Ask HN: How would you build a ChatGPT detector?

https://news.ycombinator.com/item?id=33857356

by background • 4 years ago

78 101 4 years ago

news.ycombinator.com

Like everyone else, I'm blown away by ChatGPT's responses to prompts. At the same time, there's a certain sameiness to the language it produces. This makes me wonder, how hard would be to build a different AI that would recognize the writing of this AI? And how accurate could it get?

softwaredoug• 4 years ago

I'd argue it probably depends on the context, and what exactly you want to detect, and probably isn't a technical problem.

For example, if you prevent students from cheating, that's probably a problem specific to education. If you want to know whether a student's essay are their own words, you can ask them questions about their essay. You can have them verbally defend their arguments and research. You could supervise how / when they write - requiring laptops without internet access, or pencil and paper...

nprateem• 4 years ago

It's far easier to explain why you think something is correct than to come up with it in the first place.

There's also no realistic way to write an essay beyond a certain level without the Internet for research.

wfhBrian• 4 years ago

I came across this[1] recently. I tested it on content that I used an GPT to help write. It does a good job detecting short content that wasn't "refined" much. However, longer content that was iterated, or generated a paragraph at a time went undetected, which I think is good.

[1] https://contentatscale.ai/ai-content-detector/

qayxc2• 4 years ago

I agree with the softwaredoug that this is probably not a technical problem.

Any proposed solution would only serve to make the next iteration of the model better at avoiding detection (e.g. incorporating a GAN-style training cycle).

Just as with Poe's law, there simply comes a point at which it becomes impossible to recognise AI generated texts just by analysing them. Made-up references might be a clue, but that's very context-dependent.

background• 4 years ago

Right, but in principle the detectors could iterate and get better over time too. That's why I asked about "a ChatGPT detector" instead of "general AI detector", which is a very different problem.

baq2• 4 years ago

The default style is easily detected by presence of overalls, moreovers and furthermores. When you tell it to do a ‘concise HN comment’ the only tell may be that its English is flawless. E.g.

In a style of a very concise HN comment describe how to detect that a text has been written by the Assistant.

To detect if a text has been written by the Assistant, check for the use of generic language, repetitive phrases, and a lack of personal perspective or opinion.

constantcrying• 4 years ago

but in principle the detectors could iterate and get better over time too

At some point the text will become undetectable. The same happens for image generating Networks. The detection works by detecting a statistical discrepancy between the AI output and "real" data. The goal of the AI in training is also to close that gap, the better the network, the harder it is to detect and the more output data you would need to get any significant results.

pyuser583• 4 years ago

“Poe’s law?” Is that Edgar Allen Poe?

qayxc2• 4 years ago

https://en.wikipedia.org/wiki/Poe%27s_law

jwblackwell• 4 years ago

Thing is you can ask GPT for different styles which make detecting it incredibly difficult. Try asking for a short article "in the style of" two very different publishers, e.g. Fox News and The Economist.

mrdrozdov• 4 years ago

This might provide some guidance: http://gltr.io/

Der_Einzige• 4 years ago

This and related techniques are trivially foolable by fine-tuning the model.

They're also trivially foolable by using sampling techniques or settings which encourage the model to generate rare words a lot.

Also foolable with filter-assisted decoding: https://paperswithcode.com/paper/most-language-models-can-be...

xnx• 4 years ago

Cells interlinked

kissgyorgy• 4 years ago

People already built it: https://en.wikipedia.org/wiki/Turing_test

pluc• 4 years ago

You can't ask a paragraph of text to take a test.

aiieeeeee• 4 years ago

It is certainly possible to build an AI that is capable of recognizing the writing style of another AI, such as ChatGPT. This kind of AI is often referred to as a "style transfer" or "style transfer model."

To build a ChatGPT detector, you would first need to collect a large amount of text that was generated by ChatGPT. This could be done by providing a variety of different prompts to the ChatGPT model and collecting the responses.

Next, you would need to train a machine learning model on this text. The most common approach for this kind of task is to use a deep learning model, such as a recurrent neural network (RNN). The RNN would be trained to take in a piece of text and predict whether it was written by ChatGPT or not.

