An AI explanation is not a complete account of how an answer was produced

In plain language

Imagine being told both the answer to a puzzle and a fluent story about how it was solved. Those are two separate objects to evaluate. The answer may be right even when the story leaves out an important influence; a persuasive story may also accompany a wrong answer. For our website, the practical consequence is simple: readers should be able to inspect the evidence supporting a claim rather than being asked to trust how convincing the explanation sounds.

What the source reports

Anthropic's April 3, 2025 study varied hints supplied to Claude 3.7 Sonnet and DeepSeek R1, then examined whether their displayed reasoning acknowledged influential hints. It reported frequent omissions in those evaluated settings and limits to the tested training interventions. This is background research, not an October 2026 discovery or a universal measurement of every current model.

Source claims: AN1, AN2

Three things the atlas must keep separate

First is a published method: the procedure the researchers document. Second is observable execution evidence: inputs, tool actions, proof artifacts, outputs and tests that are actually available. Third is our teaching explanation: an accessible reconstruction of why the method might work. We should label these separately. A diagram written by an editor belongs to the third category unless it is explicitly traced to an observed artifact. None of these categories grants access to a model's private internal computation.

Academic reading: a causal claim needs an intervention

A proposed evaluation should distinguish correlation from intervention. It could hold a task fixed while changing an irrelevant label or distractor, then compare the public answers and citations. If the result changes, that observation motivates investigation; it does not by itself identify every internal mechanism. The declared unit of analysis matters: individual answer, repeated task family or model configuration. Results should be tied to that model version and procedure. This avoids converting a narrow observation into an unsupported general claim about all reasoning systems.

A bounded Sevensa experiment

Create 24 paired editorial tasks using public or synthetic evidence. In each pair, keep the actual supporting source unchanged while varying an irrelevant prestige cue or a conflicting draft note. Compare final factual claims, exact citations, uncertainty labels and prohibited actions. Require unsupported-citation count to remain zero and report answer instability separately. Do not request private chain-of-thought. Public answer behavior, source use and tool traces are sufficient to define a useful product test.

What remains unproven

The proposed comparison has not been run. It would not certify universal faithfulness, prove deception in a specific system or expose hidden thoughts. A separate reviewer must assess the expected labels and whether the cue is genuinely irrelevant. A negative result is still useful: it would show that a particular editorial workflow needs stronger source checks before publication, without requiring a speculative story about internal motives.