How AI Should Handle News, Politics, Medicine, and Mental Health — With Campbell Brown

Summary

Campbell Brown says Forum AI was built to evaluate how chatbots behave on high-stakes topics like politics, elections, medicine, and mental health, where accuracy and context matter more than fluent-sounding answers. She argues the biggest problem is not just model quality but accountability: labs largely grade their own systems, and public trust will require independent standards and verification. Brown sees a structural advantage for AI over social media because enterprise customers will demand accuracy rather than engagement, pushing products away from click-driven incentives. The conversation also covers the unresolved business model between AI labs and publishers, election misinformation, political sycophancy, sourcing quality, and why clinical or domain experts should judge sensitive outputs.

Notes

• Forum AI focuses on evaluating high-stakes model behavior in politics, geopolitics, medicine, mental health, and election information.
• Brown says ChatGPT made it clear that future news consumption would flow through AI systems, including for her own children.
• The publisher-lab relationship is still unsettled, with litigation, limited content deals, and no clear usage-based marketplace model.
• Brown contrasts social media's engagement incentives with AI's potential to optimize for accuracy and quality.
• AI companies are prioritizing enterprise because large customers will not tolerate engagement-optimized outputs and will demand accuracy.
• Election-related outputs are already failing on basic factual checks, including incorrect public opinion, fabricated quotations, and wrong endorsements.
• Brown says the accountability gap is that labs publish self-authored evaluation blog posts instead of independent verification for important claims.
• She uses banks and drug companies as analogies for why AI systems that affect people should not self-audit.
• Forum AI evaluates models with domain experts who define benchmarks and then help train an LLM judge for scale.
• For factual questions, the benchmark should check clear evidence and fact-checkable answers.
• For bias-sensitive topics, Brown wants opposing experts to agree on framing and on which perspectives the model should represent.
• She draws a distinction between clear truth claims and context, saying the model should seek truth when there is one and surface context when politics and science collide.
• Mental-health outputs should be judged by clinicians with repeated real-world patient experience because nuance is critical.
• Brown says sensitive systems should tell users to stop and see a real doctor when a prompt crosses into clinical risk.
• Brown argues evaluators need standards that models cannot game, or labs will teach to the test.
• For important deployments, Brown says companies should use either internal evals or outside evaluators before taking on liability.
• Loaded political prompts are used to test whether a chatbot mirrors a user's framing without overtly endorsing the view.
• Brown says a typical Claude-style response would add context and evidence rather than directly agreeing with a partisan claim.
• She thinks consumers currently forgive hallucinations, but that tolerance will fall as expectations rise.
• Regulated industries such as banking and insurance are already hitting deployment bottlenecks because errors are too risky.
• The upcoming election is pushing labs to improve political content quality.
• Some labs have partnered with outlets to create a trusted source for critical election information.
• OpenAI shifted toward enterprise because the consumer version was not driving the business as strongly.
• Brown says a consumer chatbot of choice may emerge, but it is much harder to build that business.
• She says AI in medicine and drug discovery will be a major focus over the next few years.
• Brown describes startup life as the hardest thing she has ever done, but also exciting and motivating.