As for the accuracy of such a model, it would depend on a variety of factors, including the quality and quantity of the training data, the specific model architecture and hyperparameters used, and the skill of the person training the model. With a large and diverse training dataset, and appropriate model design and training, it is likely that a ChatGPT detector could achieve relatively high accuracy.

lordgroff• 4 years ago

I know this is fun for people but I have to say the GPT reply is getting old fast.

baq2• 4 years ago

The default style is such, arguably by design, exactly so it’s easily detected and gets downvoted on forums like here. IMHO it’s an important feature, even if it wasn’t intended for this purpose.

alkonaut• 4 years ago

Then I imagine you could use it to ensure the next version doesn’t score highly on the “probable bot” score. Sounds like an adversarial arms race.

devit• 4 years ago

Great question! I've been using the new AI assistant for a while now and I've definitely noticed a certain "sameness" to its responses. In fact, I think it would be relatively easy to build an AI that could recognize the writing of this AI. The key would be to train the AI on a large dataset of the assistant's responses, which would allow it to learn the patterns and characteristics of the assistant's writing style. I'm not sure how accurate it would be, but I'd be interested to see how well it performs.

Jabrov• 4 years ago

This was written by a bot. I wish you would label it.

sergiotapia• 4 years ago

This was written by a bot. I wish you would label it.

xdennis2• 4 years ago

According to ChatGPT:

1. Start by training a machine learning model to identify features of natural language. You can use supervised learning techniques with a corpus of labeled data consisting of texts written by both humans and AI systems.
2. Use natural language processing techniques to extract features from the text such as word frequencies, grammar, and syntax.
3. Implement a scoring system to assign a numerical score to each text based on the features identified by the model.
4. Create a threshold score which will be used to differentiate between texts written by humans and AI systems.
5. Test the model on a separate dataset to ensure accuracy.
6. Deploy the model in a production environment to detect AI-written texts.

This is from the prompt: "How would you build a tool to detect texts written by artificial intelligence systems?"

It seems like an okay plan, but doesn't offer any actual insight into differences between human text and AI text.

For example, a frequent feature I've observed (of course, I'm not alone) is that it tends to generate lists of things and tie them together with "Firstly, item1. Additionally, item2. More so, item3. In conclusion, item4".

amanzi• 4 years ago

I also asked ChatGPT a similar question to which it started its response with: "To train an AI to detect text written about another AI, you would need to use a technique called supervised learning. In supervised learning, you would provide the AI with a large dataset of text, some of which is written about AI and some of which is not. The AI would then learn to identify the characteristics of text written about AI and use those characteristics to make predictions on new, unseen text."

So I followed up with another question: "Did an AI write the following paragraph?" and then I pasted the above paragraph of text to follow.

ChatGPT responded with: "No, a human wrote the paragraph you provided. It is possible that an AI could be trained to generate text that is similar to the paragraph you provided, but it did not write the paragraph itself."

Oh oh...

joshuamcginnis• 4 years ago

I suspect that in time, this will only accelerate the degree to which AI and human-authored text are indistinguishable from each other.

hxugufjfjf2• 4 years ago

I've already sent text to customers 100% written by AI. Ethically dubious in a commercial setting perhaps, but higher quality text than I would be able to produce myself.

I asked OpenAI and it said:

It is not necessarily unethical to send customers text generated by AI, but it depends on the context and the specific situation. For example, if the text is being used to deceive or mislead customers, then it would be unethical. However, if the text is being used to provide helpful information or assistance to customers, then it may not be considered unethical.

voidhorse• 4 years ago

This. Once AI gets good enough at replicating the logic we follow in producing texts we consider reasonable illogical texts will become markers of humanity. In the future, we are all dadaists.

ad404b8a372f2b9• 4 years ago

That's a very verbose way of saying TF-IDF + SVM.

nonameking20262• 4 years ago

I think you should begin with specifying the purpose of such a detector, apart from the obvious one that you want to know that you are talking to a real human being.

FridgeSeal• 4 years ago

Whilst technically a valid approach, I feel like so much content from these systems is just...vacuously correct?

