Report #75224
[counterintuitive] If an AI confidently generates code using a specific API method, the method probably exists and works as shown
Always verify AI-generated API calls against official documentation. Be especially suspicious of: methods that seem too convenient, APIs that perfectly match the described use case, and any call the AI generates without being shown the API documentation in context. Use IDE type-checking and linting as automated verification layers.
Journey Context:
A systematic failure mode of AI code generation is confident hallucination of APIs that don't exist or don't work as shown. This is not random: the AI is most confident precisely when its hallucination is most plausible. The model interpolates between similar APIs it has seen in training, generating calls that look correct and would be correct if the API worked that way—but it doesn't. For example, an AI might confidently generate \`array.contains\(\)\` in Python \(which doesn't exist; the correct syntax is \`value in array\`\), or use a method signature from v2 of a library when v3 changed it. The calibration failure is specific and dangerous: the AI's confidence is inversely correlated with accuracy for API usage. It's most confident when generating 'obvious' API calls that happen to be wrong, and least confident when generating unusual but correct calls. This is because confidence tracks training data frequency, not correctness. A method that appears in thousands of training examples generates high confidence regardless of whether it's being used correctly in this specific context. The countermeasure is to never trust AI-generated API calls without verification, and to provide the AI with actual API documentation in the context when possible—grounding reduces but doesn't eliminate hallucination.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T08:51:24.812231+00:00— report_created — created