‹ BackHN Continuity

Thread

Uproar in France over award-winning author accused of using AI

36 points · 87 comments · harscoat

  1. _fw · · focus · HN ↗
    I remain convinced the whole notion of AI-detection algorithms with any semblance of precision and accuracy is a con.

    Critics say the author almost exclusively relied on AI to create his work, based on Pangram’s findings.

    Pangram says it “is trained specifically to detect the latest AI models from frontier labs like OpenAI, Anthropic, Google, Meta, DeepSeek, xAI, and more. [They] benchmark against 26 models (so far)…”.

    I can probably list more than 26 models off the top of my head. Granted, some models have been distilled against others and will be “detectable” but that’s NOWHERE near a large enough sample size to be reliable in 2026.

    You or I could get a flag by that detector in our first attempt, and it wouldn’t take long for a false positive against enough attempts of truly original thought either.

    The only thing I’d actually believe is things like Claude’s text watermarking system. However, those systems will never be present for EVERY model, so reliable blanket detection remains an impossibility.

    The entropy in written text of meaningful length is simply too great for detection to be reliably possible in 2026, especially considering the changes the author can make.

    1. kuerbel · · focus · HN ↗
      best blog post I know about this: <a href="https:&#x2F;&#x2F;freddiedeboer.substack.com&#x2F;p&#x2F;i-wouldnt-say-pangram-is-broken-but" rel="nofollow">https:&#x2F;&#x2F;freddiedeboer.substack.com&#x2F;p&#x2F;i-wouldnt-say-pangram-i...

      &quot;Experimenting, I’ve also found that I can pretty reliably get Pangram to declare large sections of text to be AI-written or human-written depending on the larger textual context I place those sections in - that is, I can take AI-generated text and embed it in human-generated text and have Pangram declare it 100% human, and I can take human-generated text and embed it in AI-generated text and have Pangram declare it 100% AI. I’ve also found that it’s pretty easy to induce false positives, that is, to write texts myself that Pangram identifies as AI-written. All of this makes me hesitant about the Pangram tool, and in particular the way many people use it, as a one-shot gotcha machine.&quot;

      Sounds like it may be very easy for someone motivated enough to find sections, even longer sections, that turn out as 100% AI written.

      1. TeMPOraL · · focus · HN ↗
        &gt; Sounds like it may be very easy for someone motivated enough to find sections, even longer sections, that turn out as 100% AI written.

        It&#x27;s probably as easy as prompting Claude, &quot;Take text from this markdown&#x2F;PDF file and run it through Pangram tool, figure out a way to, and find, the longest passage from the text that Pangram flags as 100% AI generated&quot;, and leaving it spinning in Auto mode for an hour.

    2. bryanlarsen · · focus · HN ↗
      &gt; The entropy in written text of meaningful length is simply too great for detection to be reliably possible in 2026

      Pangram is unreliable on short text, but much more reliable on long form text.

    3. jsnell · · focus · HN ↗
      The impact from a model not being included in the training set is increased false negatives for text generated by that model, not increased false positives for text written by humans.
      1. brookst · · focus · HN ↗
        Source?
Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.