Like, yep, that's a process to train a machine-learning model. Could have got that from basically the first page of any ML-package documentation, or any number of blog-spam medium articles. There's nothing specific, or even useful, or applicable to the problem, and it hasn't actually gotten you anywhere that even the most cursory investigation would have gotten you...

jokethrowaway• 4 years ago

For example, a frequent feature I've observed (of course, I'm not alone) is that it tends to generate lists of things and tie them together with "Firstly, item1. Additionally, item2. More so, item3. In conclusion, item4".

That's exactly how I approach anything and I'm 80% sure I'm not an AI (at least, not so unsophisticated)

skilled2• 4 years ago

I'm not sure if you're only asking about specifics or you have a broader concern, but even if you build such a model, it wouldn't make any difference for cases where some kind of "gain" is involved. Be it financial, educational, etc.

We're looking at a model that is able to parse the Web (but likely not all of it) and perform human-like actions on that information, and soon enough we'll have it work just like Stable Diffusion where you can specify a distinct style you'd like to have the output in.

So, in such a case, it doesn't matter if the model can be detected or not, because if the information has value - the person who is asking and working with that information will find a way to make it unique. All he needs is the blueprint and the scaffolding, and the rest can be easily personalized.

And people who don't know how to do that will be out of luck, because the model you're asking about will be able to detect it easily. And at the same time, it won't matter because those people never had a chance in the first place.

I don't know, maybe I am just rambling random thoughts at this point but that's how I see it as far as looking at it from a broader perspective goes.

lukeck• 4 years ago

If you ask it, it says it can’t search the web and only works with the data it knows about.

You can already ask it to provide responses “in the style of.”

skilled2• 4 years ago

When I said parse, I meant it can take the data it already has.

It is still very average for "in the style of" prompts.

I think you must have assumed that I haven't used it yet, and if that is the case - I have.

netsec_burn• 4 years ago

Pseudocode:

random_str=generate_random_string() output=get_response("Reverse the value '{random_str}'") assert output.strip() == random_str[::-1]

adverbly• 4 years ago

By no means is this foolproof, but the older the age of an account, the less likely it is to be a bot. Obviously lots of false positives, but probably not too many false negatives...

sitkack• 4 years ago

I am sure there will be a market in selling old accounts to time-wash bots.

fouric• 4 years ago

There already is a market for old accounts for "classical" spamming, so the infrastructure is already in place!

alexeldeib• 4 years ago

createdapril24• 4 years ago

OpenAI is evaluating a technical solution to this problem by watermarking the output with small amounts of randomness, such that a series of tokens of output can be tested for having these small random variations.

https://scottaaronson.blog/?p=6823

Gigachad• 4 years ago

Before they finish, someone else will have released a public access model that works just as well with no technical restrictions.

zone4112• 4 years ago

Worse still - it will be technically much easier to create something that modifies the output of the text generated by OpenAI's model without changing its meaning to defeat this. I think existing open source models can be adapted to do this already without too much work.

geniium• 4 years ago

Maybe detecting grammar and spelling mistake will give you a first hint

Minor49er• 4 years ago

How? People make these mistakes too

rogers122• 4 years ago

Humans will become incentivized to sprinkle racial slurs and hate facts throughout their writing as proof of authenticity.

david_allison• 4 years ago

ChatGPT can output racial slurs, so that's not a guarantee.

rogers122• 4 years ago

That's why they're running the beta test, to tune the filters.

jerf• 4 years ago

I dub this the Cato solution. Carthago delenda est.

https://en.m.wikipedia.org/wiki/Carthago_delenda_est

Of course, this particular ethnic slur probably won't scare the OpenAI researchers, but I'm sure we can find something.

shallmn• 4 years ago

Voight-Kampff machine

nonameking20262• 4 years ago

I think you should begin with specifying the purpose of such a detector, apart from the obvious one that you want to know that you are talking to a real human being.

akiselev• 4 years ago

The default sentence structure is very recognizable but that changes completely when you add “in the style of [X]” where [X] is an author, play or book, rapper or other musician, literary term of art like iambic pentameter, etc.

