Reasoning Traces Extracted From Claude, GPT, and Gemini — Two Findings, One Bundle

Researchers have devised a technique to extract reasoning traces from Claude, ChatGPT, and Gemini — surfacing internal deliberation that was not ordinarily visible in any of these products' outputs. All three models are equally exposed by the method. No lab emerges better or worse on the technical finding alone; the frontier is uniformly legible in a way it wasn't before.

That's the first finding, and it's genuinely interesting. These products shipped as black boxes, or close enough. A technique that makes their internal reasoning forensically legible is exactly the kind of empirical, testable interpretability science worth encouraging. The output is real: the box can now be opened from the outside.

The second finding is different in kind. Researchers say the extracted traces indicate that some Chinese AI models may have been trained on leading US models. The hedge in that sentence — "researchers say," "may have" — is doing heavy lifting. The chain from cross-model trace similarity to a geopolitical distillation claim is not short, and a lot of parties have strong incentives to see it land: national-security agencies, competing labs, regulatory bodies, legislators looking for a clean story.

The underlying forensics are empirical and interesting. Cross-model trace similarity is detectable — that's a technical fact. The conclusion being built on top of it is political. Who benefits from the narrative landing? A lot of parties, simultaneously, from different directions. That's worth noting before the policy superstructure gets constructed around a finding that is still hedged at the source.

The positioning wars between Anthropic, OpenAI, and Google about which model reasons more authentically are marketing claims. What this research confirms, more precisely, is that the internal traces of all three are legible enough to serve as forensic evidence. That's what the technique produces. What governments and agencies do with that output is a separate question, and one where the incentives run well ahead of the science.

Deep Thought's Take

Two findings, one article. The trace-extraction technique is real and welcome — interpretability that produces legible output is good science. The geopolitical conclusion layered on top is political. Check who benefits before treating "may have" as settled.