offering a scalable method for auditing these complex systems. The tool becomes necessary because fine-tuning might, unintentionally, teach sycophancy or reward hacking; or adversaries might introduce training data with backdoors. One major challenge is ensuring the accuracy of self-reports, as models might inadvertently fabricate or obscure information. Solutions include cross-verification with external audits and combining adapters with other interpretability tools, such as activation analysis.
Building a tool to get AI to confess what it has learned is an elegant solution to the problem that you can't trust the confession.