‘But there is a difference between recognising AI use and proving its use. So I tried an experiment. … I received 122 paper submissions. Of those, the Trojan horse easily identified 33 AI-generated papers. I sent these stats to all the students and gave them the opportunity to admit to using AI before they were locked into failing the class. Another 14 outed themselves. In other words, nearly 39% of the submissions were at least partially written by AI.‘

Article archived: https://web.archive.org/web/20251125225915/https://www.huffingtonpost.co.uk/entry/set-trap-to-catch-students-cheating-ai_uk_691f20d1e4b00ed8a94f4c01

  • definitemaybe@lemmy.ca
    3·
    8 hours ago

    it should not be used to seek out factual information that the prompter doesn’t already know.

    Eh… Depends on the importance and purpose of the information.

    If you’re just trying to generate ideas for fiction from historical precedents, it doesn’t matter if it’s accurate. Or if you’re using it as a starting point, then following the links to check the original source (like I do all the time for Linux terminal commands).

    Hell, I often use Linux terminal commands from Google’s search results AI box—I know enough to be able to parse what AI is suggesting (and identify when the proposed commands don’t make sense), and enough to undo what I’m doing if it doesn’t work. Saves a lot of time.

    Copilot fixed some SQL syntax issues I had yesterday, too. 100% accuracy on that, despite it being a massive query with about a dozen nested subqueries. (Granted, I gave a very detailed prompt…) But, again, this was low stakes–who cares if a SELECT query fails to execute.