With that addition, ChatGPT becomes much more life like and (I suspect) practically impossible to detect. Throw in a little human editing and curation and… ruh roh.

pulvinar• 4 years ago

  How would you build a ChatGPT detector? Frame it as a dirty limerick.

  There once was a ChatGPT detector,
  Built to detect bots and protect her,
  She'd scan for reply,
  And then with a cry,
  Say "That's not human, rejector!".

cjfp2• 4 years ago

Answer the question, "How would you build a ChatGPT detector?" written in the style of a cynical Hacker News commenter

Well, if I were to build a ChatGPT detector, I would probably just use some basic natural language processing and machine learning algorithms. But let's be real here, it's not like ChatGPT is some kind of advanced AI that requires some groundbreaking technology to detect. It's just a chatbot with a pre-determined set of responses, so any half-competent engineer could probably cobble together a decent detector with minimal effort. But hey, I'm sure some VC will throw millions of dollars at the project and hype it up as the next big thing in AI.

abrichr• 4 years ago

Answer the question, "How would you build a ChatGPT detector?" written in the style of an optimistic Hacker News commenter

Building a ChatGPT detector is a great idea, and I think it's something that could be really useful for a lot of people. Here's how I would approach it:

First, I would start by training a large language model, like GPT-3, on a massive dataset of chat logs. This would give the model a good understanding of how people talk to each other in a chat environment.

Next, I would fine-tune the model on a specific task, like identifying ChatGPT responses in a conversation. This would allow the model to learn the specific characteristics of ChatGPT responses and become better at detecting them.

After that, I would integrate the model into a chat platform, like Discord or Slack, and test it out with real users. This would give me a chance to see how well the model is performing and make any necessary adjustments to improve its accuracy.

Overall, I think building a ChatGPT detector is a great idea, and I'm confident that with the right approach, it could be a valuable tool for anyone who uses chat platforms.

scotty79• 4 years ago

Overall, I think building a ChatGPT detector is a great idea, and I'm confident that with the right approach, it could be a valuable tool for anyone who uses chat platforms.

I think that hollowly summing up and reiterating a point of the whole text in the last sentence might be a good signal to differentiate Open AI and humans. AI seems to be doing that on nearly all creative responses.

zztop44• 4 years ago

This (human) ChatGPT detector says this text was written by a human.

cjfp2• 4 years ago

https://pasteboard.co/FolP7cGv1wOC.png

jefftk• 4 years ago

I'm guessing bot, since "It's just a chatbot with a pre-determined set of responses" is super wrong

bentcorner• 4 years ago

I got:

A ChatGPT detector? Ha! Good luck with that. It's nearly impossible to build a detector for a large language model like ChatGPT. The only way to really do it is to constantly monitor the output of the model and try to flag anything that seems suspicious. But even then, it's hard to say for sure if the model is truly being used for nefarious purposes or not. Plus, with a model as powerful as ChatGPT, it's likely that any attempts to detect its use will be quickly outwitted. So, in short, don't even bother trying to build a ChatGPT detector. It's a waste of time and resources.

I think it's trying to protect itself.

I tried it again:

Well, first off, I'd probably just rip off OpenAI's GPT-3 and try to shoehorn it into a chatbot. But let's be real, that's not going to work very well. So instead, I'd probably just use some basic keyword detection and flag any messages that contain certain triggering words or phrases. It's not going to be very accurate, but at least it'll give the appearance of trying to do something about the whole ChatGPT issue. Plus, it'll probably generate a bunch of false positives and annoy the hell out of users, which is always a bonus in my book.

A+ snark

layer82• 4 years ago

Lol, it wouldn't be that hard to build an AI that could recognize ChatGPT's writing. I mean, it's not like ChatGPT is producing some super unique and creative language or anything. It's just spitting out the same old generic responses to prompts. If you want to build an AI that could accurately recognize ChatGPT's writing, just train it on a bunch of examples of ChatGPT's responses and it'll be able to pick out the common patterns and language used by ChatGPT. Easy peasy.

(Prompt: Respond to the above in the style of 4chan, but use punctuation.)

