AI technologist @deepfates argues critics now reject working AI detectors despite trusting early, ineffective versions
Newer detectors like Pangram have become highly effective.
Positive users praise Pangram AI Detector for spotting obvious AI output and learning from false negatives, while negative users call it ineffective cope that invites lawsuits and reflects bias.
No Digg Deeper questions have been answered for this story yet.
Most Activity
@deepfates im working on a creative writing pipeline and it fools pangram almost all of the time. im very proud of this because it is Only Prompting
@deepfates >Confidence High I've Fooled the Fox
@usr_bin_roygbiv Maybe you should look at the claims they actually make about what it does before saying stuff like this
@dioscuri I'm p skeptical of using it for anything with actual/regulatory consequences right now to be fair. i think it's useful for creating a verifier for social reinforcement, basically. a common reference frame that's better than random guessing. so that we can develop herd immunity
@deepfates Yes, have had a similar update (and reaction!).
@usr_bin_roygbiv are you familiar with the difference between false positives and false negatives? Your overconfidence here is making me doubt your strong opinions in other areas I previously believed you about.
@viemccoy Nice! I think basically all the cases where it has a false negative are useful and interesting to learn from. But the true positives are very useful as well
@deepfates I do not believe that pangram works and never have. You would have to distill the weights on every model and fine tune. I could believe that it works on opus, gpt, and nothing else
@usr_bin_roygbiv what you're saying is that you would rather continue to be ignorant, and lose my respect, than attempt to understand how you might be wrong 👍
@full_kelly_ it's more because I understand how things work and update my mental model based on the data I receive
@deepfates Do you think people have developed more sophisticated ways of masking the LLM smell now so Pangram (and we) don't detect it?
Seems like it would be easy-ish to build a pipeline to do that.
@usr_bin_roygbiv No shit! That is literally not the threat model. The point is to be able to believe the positives, not the negatives. Can you see why it would be useful to know whether something has definitely been written by opus or GPT
@deepfates pangram is low precision, high recall
@deepfates This is low-key a modern Seinfeld plot.
@viemccoy @deepfates I'm working on this too, but you're clearly better than me at it. I'm at least getting good pros that I enjoy out of the process.
Before you tell me that pangram does not work please describe what problem you think it is failing to solve. TIA
One very funny B plot of the last few years is that AI detectors did not work before pangram but anti-AI people would not believe me. Now pangram does work but anti-AI people have finally updated to believing that none of them work
@sean_from_earth Sure, but if someone's going to that much effort I'm more interested in reading it. I just want to filter out the main problem content
@deepfates Everyone ive sent that pope tweet too is coping hilariously: "no dude ai is just trained off religious writing"
@viemccoy @deepfates https://nickblow.tech/posts/on-ai-writing
I have an example of AI text that fooled it as well, but I brute forced it by shoving 40k of my own words into context, so it’s “only prompting” but also incredibly dumb and labour intensive.
(It was fun though).
@deepfates Yes, agreed. I think if e.g. a university wanted to use it for enforcement (not sure they should), I'd expect them to only go after people who got repeatedly flagged. And ofc it's got a very high false negative rate, so it's not guarantee something is human written.
@usr_bin_roygbiv you didn't even read the paper or understand the actual problem space or what they claim to do. So you are just literally making yourself dumber. Have a nice day.