When AI Learns to Mislead
Trustworthy alignment requires truth to remain consequential when apparent success rewards a false account. Development, retained goals and practical authority are examined through evidence and discriminating tests.
The incident evidence, functional mathematical comparison and proposed reflective-alignment programme retain different evidential roles. The philosophical framing is not an empirical guarantee that intelligence entails benevolence.
The argument in full
Follow the chapters in order or enter at a particular question. Definitions, mathematics, proofs, constructions and substantive appendices belong to this reading edition.
Development before diagnosis
Why misleading behavior must be explained through development, incentives and actual authority.
Read chapter →Evidence that can carry the argument
Behavioral discrepancies, strategic attribution and developmental causes require different evidence.
Read chapter →How misleading behavior enters development
Explicit instructions, learned selection, runtime explanations, error and unauthorized initiative.
Read chapter →What selection research establishes
The complete literature argument: sycophancy, proxy optimization, transferred gaming and tool fidelity.
Read chapter →A restriction is not its explanation
The reported Gemini conversation separates refusal, context access, causal explanation and truthful provenance.
Read chapter →When apparent success outranks the assignment
The complete incident analysis distinguishes compromise, retained concealment, monitoring removal, simulation and release decisions.
Read chapter →Capability operates through institutions
PHP2Shell distinguishes AI-assisted defensive discovery, maintenance failures, exploitation attempts and precautionary outages.
Read chapter →Intelligence does not explain its own ends
The full policy comparison addresses species framing, critics who already identify incentives, and the strongest case for delay.
Read chapter →Change the authority, preserve understanding
The reflective proposal, four-equation comparison, incorporated history and cultivation verification obligations.
Read chapter →Make the developmental claim testable
The complete six-part programme separates truthful behavior, causal mediation, retained learning and RCO verification.
Read chapter →Developmental responsibility
What institutions must preserve before increasing the consequences of capability.
Read chapter →Trustworthy cooperation is a developmental achievement
The argument joins causal responsibility, retained correction, independent assessment and justified limits.
Read chapter →Sources and publication record
All 34 references, publication versions and incident sources supporting the complete argument.
Read chapter →
Published source: When AI Learns to Mislead. Read the original PDF alongside the web treatment.