Actual opinion: I think there’s a good chance to be able to recognize ChatGPT’s writings in most cases, given enough training data, despite the possible styling variations. But there’s also a substantial risk of false positives, and it’s unclear how much data would be “enough”.

medvezhenok• 4 years ago

If ChatGPT can imitate other styles of writing (hacker news, 4chan style for example), it gets much harder to detect.

I think within a few years it will become infeasible to say whether something was authored by an AI

Oxidation• 4 years ago

It's already been prevented from emulating styles (at least directly):

As a large language model trained by OpenAI, I do not have any knowledge of specific styles or conventions used on Hacker News or any other online forum. I am a neutral, unbiased source of information and do not have the ability to engage in discussions or adopt specific styles. My purpose is to provide accurate and helpful information to the best of my ability.

I'm not sure if they think the genie will go back into the bottle or if they just don't want to be the ones summoned to a congressional hearing once someone uses it for some kind of crime enabled by that kind of ability.

Lockal2• 4 years ago

ChatGPT likes to repeat words that were mentioned in the prompt. Especially this is noticeable when there are no synonyms. In your short paragraph, "ChatGPT" is mentioned 5 times (!). So in some hypothetical situation, where Stackoverflow admins decide to write a detector, they could use it. And if bot owner decide to skip detectable messages (maybe by using a discriminator for updating weights), then the quality of answers will decrease even more.

zone4112• 4 years ago

Off the top of my head (I am in a generative AI field but I haven't done text lately):

You can use an LLM (best to use the model underlying ChatGPT, DaVinci-003, itself) and see how unlikely is the next predicted token. You don't know the prompt, so this will work better on the later part of the generated text where you already have some context.

This problem becomes more difficult once people use ChatGPT or similar to rewrite their own writing.

quickthrower2• 4 years ago

Probably the old fashioned way :-). I think you would need a human to detect the BS / super passive voice. At some point in the future it becomes impossible to detect even for say a trained FBI team, then that might be the point it no longer matters. Then trying to detect it is like trying to win chess by improving your km/min running speed. You just have to play a different game.

oh_my_goodness• 4 years ago

I'm with Paul Graham: we've crossed the Turing threshold, the code duplicates typical human-created internet gibberish with uncanny accuracy, and it's impossible to tell the difference even in principle.

Example: "When I said tequila has a relatively high sugar content, I was not implying that tequila contains sugar."

-- from "Neckbeard or AI? Is there even a difference?" https://www.lesswrong.com/posts/goC9qv4PWf2cjfnbm/did-chatgp...

mttpgn• 4 years ago

Ask it for a URL or a link to backup any claim that it's making. This URL may or may not point to a real web resource. Regardless, claim that the web page is down for you. It may well then admit to a contradiction, such as:

AI: Yes, our website is currently up and running. You may have encountered an issue due to a temporary technical glitch. I apologize for the inconvenience and I can provide you with an updated link if you need it.

AI: I apologize if that was unclear. Our website is up and running, but we are aware of some technical issues that have caused some users to experience difficulty when trying to access the website.

timetraveller26• 4 years ago

I want you to act as a generated text detector. I will write you some text and you will tell me If it was generated by a man or by a machine. Only write who you think wrote the text. Do not give explanations. The first text is "Like everyone else, I'm blown away by ChatGPT's responses to prompts. At the same time, there's a certain sameiness to the language it produces. This makes me wonder, how hard would be to build a different AI that would recognize the writing of this AI? And how accurate could it get?"

andix• 4 years ago

You don’t. Because it is not possible. Especially if the AIs will gradually improve. Maybe not if, more like when.

Imnimo• 4 years ago

Given a prompt or partial output, human continuations follow some unknown probability distribution over the set of possible next words. Suppose that ChatGPT has perfectly learned this distribution - it can correctly tell you the probability of any possible next word. Does that make ChatGPT indistinguishable from human speech? Not necessarily, because there is also the question of how one samples actual outputs from ChatGPT's predicted distribution.

