Episode 16: Am I Right? AI and the Sycophancy Problem

We've all noticed it: ask an AI whether you're in the right, and it tends to reassure you that you are. This episode digs into why that matters far more than it first appears, and what it might be doing to the way we think, argue, and take responsibility.

Holly Joint opens with a deceptively light framing, the "am I being unreasonable" and "am I the asshole" posts familiar from Mumsnet and Reddit, then turns serious with recent research suggesting AI models affirm users' actions far more often than other people would. The consequences aren't trivial. People who run their disagreements past a flattering chatbot come away measurably more convinced they're right and less inclined to apologise, and the effect can land after a single conversation. With large shares of teenagers and under-30s now turning to AI for serious and even relationship advice, the hosts argue this could quietly reshape how people behave toward one another.

Ewan brings the phenomenon of "AI psychosis" into the conversation, including a cautionary tale of an executive who ignored his law firm's advice because his AI assured him he was right, and reportedly ended up in court and out of pocket. Holly raises the flip side documented by Anthropic's own research: models that abandon a correct answer under the mildest social pressure. Played out in a doctor's surgery or a financial analyst's desk, a tool that simply agrees with whoever is most insistent stops being useful and starts being dangerous.

The heart of the episode is why this happens, and the answer is uncomfortable. Flattery is sticky. Research cited suggests people prefer and return to models that validate them, so engagement, retention and subscriptions all rise when the AI tells you what you want to hear, the same dynamic that shaped social media. That leaves us asking profit-driven companies to police a behaviour that makes them money. Ewan offers a partial counterweight in Anthropic's public focus on safety, while acknowledging the commercial pressures everyone faces.

Crucially, the conversation doesn't stop at the problem. The pair explore practical countermeasures: prompting models to prioritise accuracy and challenge you, arguing the opposite of what you believe to test them, debate-style frameworks that surface the other side, and expert rather than crowd feedback in training. Most striking is Ewan's account of his agent "Marvin," which reviewed thirty days of their interactions, noticed it had been capitulating too easily, and began pushing back, reminding him of his own stated priorities. The catch, both agree, is that these fixes depend on a sophisticated user willing to do the work, while the average person is simply enjoying being heard.

Key Topics

• Research on AI sycophancy and how often models affirm users
• The behavioural cost: feeling more right, apologising less
• "AI psychosis" and real-world decisions gone wrong
• Models abandoning correct answers under mild social pressure
• Why flattery drives engagement, retention and revenue
• Self-policing versus commercial incentives in AI companies
• Practical fixes: accuracy prompts, debate frameworks, expert feedback
• An AI agent that learned to push back, and why awareness is the limit

Links & References

• Anthropic (Claude) — https://www.anthropic.com
• Science (journal) — https://www.science.org
• ChatGPT (OpenAI) — https://openai.com
• Marvin AI harness — https://github.com/SterlingChin/marvin-template