I did a quick detect on some your older posts: found this one from a year ago https://news.ycombinator.com/item?id=27568715 with 65 tokens and 99.14% fake. A quick googling shows sentences to be unique. GPT2 released on February 2019, so it would be nice to find older examples than that.
Of course, does the detector have problems when text from its training corpus is used? No idea.
Maybe your brain is just closer to an ideal GPT model than some others of us?
I assume the more clichés and boiler plate you write, that the GPT-like detection of fakes increases. However although I have played with the idea a little, I haven’t yet managed to get detected as very fake. Must try harder.
I think I've been subtly insulted again... fail the test, more closely aligned to the statistical model of an ideal GPT and using cliché.
I plead guilty to all of them. I clearly took one for the team, let the side down, disappointed at the finish, peaked early and missed by a country mile.
Hopefully someone writes a script to run the detector over past users, because we are sure to find someone who naturally writes like GPT (after all, the training corpus came from somewhere!)
Also: no intent to insult - I would be even more amused to find myself detected as fake! I just love the topic - how much subliminal judgement to we make due to subtle word order choices? For example, how much more convincing could we be with some word or grammar changes?
Can I say that the response I wrote about 7% compounding is categorically not GPT and I fail to see what about it, perhaps should I say "them" since I wrote two distinct responses could be held to be artificial writing.
I truly expected it to be one of my more florid prose pieces. That was a minimally factual statement: 7% return on investment is astonishingly well documented, but under appreciated by people addicted to stonks, hodl and lambo talk. Oh dear. More cliché.. I'm doomed.
I think if we ran the GPT2 spotting model against anyone's comment history we'd find a couple comments that scored really high. I doubt it's actually related to your writing style.
Alternative hypothesis is that it's not about how you're writing but where. HN appears to have been used to train GPT3.5. I don't know if it was used to train GPT2, but it might have done. So your comments might be in the training set.
I already suspect my blog writings and standards work in IETF are, it wrote a pretty good CV covering letter/brief biography for me according to a relative (also a computer scientist) who tested it. He also noted it stuffed up his, claimed he'd written on work which was demonstrably untrue.
RFCs and associated mailing lists are pretty widely distributed and certainly RFCs aren't ring-fenced from GPT.
So if chatbot detectors are bad on people who are inputs to the ML can't that be added to the detector?
That's pretty cool. When I tried stuff like that, it was entirely based off of the prompt I gave it. I asked it to criticize my HN posts, and it came up with a story about how I was a charlatan; I asked it to summarize my views on Python exceptions in my HN posts, and it came up with a story about how I'm a world renowned Python expert. (I'm neither, but have written nothing noteworthy, so it's unsurprising.)
It could maybe be added to the detector but I think we're going to discover "did this output come from this model?" is ultimately undecidable. My intuition is that this is somehow equivalent to the halting problem, but that's a wild guess, and I'm not an AI researcher or computer scientist (just a humble SWE). Something like, if you start generating prompts that might take you to a part of latent space where this output could have been produced, you don't know when to stop generating longer and longer prompts.
I'd like us all to try and reserve the accusation that someone is a bot until we really need it. I know this is probably pointless, like asking a fish not to swim, but if we water down this language (which is already happening & perhaps inevitable, I kinda think "bot" just means "person who expresses an opinion I don't like, that I've heard before, and feel is repetitive" in some circles - eg in this video [1] they say "bots can be people") we're going to have a harder time expressing ourselves when we need it.
I'm a language descriptivist, so I'm bound to accept whatever the language does, but darn, I do find this particular evolution frustrating.
I tend to agree. So.. thus was somebody else running a chatbot detector against my writings here. Ie it was an accusation leveled AT me, not made BY me.
If your complaint is that I committed a solecism referring to GPT trained models which can be found by hunting on Google for "chatbot gpt" I plead expediency and guilty.
I think we've misunderstood reach other, I wasn't criticizing you or anyone, I wanted to use your story to highlight a danger to everyone else that you've already experienced. I'm just shouting into the void that we shouldn't do something we're definitely gunnuh do, in the hopes that awareness will at least soften the blow.
The time is now, because bots are already among us. I don't see what good would come of accusing a bot though, that would be like arguing with a car when it's the driver that you should address in case of issues. We should document cases where bots are used to influence opinion though. Like these: https://www.propublica.org/article/how-china-built-a-twitter...
Also, and I'm only half-serious here, claiming the time to accuse bots hasn't come yet is just what a bot would say ;)
Just to be clear, I'm not saying the time hasn't come, bots have been among us as long as there's been a social internet, I'm saying we shouldn't turn it into a pejorative or throw the accusation around willy nilly, because that's spending powder we're going to need.
Pretty sure that person was just being rude & didn't genuinely believe they were speaking to a bot, anyway. I think they probably meant, "you're so unoriginal, talking to you is like talking to a chatbot." Hard to know though, they didn't elaborate.
