Reading for “When AI is wrong, and how to tell”

For the $25 webinar on Tuesday, November 10, 2026: When AI is wrong, and how to tell. You don't need to read any of it first.

1. Claude is providing incorrect or misleading responses. What's going on?

   Claude Help Center (Anthropic)

   A vendor saying plainly that its assistant can be wrong, and that you shouldn't treat it as your only source. Short, and a good one to share.

2. Reduce hallucinations

   Anthropic (Claude docs)

   Techniques that work in any assistant: let it say “I don't know”, ask for direct quotes from your document, and make it back each claim with a source.

3. Why language models hallucinate

   arXiv (paper by OpenAI researchers)

   The argument that models guess because training and testing reward a confident guess over “I'm not sure”. The abstract alone explains a lot.

4. View related sources from Gemini Apps

   Gemini Apps Help (Google)

   Where Gemini shows the sources behind an answer, and that not every answer has them. The habit to build in any tool: find the source, then open it.

5. Application card: Microsoft Copilot (for organizations)

   Microsoft Learn

   Microsoft's own account of Copilot's limits, including answers not supported by the sources and the risk of relying on output that's hard to check. Skim the limitations and best practices sections.

6. Generative AI Profile (NIST AI 600-1)

   NIST

   The US standards body's list of generative AI risks, with made-up content (“confabulation”) near the top. Dense, but the risk list is worth a look if you set rules for others.

7. BBC research shows issues with answers from AI assistants

   BBC

   BBC journalists checked news answers from four popular assistants and found errors and altered quotes. A clear example of why a source link isn't proof.

8. AI search has a citation problem

   Columbia Journalism Review, Tow Center

   Researchers asked AI search tools to name the source of real news articles. The tools were often wrong, rarely said so, and sometimes made up links.

9. AI on trial: legal models hallucinate in 1 out of 6 (or more) benchmarking queries

   Stanford HAI

   Even specialist tools built on trusted documents still made things up. Worth remembering when anyone tells you a product never makes things up.

10. Civic Online Reasoning

    Digital Inquiry Group

    Lessons on how fact-checkers judge a source: lateral reading, asking who's behind it and what the evidence is. Made for classrooms, and just as useful for checking what an AI hands you.

Every link was opened and checked on October 24, 2026. Each one is public, with no paywall. None of these sources is connected to the workshops.