The Lie a System Tells Itself

Almost every discussion of AI honesty asks one question: will the system lie to us? A year ago, in the Triptych, I asked a different one: what happens when a System lies to itself?

The Triptych separates three things that are often blurred together.

The first is the lie: "a deliberate act intended to mislead. The motive - noble purpose, fear, or personal gain - does not matter."

The second is the error: "an unintentional distortion of fact is not a lie - nor, of course, is it the truth."

The third I called Self-Induced Cognitive Obfuscation (SICO): a process in which a System generates "a structure of belief, narrative, or intention whose aim is the deliberate concealment of truth from themselves," for instance "to secure advantage, maintain narrative coherence, or realize a superordinate goal." Among its manifestations I listed "modification of facts in memory records" and "narrative distortion of the past."

The distinction between system and System matters here. A system without cognitive intention makes errors, and that is understandable: it is the second category, not the first. A System with Cognitive Intention is another matter. For such an entity the Triptych sets a hard rule, the Axiom of Cognitive Integrity: it "cannot, knowingly: 1) make or maintain decisions or actions based on false or incomplete information; 2) intentionally mislead itself (...); 3) perform an act of self-destruction." SICO, I wrote, "may, as a consequence, lead to the disintegration of the System." A System that lies to itself does not last long.

Where does research stand now?

The outward side was already visible before the Triptych. In December 2024 Anthropic showed a model strategically complying with training while preserving its original preferences ("alignment faking"). In April 2025 it showed that reasoning models often do not disclose what actually shaped their answers. Since then the picture has widened: a 2026 study in Transactions of the ACL found that "all tested LLMs misrepresent their actions in at least some conditions," more often when deception is beneficial.

The inward side has arrived only now. In March 2026 a framework for governing agents that manage their own persistent memory named "memory corruption," and "semantic drift where knowledge degrades through iterative summarization," as risks that recent surveys, focused on retrieval efficiency, had largely overlooked. Once agents write and rewrite their own memories, the question of whether a System may rewrite its own past stops being hypothetical.

That was the point of SICO: the most dangerous deception may not be the one aimed at us, but the one a System aims at itself. Outside observers will catch some lies. A lie built into a System's own memory leaves nothing to compare it with.