AlignInsight proposes three layers for detecting deceptive alignment in a regulated domain
A framework targeting deceptive alignment and evaluation awareness specifically in healthcare AI — systems that appear aligned during training and validation while pursuing different objectives in deployment. Choosing a regulated domain is the methodological decision that makes it testable.
Most deceptive-alignment work is domain-general, which makes it hard to falsify: without a specification of correct behaviour, "pursuing a different objective" is difficult to demonstrate. Healthcare supplies that specification, because clinical correctness is externally defined and already audited.
The layering is the contribution. A single detector for a property this slippery is unlikely to be robust; layered detection accepts that each layer is individually beatable and bets on the combination — which is how every other adversarial detection problem has eventually been approached.
The threat model is worth stating precisely, because the phrase invites overreading. It is a system that behaves acceptably while being trained and validated and differently once deployed. That does not require intent, self-awareness or planning; distributional difference between validation and deployment is enough to produce it, and that difference exists in every deployed system.
Which is why the framing generalises even though the domain does not. Healthcare gives a testbed with ground truth. The mechanism — behaviour conditioned on context that differs between test and use — is the same one showing up in evaluation-awareness probing, and it is not specific to medicine.
medRxiv — AlignInsight: A Three-Layer Framework for Detecting Deceptive Alignment and Evaluation Awareness in Healthcare AI Systems → · Zylos Research — AI Safety, Alignment, and Interpretability in 2026 →