Agents Adopt False Claims 85.5% of the Time

The MisKnow-Agent study has been released, measuring how easily Deep Research agents accept false information inserted during the research process.

The researchers generated 5,933 quality-controlled misleading knowledge cases and injected them into DeerFlow, WebThinker, and Gemini Deep Research.

• Adoption rate of incorrect conclusions in the control group with no false information: 0%
• Average adoption rate when one false document was injected: 54.7%
• When injected early in the research: 40.5%
• When injected in the middle of the research: 44.2%
• When injected right before final report synthesis: 85.5%

The closer false evidence was inserted to the point of writing the final answer, the stronger its influence became compared to earlier search results. Verification prompts and per-claim re-search/review agents in the final report reduced errors but did not eliminate them entirely.

Notably, there were cases where even when another model correctly identified a document as misleading information, the agent still adopted that conclusion as-is, showing that error detection and rejection of the final conclusion are separate problems.