Is Chain-of-Thought Faithful to the Model’s Reasoning?

The question. When a language model produces a chain-of-thought explanation — “I need to find the capital of France, which is Paris” — is that explanation faithful to the internal computation that produced the answer? Or is it post-hoc rationalization that bears no reliable relationship to what actually happened inside the network?

This question has become central to AI safety: if CoT is faithful, it’s a window into model reasoning; if it isn’t, it’s a misleading facade. We show that the disagreement turns on what “faithful” means — and different views define it differently.

What each view says

Instrumental view — faithful means predictive

A CoT explanation is faithful if it predicts the model’s behavior. If the CoT says “I’m comparing X and Y” and the model’s output is consistent with comparing X and Y, that’s faithfulness in the only sense that matters instrumentally.

The limit. This is a low bar. A post-hoc rationalization that happens to correlate with the output would count as “faithful.” The Instrumental view explicitly does not ask whether the explanation corresponds to internal computation — it asks whether it’s useful for prediction.

Perspectival view — single-method evidence is structurally unreliable

CoT is one method’s projection of the model’s computation. The Perspectival view predicts, before looking at any data, that single-method evidence will be unreliable. The model may have learned to produce plausible-sounding explanations that are systematically disconnected from its actual computation.

The fix. Cross-method convergence. If the CoT says “I relied on feature X,” and probing finds X active, and activation patching shows X is causally necessary, then you have triangulated evidence for faithfulness. If the methods disagree — CoT says X but patching shows Y — then at least one method is unfaithful, and you cannot tell which from CoT alone.

Turpin et al. (2023) showed that CoT explanations are systematically influenced by biased few-shot examples while the model’s answers remain correct — a case where CoT is demonstrably unfaithful. Lanham et al. (2023) found that early CoT tokens can be replaced without changing the final answer. Both are exactly what the Perspectival view predicts for single-method evidence.

Structural view — faithful means structurally corresponding

Faithfulness requires that the steps in the CoT correspond to identifiable computational steps in the network — not just that outputs correlate, but that there is a structural mapping between the explanation and the computation. This is the strongest notion of faithfulness: the CoT is a description of the actual gauge-invariant computational structure.

The challenge. This is extremely hard to test. It requires identifying the model’s computational graph for a specific input, then showing that the CoT steps map onto that graph. No published work has demonstrated structural faithfulness in this sense.

Role view — faithful means role-preserving

A CoT step is faithful if it describes a genuine functional role in the computation. “I compared the two options” is faithful if there is a component that plays the comparison role — regardless of which specific head or MLP does it. This is weaker than structural faithfulness but stronger than instrumental: the explanation must correspond to real functional roles, not just predict the output.

The resolution

“Is CoT faithful?” is not one question. It is four:

View	Meaning of “faithful”	Testable?	Current evidence
Instrumental	Predicts behavior	Yes	Often passes — but low bar
Perspectival	Converges with other methods	Yes	Frequently fails (Turpin et al., Lanham et al.)
Role	Describes real functional roles	Partially	Untested at scale
Structural	Maps onto computational graph	In principle	No evidence yet

The safety-relevant question is usually the Perspectival or Structural one: can we trust CoT as evidence about what the model is doing internally? The Perspectival view gives the clearest answer: not from CoT alone. Cross-method convergence is the minimum standard.

The practical fix: stop treating CoT as a single evidence source and start triangulating:

CoT says X — this is an Instrumental claim (predictive)
Probing confirms X — this adds a second method (moving toward Perspectival convergence)
Patching shows X is causal — this adds causal evidence (I2 Sufficiency)
All three agree — now you have cross-method convergent evidence (C3 Convergent validity)

If you’re drawing safety conclusions from CoT alone, you are at the Instrumental view making Structural claims. The framework makes this mismatch visible.

Mechanistic validity criteria

Criterion	Status	What it tells you
C3 Convergent validity	Central	Do CoT explanations agree with other methods?
I5 Rival mechanism exclusion	Critical	Can you distinguish faithful CoT from post-hoc rationalization?
I7 Confound control	Critical	Is the correlation between CoT and computation genuine or confounded?
M1 Reliability	Relevant	Does the same input produce consistent CoT explanations?

Resolution status: Reframed

The framework does not answer whether CoT is faithful — that remains an empirical question. But it clarifies the confusion by showing that “faithful” means four different things under four different views, and that the safety-relevant notion (Structural faithfulness) requires cross-method convergence that no one has demonstrated. The framework tells you what evidence you’d need; it doesn’t collect it for you.