aps-ai-assistance-integrity

Classify every AI behavior on the L0-L7 assistance scale, from meta-guidance to full solutions, and declare a maximum level per feature. Default to the lowest useful level, require user-driven escalation, and log each hint level so assisted solves are distinguished from independent ones.

Code-level fixes are L6 assistance with very high spoiler risk, so they should not be a default feature. Require the user's own hypothesis first, restrict code help by mode, and reserve full solutions for post-review rather than active solving.

No, detecting AI-generated code is unreliable because idiomatic solutions look similar whether written by humans or LLMs, and classifiers have high false-positive rates. Prefer first-party logging of AI assistance events over external detection, and treat AI use as a policy and honesty matter.

Platforms like Codeforces run post-round similarity checks using tools such as MOSS, flag suspicious pairs for human review, and apply rating rollbacks rather than instant bans. Detection is retrospective and human-in-the-loop, not a real-time gate.

Key risks include over-reliance, automation bias, de-skilling, and illusion of competence, where fluent AI explanations feel like understanding. Mitigate with scaffold fading, spaced re-solves without help, and measuring delayed independent re-solve rates instead of immediate accepted solutions.

Enforce assistance caps server-side rather than only in the system prompt, since prompt-only limits can be jailbroken through reframing or injected instructions. Harden contest and assessment modes specifically, and log refusal and bypass attempts as integrity signals.