3 ms·

> Hancock and his collaborators set out to explore this problem space by looking at how successful we are at differentiating between human and AI-generated text

• jnovek 4y ago
  According to academic friends of mine, tools like ZeroGPT still have too much noise in the signal to be viable way to catch cheaters. It seems to be better than these short form pieces or content, but if even if it’s “only” 80% accurate, some of those 20% will be false positives which is problematic.
  • rchaud 4y ago
    In an econometrics class in college, we have a team project and a final exam. The exam contained a question specific to the analysis method used in the team project. Answers to this question identified who genuinely worked on the project and who coasted on their team's work. Same thing can happen here: students can submit their term papers, but they have to do a 5-minute oral exam with an instructor or TA to discuss their paper.
  • PeterisP 4y ago
    Over the course of a year, I may get almost 500 assignments. If there is no reasonable way to verify if a submission flagged by a tool actually is AI-assisted or not (and IMHO there isn't), then even a 99% accurate tool is useless - I can't simply make 5 high-impact false accusations of innocent students each year, so these 'detections' are not actionable.
    • inimino 4y ago
      Or you could wait until you have 5 flagged submissions from the same student, and then, I don't know, apply human scrutiny and your own good judgment?
    • mcv 4y ago
      You could ask the the flagged students questions about what they wrote.
    • YeGoblynQueenne 4y ago
      Isn't it the same thing with plagiarism detection tools? Don't they also have false positives, too high to act on them? I don't know how that works in practice (I don't teach, though I've TA'd and noticed how there were clusters of assignments that were similar not only to each other, but also to the previous years' ones). So I'm curious what's the done thing. Does a tutor just confront a student with the "evidence" in the form of the plagiarism detector's score, say, and wait for the student to confess? How, exactly, can you prove not only whom, but how, wrote an assignment? I mean in the general case?
      • PeterisP 4y ago
        No, it's not the same thing because for plagiarism detection tools we don't really base a decision on the tool's output but rather compare the submitted document with the supposedly plagiarized original that the tool identified, and in those cases both verbatim copying without attribution and trivial changes can be identified by a human, the tool is just generating suggestions for human review. On the other hand, if a tool thinks that the text has an x% probability[1] of being AI-generated text, I have nothing to base a decision other than the output of that tool - if I'm not willing to fully delegate the decision to that tool (and I'm not) then I have no practical means to verify if the tool's suspicion is right. [1] And as someone working on machine learning, let me spare you a lecture on that a ML classifier saying "x%" can not be safely interpreted as x% probability, there's a million caveats that the vast majority of papers and tools don't try to address.
        • YeGoblynQueenne 4y ago
          Thanks for the clarification!
• ouid 4y ago
  I'm pretty sure the people who write short form marketing content don't pass the turing test either.
• civilized 4y ago
  > The study evaluated short-form generic marketing-style content, most of which is manicured and optimized to within an inch of its life. This is also the kind of human-written content that is closest to how LLMs sound. The tonal and structural similarity is so glaring that I have often wondered if a large percentage of the GPT training corpus is made up of text from spam blogs. I think if I was given, say, a couple pages from an actual physics textbook and then a GPT emulation of the same, I would be able to tell the difference easily. Similarly with poetry - GPT's attempts at poetry are maximally conventional and stuffed with flat and stale imagery. They can easily be separated from poetry by a truly original human writer. If AI developers want to impress me, show me an AI whose writing style departs significantly from the superficiality and verbosity of a spam blog. Or, in the case of Bing, an unhinged individual with a nasty mix of antisocial, borderline, and histrionic personality disorders.
  • rchaud 4y ago
    > The tonal and structural similarity is so glaring that I have often wondered if a large percentage of the GPT training corpus is made up of text from spam blogs. This is almost certainly the case, because the shifts in tone and vocabulary between an Inc.com or Buzzfeed article vs a London Review of Books article is far too wide to allow an AI to simply weigh them equally. AI speaks a kind of global English that's been trained on not just blogs and Wikipedia, but also Quora answers and content marketing pieces, a lot of which is written by non-native speakers. It isn't grammatically wrong, but as it targets the widest possible audience, its voice also isn't very interesting.
  • meh8881 4y ago
    You’re not interacting with the raw model. You’re interacting with a service that has intentionally designed it to work that way.
    • civilized 4y ago
      But if you ask ChatGPT to assume some other voice, it always just sounds like ChatGPT making a perfunctory effort to sound like something else, not actually like another voice. And from what I've seen of the raw model, when you ask it to depart from this voice, it can sometimes, but the bigger the departure, the more the results are weird and inhuman.
• shvedsky 4y ago
  Totally agree. Just yesterday, I was finishing up an article [1] that advocates for conversation length as the new definition of a "score" on a Turing test. You assume everyone is a robot and measure how long it takes to tell otherwise. [1]: http://coldattic.info/post/129/ http://coldattic.info/post/129/
  • meh8881 4y ago
    Such a metric is clearly useless if you cannot tell otherwise. I am very frustrated by the way this article repeatedly asks chatgpt to guess if something is a bot, gets told “well, we can’t know for sure but this is at least the sign of a crappy bot or human behavior” and then the author says “Aha! But a human could act like a crappy bot or a you could train a bot to mimic this exact behavior”. Well yeah. No shit.
    • 6510 4y ago
      I never imagined I could pretend to be a human. Thanks for the insight.
• dr_dshiv 4y ago
  Practically, long form content involves people. Can people tell on a sentence by sentence level what was written by humans and what by AI? My guess is that we all become more sensitive to this in a year or two. Look at how awful DALLE looks now, relative to our amazement last year.
  • tonguetrainer 4y ago
    DALL-E looks awful? I think results depend on the prompt modifiers you use. Personally, I'm happy with DALL-E, and generally prefer it to Midjourney.