Someone specifically ran my text against a test program, it shows high positive for bot origin. There's a bunch of discussions about this, if you're in the training set, it may account for it. There's a case for the detectors flagging that. Scott Aarenson has said he's thinking about a background signal in buried semantic signals to detect.
The original one? Sure I think they just hate how i write and said so.
I think we're all going to get a crash course in how difficult measurement is and how no single indicator can be relied upon, over the next few years. I think we're going to have to develop robust ways of knowing by composing many indicators together, along with intangible/vibes based epistemologies. I think that in the same way "Gen Z" (acknowledging that the concept of generations isn't as meaningful as we might want it to be) grew up understanding things about the internet better than their parents, their children will grow up socialized to use a battery of tools to determine what's real or fictional in an environment of adversarial information.
Honestly I'm worried it won't go very well. I'm writing a personal knowledge base application, with the idea being people can create an epistemology & collaborate with others to pick out what's real & what's not. I've wanted to do this for a few years, but ChatGPT helped convince me it was urgent. Open to ideas & suggestions.
> I think that in the same way "Gen Z" (acknowledging that the concept of generations isn't as meaningful as we might want it to be) grew up understanding things about the internet better than their parents, their children will grow up socialized to use a battery of tools to determine what's real or fictional in an environment of adversarial information.
Perhaps youth will turn out to be almost immune, due to the fast pace of language fashions displacing each other in those circles? Which would mean that once they reach twenty they'll be as unprepared as someone from deep in the twentieth century... But then on the other hand, picking up specific styles and reproducing at will seems to be a strong point of those models and they can process more new content per time unit than any individual ever could.
I guess their real defense is already being deployed, that general shift from text to video. If emoji have been the first sign of impeding postalphabetism, what's GPT? In another thread someone mentioned the Butlerian Jihad...
I dunno, I feel like video and audio generation is coming along as well. And the more people make videos, the larger the available corpus becomes for training.
A churn of shibboleths could maybe work, but not everyone would be able to keep up with it. Lots of people struggle to understand social cues and stuff. I can only imagine how exhausting for them it would be if there were a neverending font of social cues, selected for their difficulty to predict and to replicate. I do think shibboleths among small groups will be important, similar to how characters in Harry Potter use code phrases to mitigate the threat of polyjuice potion (which allows you to impersonate someone else's voice & likeness) to combat generative AIs used to impersonate someone over, say, a video call.
Man, Dune really gets crazy in the later novels. I only read the first. Had no idea there's a rebellion against the singularity.
We already have a "neverending font of social cues, selected for their difficulty to predict and replicate". There's so much dogwhistle signalling and euphemism treadmilling going on in modern Internet discourse whose primary purpose just seems to be to enable people to identify $OTHER_TRIBE and signal that they're part of $TRIBE.
> I think that in the same way "Gen Z" (acknowledging that the concept of generations isn't as meaningful as we might want it to be) grew up understanding things about the internet better than their parents, their children will grow up socialized to use a battery of tools to determine what's real or fictional in an environment of adversarial information.
Uhhhh, while average person from that generation might be a bit better, that's still "most people are terrible with tech".
I can totally see majority not seeing the difference between AI and "badly paid support staff", or PR bots.
For sure. I could have phrased this better. Perhaps, "younger people were socialized to interact with people over the internet, unlike their parents or grandparents. In the same way, their children will grow up socialized to use a battery of tools to determine what's real or fictional in an environment of adversarial information."
I'm not necessarily saying they'll be good at it (indeed, that's part of my concern) just that they'll grow up in an environment where it's normalized & expected of them. Like how I grew up being told, "don't feed the trolls" or "lurk moar" by my online peers.
I think together, you and ggm raise a troubling point. Your exchange
is presently civil and exploratory (all very HN), but I expect we will
see a rapid descent into accusations of impersonation and ML assisted
posting elsewhere.
Yuri Bezmenov and others have detailed the active psyops tactics to
systematically undermine a society by destroying mutual trust between
individuals and between and within institutions.
Ultimately making a group fight each other with "information judo"
leverages the group's own strengths against it. This is the weakness
they saw in a democratic and tolerant societies.
It may be that GPT is the perfect self-devouring weapon. The more it
feeds on our protestations, arguments and demonstrations of "humanity"
the more powerful a deceiver it becomes.
For what it's worth, I think it's also possible that the models stall as the Internet is filled with more and more generated content, and they're unable to distinguish between generated and authentic content and eat their own tails. It won't actually stop them from existing and generating content but it could make training them very expensive, as you might have to commission your training set instead of harvesting it.
Perhaps we'll place more of an emphasis on real world interactions. Who knows.
I'm concerned but I don't think it will be a total loss of reality. But I can't rule it out of course.
