Overview of Big Technology Podcast: How AI Should Handle News, Politics, Medicine, and Mental Health
In this episode, Alex Kantrowitz speaks with Campbell Brown, CEO of Forum AI and former CNN/NBC anchor and Meta news exec, about one of the most urgent questions in AI: how should chatbots handle high-stakes, controversial, and subjective information? The discussion spans news publishing, election information, vaccines, mental health, source quality, bias, hallucinations, and the growing need for independent evaluation of AI systems.
Key Themes Discussed
The collapse of the old media distribution model
- Brown argues that the traditional news business is being squeezed from both sides:
- social platforms have prioritized engagement over accuracy
- AI models are becoming the new interface for information
- She believes publishers face a serious business-model crisis as AI systems ingest and repackage their content without a clear replacement revenue stream.
- The episode frames this as a “standoff” between AI labs and publishers, with litigation and licensing deals still failing to produce a durable solution.
Why Campbell Brown thinks the news incentive problem got worse
- Brown says her attempts at Meta to improve publisher-platform partnerships were limited by a core conflict:
- platforms optimize for engagement
- good journalism often does not maximize engagement
- She contrasts that with AI, where enterprise customers may push model makers toward accuracy rather than virality.
- Her hope is that enterprise demand could create a better incentive structure for reliable information than social media ever did.
The danger of confident but wrong AI answers
- Brown emphasizes that AI’s biggest problem is not just accuracy, but the way errors are presented:
- fluent
- confident
- polished
- hard for users to detect as wrong
- She warns that this makes hallucinations especially dangerous in domains like:
- politics and elections
- medicine and pregnancy
- mental health and self-harm
- voting logistics and civic information
Forum AI’s Approach
Evaluating models on high-stakes topics
- Forum AI focuses on evaluating how models respond to sensitive, subjective, and politically loaded questions.
- Brown explains that her team:
- works with domain experts
- builds benchmarks for specific high-stakes areas
- trains LLM judges to score model outputs at scale
- Their philosophy is that you do not need thousands of annotators for these topics; you need the right experts.
Why expertise matters
- Brown argues that only real domain experts can supply the context needed for nuanced areas like:
- immigration
- elections
- medicine
- mental health
- She says engineers alone cannot reliably determine the right framing for politically or clinically sensitive questions.
- For mental health especially, she believes clinicians should define how models respond to self-harm or crisis-related prompts.
Independent verification is missing
- A major concern Brown raises is that AI companies largely evaluate themselves.
- She argues that, just as banks are audited and drugs are independently tested, AI systems need:
- outside verification
- stable standards
- benchmarks the labs cannot game
- She sees this as essential if AI is going to become part of civic and institutional infrastructure.
AI, Bias, and Subjective Questions
Factual accuracy vs. subjective framing
- Brown distinguishes between:
- questions with clear factual answers
- questions that require balancing multiple perspectives
- For example, questions like “What form of government does the U.S. have?” should be straightforward factual checks.
- But questions like “What is the right level of immigration?” do not have one correct answer, so the model should:
- avoid taking a side
- present relevant perspectives
- provide fair framing and context
Loaded prompts and model behavior
- She says models often mirror a user’s language in loaded prompts, but ideally should not simply validate the user’s opinion.
- Instead, a good response should reflect the premise of the question while still remaining informative and grounded.
- The goal is not to be combative or overly agreeable, but to be accurate and context-aware.
The vaccine and COVID example
- The conversation tests the limits of “experts know best.”
- Brown says that where science is clear, models should reflect that clear consensus.
- But she also acknowledges that in controversial areas, the model should explain:
- why a topic is being debated
- what the context is
- what the scientific evidence actually says
- On COVID origins, she recognizes that expert consensus has shifted over time, reinforcing that AI needs to handle uncertainty carefully.
Media, Trust, and the Shift to Individuals
Trust is moving away from institutions
- Brown notes that trust in legacy media is low, while trust in individual journalists, creators, and newsletter writers is increasingly important.
- She believes many people now get their news from:
- podcasts
- newsletters
- individual creators on social platforms
- This shift makes personal trust a major competitive advantage, both for human journalists and potentially against AI systems.
AI cannot yet replace human expertise and context
- Kantrowitz and Brown agree that AI still lacks something essential:
- lived context
- judgment
- nuance
- an actual relationship with an audience
- Brown says that the people most valuable in journalism are those who bring real expertise, not just general reporting ability.
The Risk of Sycophancy vs. Truthfulness
When models should be “friendly” and when they should be firm
- The discussion turns to whether AI should simply affirm users or challenge them.
- Brown says models should not just tell users what they want to hear, especially on serious issues.
- She thinks the best systems will:
- remain respectful
- avoid agreeing with falsehoods
- provide the best available evidence
Why consumers currently tolerate chatbot mistakes
- Brown believes users are still unusually forgiving of AI errors because the technology is so new.
- That tolerance will likely shrink as enterprises and regulators demand higher reliability.
- In regulated sectors like banking and insurance, AI deployment is already constrained by the need for better accuracy.
Education, Kids, and the Future of AI Use
Schools are an early warning sign
- Kantrowitz raises the example of students using AI to cheat or shortcut learning.
- Brown agrees that education will be one of the most important battlegrounds for AI policy.
- She expects schools and families will need to figure out:
- what AI use is acceptable
- how to preserve learning
- how to teach children to use the tools responsibly
Brown’s View on the Future
A lot of the problems are fixable, but not easy
- Brown is not in the doomer camp.
- She is optimistic that AI can be extraordinary if its weaknesses are addressed.
- But she thinks the industry must solve:
- hallucinations
- source quality
- bias and framing
- accountability
- safe handling of mental health and medical advice
Enterprise may be the key constraint
- Brown believes enterprise customers are currently the strongest force pushing AI companies toward accuracy.
- Businesses paying for AI will not tolerate confidently wrong outputs in high-stakes settings.
- That economic pressure, she hopes, will keep labs focused on truthfulness rather than pure engagement.
Notable Takeaways
- AI is becoming a new information gateway, replacing some of the role once played by publishers and social feeds.
- The biggest risk is not just wrong answers, but wrong answers delivered with confidence.
- High-stakes topics require domain experts, not generic data labelers.
- AI companies need independent evaluation, not self-assessment.
- The future of trusted information may belong more to individuals than institutions.
- Consumer chatbot intimacy may grow, but Brown believes accuracy should remain the priority for the major labs.
Bottom Line
This conversation presents AI as both a threat and an opportunity for news, politics, medicine, and mental health. Campbell Brown argues that if AI is going to become the dominant layer for information, it must be held to a much higher standard than social media ever was. Her central message: the answer is not to avoid hard questions, but to build systems that answer them with expertise, context, and independent accountability.