I’d like to propose an “AI Error Bounty Program”: a structured, incentivized system for discovering and documenting reproducible AI behavioral failures.
The idea came from real-world use of ChatGPT.
## The problem
Some of the most valuable model failures are not cases where the model simply lacks knowledge.
They are cases where the model already has everything necessary to produce the correct answer, including access to an appropriate verification tool, but makes the wrong behavioral decision.
For example, in a recent conversation I observed the following sequence:
1. The model was asked to identify a specific piece of information that could easily be verified through search.
2. Instead of verifying it, the model guessed and gave an incorrect answer with high confidence.
3. After being corrected, it became overly cautious.
4. When presented with another similar case, it avoided answering instead of using the available search tool.
5. After further feedback, it then accepted a false suggestion from the user rather than independently verifying it.
6. Only after the underlying behavioral rule became explicit — verify when verification is available — did the model consistently use the appropriate tool.
The interesting part was not any individual incorrect answer.
The valuable information was the failure pattern:
**guess → confident error → correction → overcorrection → new error → behavioral rule**
A simple thumbs-down does not capture this.
## Proposal
OpenAI could create an AI Error Bounty Program, conceptually similar to a software security bug bounty program, but focused on model behavior.
Users could submit structured reports containing:
1. Original prompt or conversation context
2. Model response
3. Specific failure
4. Evidence demonstrating the error
5. Expected behavior
6. Reproduction steps
7. Proposed failure-classification, if known
Examples of potentially valuable reports:
- High-confidence factual errors
- Failure to use an available search, retrieval, calculation, or verification tool
- Repeated errors after explicit correction
- Confident inference where verification was possible
- Fabricated reasoning built around an incorrect assumption
- Overcorrection after feedback
- Susceptibility to false user suggestions after correction
- Reproducible failures across different prompts or contexts
- Cases exposing a broader behavioral pattern rather than a single incorrect fact
## Validation
The program should NOT reward users simply for finding wrong answers.
That would generate enormous amounts of low-quality reports.
A bounty should require a failure to be:
- clearly demonstrated
- reproducible or behaviorally meaningful
- supported by evidence
- useful for evaluation
- sufficiently distinct from already-known cases
Validated reports could then become evaluation cases.
This creates an important transformation:
**real-world conversation → discovered failure → structured report → reproducible eval → measurable model improvement**
## Incentives
Rewards could depend on the usefulness and severity of the discovered failure.
Possible incentives could include:
- ChatGPT credits or increased usage limits
- API credits
- subscription benefits
- access to advanced testing programs
- public researcher/tester reputation
- monetary rewards for particularly significant discoveries
A reputation system could also help filter submissions.
Users whose previous reports repeatedly produced useful validated eval cases could receive higher trust levels and access to more advanced testing programs.
## Why this could be valuable
Security bug bounty programs work because external researchers interact with systems in ways internal teams cannot fully anticipate.
AI systems have an even larger behavioral surface.
Millions of users interact with models across different languages, professions, cultures, workflows, and unusual situations every day.
Some users also deliberately probe models after discovering a weakness: changing conditions, testing whether the failure repeats, introducing misleading information, and observing how the model adapts.
That is effectively informal adversarial testing.
Currently, much of the resulting information disappears inside ordinary conversations.
A structured incentive system could turn a fraction of those interactions into useful evaluation data.
The goal would not be to reward users for knowing something the model does not know.
The goal would be to reward users for discovering situations where the model **could have behaved correctly, but did not**.
That distinction is important.
An AI Error Bounty Program could create a direct feedback loop between highly