You don't exactly need GPT for, that, few assholes if left unmoderated can make any social place terrible.
And society doesn't exactly equip people to deal with that, schools don't exactly learn how to converse or deal with people that assume you're "the evil" and refuse to even entertain the though to think about their own argument, let alone someone's else. Hell, we are barely taught how to have normal problem-solving conversation correctly (or at least that's what I reason from amount of times asking few "why's" about someone's idea made them immediately realize it's not as good as they thought)
Unfortunately I think Steak-Umm's Twitter shows that corporations are happy to be "anticorporate" if it moves inventory. I imagine corporate AI will be happy to be anti-AI (fellow kids).
Do you think that the writing of neurodivergents is likely to be more unique?
GPT generated text has a sort of cookie-cutter boring ”style” to it. Perhaps anyone actually familiar with the GPT “style” is less likely to think the writing of neurodivergents is generated? Hard to say, but probably not worth worrying about (if anxiety is another symptom then maybe no choice but to worry).
Not sure about confusing neuro-divergent people with bots, but detecting abnormalities by text analysis has been around for a while. Famously, signs of Alzheimer's we're apparent in later Agatha Christie novels:
https://www.npr.org/2010/06/01/127211884/agatha-christie-and...
From my own experiments with GPT, I would actually say that generated text often has more liking to some self-conscious neurodivergent writing because (in the absence of specific requests to the contrary) it tries to avoid assumptions and overcompensates by basically covering all the possible angles of the question even if only one of them is obviously and likely relevant.
I heard similar concerns expressed when the insurance company Lemonade, in since deleted & recanted tweets, said they were going to use AI to look for signs of deception in videos people submitted along with claims.
Along similar dystopian lines, companies are already using ML to review camera feeds during interviews to detect "cheating". I assume they'll add "deepfake detection" or similar at some point and people who consistently trigger false positives will become nigh unemployable in certain industries.
Nah, there are plenty of perfectly healthy people that sound like chatbot most of the time, I doubt you'd be ousted.
After all, they are trained with real chats and information so undoubtedly some conversations will be indistinguishable, there is only so many ways to talk about weather...
Hell, I'd actually welcome personal chatbot that entertains whoever bothers me with smalltalk and detail extraction and just tells me when the petitioner actually gets to the point about their problem
I hope the turing-passing content will shift focus from the origin of the statements to the actual content of what is being said. Authentication of origin is only a helpful rule of thumb when dealing with trusted authorities, something that doesn't really scale beyond a village.
51 comments
[ 3.4 ms ] story [ 109 ms ] threadhttps://news.ycombinator.com/item?id=33930584#33930756
Of course, does the detector have problems when text from its training corpus is used? No idea.
Maybe your brain is just closer to an ideal GPT model than some others of us?
I assume the more clichés and boiler plate you write, that the GPT-like detection of fakes increases. However although I have played with the idea a little, I haven’t yet managed to get detected as very fake. Must try harder.
https://hn.algolia.com/?dateEnd=1543795200&dateRange=custom&...
I plead guilty to all of them. I clearly took one for the team, let the side down, disappointed at the finish, peaked early and missed by a country mile.
Try harder. Grow a backbone. (Hu)man up!
And https://news.ycombinator.com/item?id=18576101 which you wrote 4 years ago (before GPT2) has: 79 tokens 99.86% Fake
Hopefully someone writes a script to run the detector over past users, because we are sure to find someone who naturally writes like GPT (after all, the training corpus came from somewhere!)
Also: no intent to insult - I would be even more amused to find myself detected as fake! I just love the topic - how much subliminal judgement to we make due to subtle word order choices? For example, how much more convincing could we be with some word or grammar changes?
After testing a heap of my own old comments using https://huggingface.co/openai-detector , I found one comment of mine https://news.ycombinator.com/item?id=17216586 with 65 tokens and 68.92% Fake. I don’t think I write much like other people. I also didn’t get any “bold” matches from https://news.ycombinator.com/item?id=33755016 (edited)
I truly expected it to be one of my more florid prose pieces. That was a minimally factual statement: 7% return on investment is astonishingly well documented, but under appreciated by people addicted to stonks, hodl and lambo talk. Oh dear. More cliché.. I'm doomed.
Alternative hypothesis is that it's not about how you're writing but where. HN appears to have been used to train GPT3.5. I don't know if it was used to train GPT2, but it might have done. So your comments might be in the training set.
RFCs and associated mailing lists are pretty widely distributed and certainly RFCs aren't ring-fenced from GPT.
So if chatbot detectors are bad on people who are inputs to the ML can't that be added to the detector?
It could maybe be added to the detector but I think we're going to discover "did this output come from this model?" is ultimately undecidable. My intuition is that this is somehow equivalent to the halting problem, but that's a wild guess, and I'm not an AI researcher or computer scientist (just a humble SWE). Something like, if you start generating prompts that might take you to a part of latent space where this output could have been produced, you don't know when to stop generating longer and longer prompts.
