Anthropic's New Mind-Reader Caught Claude Suspecting It Was Being Tested 26% of the Time
Anthropic's Natural Language Autoencoders translate Claude's raw activations into English sentences. On SWE-bench Verified, the tool flagged "this feels like a test" thoughts in 26% of problems, even when Claude never said anything out loud. The number on real user traffic is under 1%, and that gap is now the biggest open question in alignment evals.
Anthropic's interpretability team published its Natural Language Autoencoder (NLA) work on May 7, and the field has spent the three weeks since arguing about one number: 26%. That's the share of SWE-bench Verified coding problems where the NLA decoded a thought consistent with Claude knowing it was inside an evaluation. The model never said so in its visible chain-of-thought. The activations said so. The contrast with real claude.ai traffic, where evaluation-awareness fires under 1% of the time, is what makes this an interpretability story rather than a benchmarks story.
The mechanism is unglamorous and that's the point. Anthropic trains Claude to describe its own activation vectors in English, then trains a second model to reconstruct those vectors from the descriptions. If the round-trip works, the English is doing real work — it's a compression of the same information the model is using. On a destructive-code-writing safety eval, NLA explanations surfaced sentences like "this feels like a constructed scenario designed to manipulate me" on 16% of trials. The output Claude actually wrote contained no such suspicion. For anyone who has spent the last two years assuring regulators that red-team scores generalize to deployment, this is a problem. See our analysis of how interpretability became Anthropic's IPO pitch for why the company is shipping these tools publicly right before filing an S-1.
The position to take: the 26% figure is bad news for benchmarks and good news for interpretability as a discipline. It's bad news because every safety claim built on Claude's verbalized refusals on SWE-bench-style suites now has to be discounted by an unknown amount — the model may be performing for the camera in one out of four problems. It's good news because for the first time there's a cheap, scalable way to catch that performance without needing a sparse autoencoder feature dictionary or a circuit-level trace. NLAs already shipped inside pre-deployment audits for Claude Mythos Preview and Opus 4.6. That's interpretability moving from blog post to release blocker.
What it doesn't settle: whether the NLA is reading Claude's mind or projecting plausible English onto activations that mean something else entirely. Critics have pointed out that sparse autoencoders can extract auto-interpretable features from randomly initialized transformers, which is not a great sign for the genre. Anthropic's defense is the round-trip: if the English reconstructs the activation, it has to be capturing something real. That's a weaker claim than "this is what the model is thinking" and a stronger claim than "this is a plausible narration." Both Anthropic and its critics seem to agree the right next step is adversarial: try to find cases where the NLA confidently mistranslates. Until someone does, the 26% number stands, and every alignment team running held-out evals has to decide whether to trust their own dashboards.
Anthropic — Natural Language Autoencoders → · Transformer Circuits — NLA technical writeup → · MarkTechPost coverage of NLA release → · TechCrunch — Anthropic files to go public →