The tools we're building to make knowledge more accessible might be making us easier to mislead.
The Summary
- A KFF poll of 2,480 US adults found frequent AI chatbot users are more likely to believe vaccine misinformation, including debunked claims like vaccines causing autism
- The correlation held even after controlling for age, race, education, and political partisanship
- This isn't about what chatbots say. It's about who's using them for medical advice and why.
The Signal
The May 2025 survey from health research firm KFF reveals a pattern that should worry anyone building in the agent economy. Adults who frequently consult AI chatbots for health advice are more likely to believe false claims such as vaccines causing autism or the measles vaccine being more dangerous than measles itself. The correlation persisted across demographic controls.
This is not a chatbot accuracy problem. This is a selection problem. The people turning to AI for medical guidance instead of doctors aren't getting bad answers from bots. They're already looking for alternatives to mainstream medical advice.
"The correlation remained while controlling for factors such as age, race, education and political partisanship."
Think about the user journey. Someone skeptical of institutional medicine doesn't book a doctor's appointment. They open ChatGPT at 11pm. The chatbot becomes the path of least resistance for people who've already decided traditional sources aren't trustworthy. The tool didn't create the distrust. It gave it a friction-free outlet.
Key pattern emerging:
- AI chatbots lower barriers to information-seeking behavior
- People with institutional distrust find AI more neutral/accessible than doctors
- Confirmation bias gets turbocharged when you can iterate queries until you get the answer you want
This matters for Web4 because we're building agent systems that will handle increasingly complex personal decisions. Health, finance, legal advice. If the early adopters of AI-for-everything are disproportionately people seeking alternatives to expert consensus, we're not democratizing expertise. We're automating confirmation bias at scale.
The Implication
If you're building consumer AI agents, understand who your actual users are versus who you think they are. The people most eager to delegate decisions to AI might be the ones least equipped to validate its outputs. Design for skepticism, not efficiency.
For everyone else: if you're using chatbots for medical questions, you're not getting a second opinion. You're getting a language model trained on the entire internet, including the parts that are confidently wrong. The agent economy only works if we build better human judgment alongside better agents.