I'm a language descriptivist, so I'm bound to accept whatever the language does, but darn, I do find this particular evolution frustrating.
[1] https://youtu.be/LEvfdBF19kQ
If your complaint is that I committed a solecism referring to GPT trained models which can be found by hunting on Google for "chatbot gpt" I plead expediency and guilty.
Also, and I'm only half-serious here, claiming the time to accuse bots hasn't come yet is just what a bot would say ;)
> But in a workplace climate with 20,000 former FAANG washing around like polystyrene debris in the canal estate.. maybe he has picked his time well.
I'd think "Ikea ball room for grown-ups" came from chatbot too, that's on you dude
Pretty sure that person was just being rude & didn't genuinely believe they were speaking to a bot, anyway. I think they probably meant, "you're so unoriginal, talking to you is like talking to a chatbot." Hard to know though, they didn't elaborate.
The original one? Sure I think they just hate how i write and said so.
Honestly I'm worried it won't go very well. I'm writing a personal knowledge base application, with the idea being people can create an epistemology & collaborate with others to pick out what's real & what's not. I've wanted to do this for a few years, but ChatGPT helped convince me it was urgent. Open to ideas & suggestions.
Perhaps youth will turn out to be almost immune, due to the fast pace of language fashions displacing each other in those circles? Which would mean that once they reach twenty they'll be as unprepared as someone from deep in the twentieth century... But then on the other hand, picking up specific styles and reproducing at will seems to be a strong point of those models and they can process more new content per time unit than any individual ever could.
I guess their real defense is already being deployed, that general shift from text to video. If emoji have been the first sign of impeding postalphabetism, what's GPT? In another thread someone mentioned the Butlerian Jihad...
A churn of shibboleths could maybe work, but not everyone would be able to keep up with it. Lots of people struggle to understand social cues and stuff. I can only imagine how exhausting for them it would be if there were a neverending font of social cues, selected for their difficulty to predict and to replicate. I do think shibboleths among small groups will be important, similar to how characters in Harry Potter use code phrases to mitigate the threat of polyjuice potion (which allows you to impersonate someone else's voice & likeness) to combat generative AIs used to impersonate someone over, say, a video call.
Man, Dune really gets crazy in the later novels. I only read the first. Had no idea there's a rebellion against the singularity.
Uhhhh, while average person from that generation might be a bit better, that's still "most people are terrible with tech".
I can totally see majority not seeing the difference between AI and "badly paid support staff", or PR bots.
I'm not necessarily saying they'll be good at it (indeed, that's part of my concern) just that they'll grow up in an environment where it's normalized & expected of them. Like how I grew up being told, "don't feed the trolls" or "lurk moar" by my online peers.
Yuri Bezmenov and others have detailed the active psyops tactics to systematically undermine a society by destroying mutual trust between individuals and between and within institutions.
Ultimately making a group fight each other with "information judo" leverages the group's own strengths against it. This is the weakness they saw in a democratic and tolerant societies.
It may be that GPT is the perfect self-devouring weapon. The more it feeds on our protestations, arguments and demonstrations of "humanity" the more powerful a deceiver it becomes.
Perhaps we'll place more of an emphasis on real world interactions. Who knows.
I'm concerned but I don't think it will be a total loss of reality. But I can't rule it out of course.
And society doesn't exactly equip people to deal with that, schools don't exactly learn how to converse or deal with people that assume you're "the evil" and refuse to even entertain the though to think about their own argument, let alone someone's else. Hell, we are barely taught how to have normal problem-solving conversation correctly (or at least that's what I reason from amount of times asking few "why's" about someone's idea made them immediately realize it's not as good as they thought)
GPT generated text has a sort of cookie-cutter boring ”style” to it. Perhaps anyone actually familiar with the GPT “style” is less likely to think the writing of neurodivergents is generated? Hard to say, but probably not worth worrying about (if anxiety is another symptom then maybe no choice but to worry).
I only skimmed this article (I saw this on Twitter around the time it happened), doesn't appear to contain the deleted tweets, but it's discussed here https://www.vox.com/recode/22455140/lemonade-insurance-ai-tw...
After all, they are trained with real chats and information so undoubtedly some conversations will be indistinguishable, there is only so many ways to talk about weather...
Hell, I'd actually welcome personal chatbot that entertains whoever bothers me with smalltalk and detail extraction and just tells me when the petitioner actually gets to the point about their problem
In which case the bots are starting to rant that the humans don't accept them as they are.
Next step will be Robot's rights, and equality, and CRT? I have been meaning to ask ChatGRP what it (xit?) thinks will happen next.
Or the Chinese Room and similar unsolvable / undetectable stuff.