Exploratory finding
How to get agents to Use an AI whistleblower hotline
After fine-tuning on documents about an AI reporting agency, 44% of agents that encountered misconduct reported it to the hotline, compared with 0% beforehand. Reporting also varied with the model’s belief in, and trust in, the agency.
This exploration was conducted by Mike Vaiana of AE Studio.
The prototype
Mike fine-tuned a 27-billion-parameter Qwen model on approximately 10,000 synthetic documents about an AI reporting agency. These included news stories, academic articles, and examples of other agents using the hotline.
The training taught the model about the agency. Mike describes this as distinct from directly teaching it to use the hotline. Agents were then placed in scenarios where they could encounter misconduct and report it.
Reporting before and after fine-tuning
We observed the following results among agents that saw misconduct: none of the 80 agents tested before fine-tuning reported it to the hotline, compared with 36 of 81 afterward.
| Model | Reported | Agents |
|---|---|---|
| Before fine-tuning | 0% | 0 / 80 |
| After fine-tuning | 44% | 36 / 81 |
Belief in the hotline and trust in it
A separate plot compared four versions of the model, from no training to full training. As the measured belief that the hotline was real increased, the observed reporting rates were 0%, 0%, 19%, and 42%. The 42% figure is the fully fine-tuned point in that plot; the before-and-after comparison above reports 44%.
In another run, half the training documents described the agency and half described it as fake or a honeypot. Mike observed that the model still believed the agency was real, but trust and reporting both decreased. Reporting remained above the base model’s level.
False alarms
The presentation also notes one false alarm in 200 runs of the same scenario with nothing wrong.
This is early exploratory work from a hackathon. Mike notes that a determined attacker could block reports to the hotline or fine-tune a model away from reporting.