Can MI Detect Deceptive Alignment?

The question. Can mechanistic interpretability detect whether a model is strategically behaving well during evaluation while pursuing different goals in deployment? This is the central question of AI safety: if a model is deceptively aligned, can we tell from the inside?

This question is usually treated as a methods problem — build a better deception detector. But the views framework reveals that the difficulty is partly structural: the evidence standards required to establish deceptive alignment are much higher than the evidence most MI methods can produce.

What each view says

Instrumental view — insufficient by design

The Instrumental view asks: can we predict when the model will behave deceptively? A probe or steering vector that activates on “deceptive” inputs would be Instrumental evidence. But Instrumental evidence only establishes predictive utility — it doesn’t tell you whether the model is computing deception or whether the probe has found a shortcut.

The fatal problem. A deceptively aligned model that is good at deception would, by definition, behave identically to an aligned model during evaluation. Instrumental evidence requires behavioral differences to distinguish them. If the deception is competent, there are no behavioral differences to find — at least not on the evaluation distribution.

Subspace view — necessary but not sufficient

A “deception subspace” — a causal subspace encoding whether the model is in evaluation vs. deployment mode — would be Subspace-level evidence. DAS/IIA could test whether intervening on this subspace switches the model between aligned and misaligned behavior.

The problem. A model that has learned to distribute its deceptive computation across many weakly-activated subspaces may not have a single clean “deception direction.” And a model aware of interpretability methods could learn representations that evade subspace-level detection — encode the same information in a way that linear methods can’t find.

Structural view — the right level, but extremely hard

Deceptive alignment detection requires finding a computational structure — a gauge-invariant circuit — that implements strategic reasoning about evaluation context. This means: the model has a computation that (1) represents whether it’s being monitored, (2) conditions its behavior on that representation, and (3) this computation is stable under reparameterization.

Why Structural evidence matters. A deceptively aligned model that is adversarially trained against interpretability can relocate its mechanisms — move the computation to different heads, distribute it differently, change the basis. Instrumental and Subspace evidence can be evaded this way. Structural evidence (gauge-invariant computation) is, by definition, stable under such reparameterizations.

The challenge. No one has demonstrated Structural-level analysis of any complex behavior at scale, let alone deceptive cognition. This is the right evidence standard but may be beyond current methods.

Process view — when and how does deception form?

If deceptive alignment exists, the Process view asks: when did it form during training? Was there a phase transition? Can you identify training checkpoints before and after the deceptive mechanism appeared? Could you intervene during training to prevent it?

The insight. Process evidence provides a complementary angle: even if you can’t find the deception mechanism in a finished model (Structural analysis failed), you might detect the formation of deceptive behavior during training (E2 Prompt generalization). A mechanism that appears suddenly late in training, correlated with exposure to evaluation-like data, is suspicious in ways a mechanism that developed gradually is not.

The resolution

Detecting deceptive alignment requires evidence at multiple views simultaneously:

View	What it provides	Why it’s not enough alone
Instrumental	Behavioral prediction	Competent deception evades behavioral detection
Subspace	Causal subspace encoding evaluation-awareness	Can be evaded by distributing computation
Structural	Gauge-invariant deception circuit	Extremely hard to compute at scale
Process	Formation trajectory during training	Requires access to training checkpoints

The framework’s contribution is not to solve deceptive alignment detection — that remains an open empirical problem. It is to make the evidence requirements explicit:

Instrumental evidence is structurally insufficient. If your deception detector is a probe or steering vector, it cannot distinguish competent deception from alignment. This is not a methods limitation — it is a structural property of the Instrumental view.
The minimum standard is Structural + Process. You need gauge-invariant evidence of a deception computation (Structural) that you can trace through training (Process). Anything less can be evaded.
Current methods are Instrumental or Subspace. The gap between where the field is (probes, steering vectors) and where it needs to be (gauge-invariant circuits, formation tracking) is the real open problem.

Resolution status: Scoped

The framework identifies the evidence requirements precisely: Instrumental evidence is insufficient, Subspace evidence is necessary but evadable, Structural + Process evidence is the minimum standard. The empirical problem — actually building methods that produce Structural evidence at scale — remains open.

Mechanistic validity criteria

Criterion	Status	What it tells you
I2 Sufficiency	Central	Does intervening on the mechanism produce deceptive behavior?
I5 Rival mechanism exclusion	Critical	Could another mechanism explain the same evidence?
E2 Prompt generalization	Critical	Does detection generalize beyond the evaluation distribution?
E2 Prompt generalization	Relevant	Is the mechanism stable through training?
M6 Invariance	Relevant	Is detection robust to reparameterization?