In all the debates about human and machine intelligence, one thing has become clear: We humans love to talk about AI. We’ve speculated whether AI will create or eliminate jobs, revolutionize or destroy science, comfort or torment us, entertain us brilliantly or rot our brains, or some combination of all these things. But amid the noise, we may have paid too little attention to a subtler impact: how AI is disrupting our relationship to information and making it harder to know what to believe.
The problem isn’t just deepfakes, the malicious manipulation of images, video, and audio. Even benign, well-intentioned AI applications can be profoundly misleading. The nature of the large language models that power chatbots means that they are always guessing, however confident they seem. How can we use AI to produce reliable information that improves public discourse rather than distorting it?
To explore this dilemma, I moderated a webinar on Trust in Information in the Age of AI, co-sponsored by NAPA, CODE, and the Bridge Alliance, with four leaders in the field: Richard Coffin of USAFacts, Professor Renée Cummings of the University of Virginia, Meghan Herbst of WIRED, and Dr. Alan Shark of NAPA and George Mason University. I also interviewed Claude (Opus 5.5 Extra), which assessed its own limitations and offered some useful advice.
The human experts and the machine intelligence agreed on several points.
A significant AI trust gap exists. Several surveys show that more Americans use AI, but few trust its results. However, many of us still have no firm opinion on AI, suggesting that trust in AI will depend on how it performs in the future.
Our skepticism is justified: AI often gets things wrong. Large language models don’t actually “know” anything – they answer queries based on patterns they deduce from the data they’ve been trained on. While accuracy rates depend on the task and how you measure success, rough estimates hold that chatbots make errors about half the time. The general consensus is that these AI hallucinations can be reduced but may be impossible to eliminate.
AI is often biased in ways that can be hard to detect. Algorithmic bias has been a concern for years, long before today’s chatbots were developed. Because AI is trained on historical data, it can reinforce old biases about gender, race, or ethnicity in hiring, criminal justice, and other areas. This can be a serious issue in any AI application used for public policy.
AI doesn’t quite know how to use government data. Chatbots often rely on secondary sources even when more reliable primary government data is available. The problem stems from the fact that much government data isn't “AI-ready,” and AI models aren't designed to go to government sources first, even when they are fit for use.
AI’s impact on information needs regulation. The debate over AI regulation has focused on the big picture: the need for guardrails to ensure AI doesn't turn on or even destroy humanity. While that priority is hard to argue with, we also need regulations to prevent AI from misleading or misinforming us. The European Union now requires a watermark on images or text produced by AI. In the U.S., several states are moving to prevent bias in public uses of the technology.
AI may offer real benefits for all kinds of human endeavors, but its flaws remind us to treat the technology with caution and care. This is especially true for using AI in public policy and decision-making, where poorly informed plans can have major repercussions.
For now, the best advice is to use AI with an awareness of its shortcomings, don’t put human judgment aside, and ask for and review your chatbot’s sources. As Claude said about our interview at the end: “If your readers take away one habit from it, I hope it’s checking the sources I’ve cited rather than taking my word for any of this.”
You can watch the full video of the webinar “Trust in Information in the Age of AI” on YouTube at: https://napawash.org/event/napa-technology-leaders...
Joel Gurin is the president and founder of the Center for Open Data.