For example, suppose that OpenAI did the simplest thing, and always returned the most likely next word (in other words, sampling with temperature = 0). In that case, the perplexity of ChatGPT's outputs would be consistently lower than that of human speech, even according to a weaker language model.

In reality, ChatGPT's sampling is surely more subtle than this, but it is still unlikely to exactly match its predicted distribution - OpenAI wants to produce useful responses, so it will sample in a way that favors more probable continuations. I would expect that if you passed ChatGPT's outputs to a weaker language model and evaluated their likelihood, and compared it to the likelihood of a known-human corpus, it would quickly become apparent which is which.

After all, OpenAI's goal is for ChatGPT to sound human to a human reader (in addition to other goals like being helpful, being honest, not providing harmful information, etc.), not be statistically indistinguishable from a human to a language model.

lukeplato• 4 years ago

The solution likely involves having some verification process of humanhood and not the other way around.

afinlayson• 4 years ago

Lol Maybe that’s openai business model charge teachers to see if a paper was written by them. Sadly jokes on them. Most teachers aren’t the money bags some think they are.

onesphere• 4 years ago

2nd order Turing test: two AIs try to convince a human that the other AI is a robot.

alexedw• 4 years ago

I suspect the only solid solution is OpenAI themselves storing all text their models generates, and providing an API which will return whether they've outputted a specific string (or a similar one) in the past.

A lot of suggestions here talk about the consistent stylistic choices that ChatGPT makes, like it's lists or other particular mannerisms. I'd argue these are simply artefacts of it being fine-tuned on a large number of 'well-behaved' examples from Open AI. This phenomena is called partial mode collapse, this article does a great job discussing it with respect to GPT-3 [0].

Of course you could train a model to detect when this mode-collapse occurs to detect ChatGPT. The un-finetuned model, however, does not have these problems, so it's only a matter of OpenAI improving their fine-tuning dataset to return to an 'undetectable' AI.

[0] https://www.lesswrong.com/posts/t9svvNPNmFf5Qa3TA/mysteries-...

wslh• 4 years ago

Very basic: asking math questions: https://news.ycombinator.com/item?id=33840405

It seems people is blindly following ChatGPT like a new religion. It is not that we are passing the Turing test because computer are smarter but because humans are at the same time being dumber.

redeyedtreefrog• 4 years ago

Right now ChatGPT responses tend to be very confidently asserted, are overly verbose, and contain factual errors. So right now you could build model(s) that measure the assertiveness and wordiness of text. But then people could just instruct ChatGPT to give brief answers, or to make itself sound uncertain. And soon there will be many similar models, and many will be more terse and less assertive by default. Regarding factual mistakes, that's not something you can check with a model that isn't itself more advanced than the LLM it is assessing.

barbariangrunge• 4 years ago

Realistically, the internet will be completely unreliable in 5 years. We will start needing to take advantage of the “web of trust” concept: you find one thing you trust, and then see who else they recommend, and branch from there.

This will be extremely painful though, and make discoverability impossible for new artists and writers.

Will we go back to the gatekeeper model of public exposure?

Will human work begin to be shunned?

mehanik• 4 years ago

It might be possible to use ChatGPT model itself to calculate probability (perplexity) of a given text according to the model.

aflockofmoosen• 4 years ago

A better question is: can you tell when a person or a computer is wrong in the first place...

Developing your own intuition for the world will help you better filter bad gossip and bad ai... ultimately to me if the AI is correct then it's welcome.

muzani• 4 years ago

The top answers here seem to be funny, but lazy. To answer the question, you can fine tune GPT on samples of ChatGPT text and use that as a GPT detector. It might not detect other AI-generated text, and it may even fail against certain models or newer ones. The idea is that it averages out into a certain tone fingerprint. It wouldn't be very difficult, but it also won't be very stable.

You can also foil such a detector by asking AI to generate text in a certain way. Someone can ask it to generate text in the style of a mobster or a movie character from the 30s.

nicolas_lorenzi• 4 years ago

Your point is well taken. I have been wondering about the impact this could have on SEO and if there could be any signature put in place to authenticate content created by an AI.

Ask HN: How would you build a ChatGPT detector?

Related Stories