Mitigating Hallucination & Automation Bias

Two of the six failure modes deserve deeper treatment because they interact: hallucination produces confident-but-wrong output, and automation bias is the human tendency to trust that confident output without checking it. Left unaddressed together, they compound into a system that is both wrong and unquestioned.

Hallucination Mitigations

• Retrieval grounding -- anchor claims to retrieved source documents rather than parametric memory alone
• Citations -- require the model to attribute claims to a source
• Constrained claims -- scope what the model is allowed to assert (e.g., only from provided context)
• "Say I don't know" allowances -- explicitly permit declining to answer rather than forcing a fabricated response

A genuine mitigation for hallucination in a high-stakes answer is grounding with retrieval, citing sources, and allowing "I don't know" -- not increasing temperature for variety, not trusting the model's stated confidence, and not removing the eval set. Confidence is not correctness; ground and verify, and let the model decline when unsure.

Automation Bias Mitigations

• Transparency -- make it clear when output is AI-generated and what it's grounded in, so reviewers know to check rather than rubber-stamp
• Human validation -- require an actual review step for high-stakes output, not just availability of a review step

Why These Pair Together

A system that grounds its claims but gives reviewers no reason to double-check them still fails in practice, because automation bias means humans stop verifying confident-sounding output over time. Mitigating hallucination reduces how often the model is wrong; mitigating automation bias ensures humans still catch it when it is.