The rise of AI for information-seeking has prompted a parallel rise in concerns about related risks. Such risks can be highly individual, such as psychological or medical impacts or malicious use by bad actors. But there are also concerns about societal risks if systemic problems compound: for instance, if political leanings or source selection in AI answers mediate how citizens receive news or political information en masse.Such risks have also led to a rise in auditing of chatbot responses. By the standards of technical research, the basics are relatively straightforward. A user can prompt an AI tool and assess its answers on various axes (inaccuracies, bias, sources used, etc.). There are various methods for scaling up both the prompting and analysis, with advantages and disadvantages: for instance, automating answers via official APIs may not be fully reflective of answers chatbots give when used normally, but “unofficial” automating is often blocked by companies. Nonetheless, general barriers to entry for this sort of research are relatively low, and we have seen a proliferation of such audits.However, auditing answers to prompts gives only limited insight into societal risks. Even if some sets of prompts gave potentially harmful answers, this would have limited impact if no real person was likely to ever make those prompts. On the other hand, prompts which give potentially harmful answers one percent of the time could be high-risk if made by millions of…

Because you read this