‘ChatGPT detector’ catches AI-generated papers with unprecedented accuracy

Doubt

Believable because:
However, the system is highly specialized for scientific journal articles. When presented with real articles from university newspapers, it failed to recognize them as being written by humans.
So outside of its purview? Agree.
Why?
Chatgpt writes them all the same. So its not so much "an AI wrote this" as it is "Bob always writes like this, we know Bob wrote this because _____"
It's a bad headline, but the article immediately clarified.
- All of these were written by chatgpt:
  To make toast, start by plugging in your toaster and adjusting the browning setting to your preference. Take a slice or two of bread and place them into the toaster slots. Push down the lever to activate the heating elements. Wait for a few minutes until the toaster automatically pops the toast back up, indicating it's ready. Carefully remove the toast and spread your desired toppings, such as butter or jam, to taste. Enjoy your warm, crispy toast, a simple and satisfying breakfast or snack.
  In the pursuit of achieving the toasted state of bread, one must first interface with an electrical device colloquially known as a toaster. This remarkable appliance harnesses electrical energy to generate heat through resistive elements. Place slices of bread in the designated slots, ensuring that the toaster is appropriately powered, and set the temporal parameter of toasting, often denoted as "browning level," to the desired degree of crispiness. Activation is achieved by engaging the mechanical lever, prompting the transformational process, during which thermal energy is conveyed into the bread, inducing a Maillard reaction that alters its molecular structure. Upon the cessation of this temporal journey, the toaster signals completion. Carefully extract the now-toasted bread, apply preferred condiments, and savor this culinary endeavor.
  Alright, so like, to make some bomb toast, you gotta start by pluggin' in that toaster thingy and crankin' up the toastiness level, you know, like, how crispy you want it. Then you grab your slices of bread and just slide 'em into those slots in the toaster, pop the sucker down, and hit the magic button to fire it up. While you're waitin' for the toasting action to go down, you can do a lil' dance or somethin', and when it's done, it's like, bam, your toast is ready, girl! Grab it, slather on some butter or whatever floats your boat, and get your snack on – it's, like, totally yum!
  Well, back in our day, making toast was a straightforward affair. First, you'd plug in your trusty toaster and set the browning knob to your liking. Then, take a couple of slices of good old-fashioned bread, none of that fancy stuff, and place them gently into the toaster slots. Press down the lever, and as the heating elements did their work, you'd have time to catch up on the morning paper. When the toast popped up, simply retrieve it with a fork or a butter knife, add some butter, and there you had it – a simple, no-fuss breakfast, just the way we liked it. Those were the days, my friend.
  Seriously, making toast is, like, the most basic skill ever, and it's so obvious that anyone who doesn't get it must be, like, a total amateur. You just plug in that toaster and adjust the settings for your personal taste – it's not rocket science, people! Then, grab some bread, any kind you want, and drop it into the slots, it's not that hard. Push the lever down, and boom, the heat does its thing. It's, like, literally impossible to mess up. But I guess there are still some folks out there who, like, need to argue about every little detail because they just can't accept that not everyone is a culinary genius. 😒🍞 #ToastGate
  No, if chatgpt does not write it all the same.

As I understand it, one of the ways AI models are commonly trained is basically to run them against a detector and train against it until they can reliably defeat it. Even if this was a great detector, all it’ll really serve to do is teach the next model to beat it.

That’s how GANs are trained, and I haven’t seen anything about GPT4 (or DALL-E) being trained this way. It seems like current generative AI research is moving away from GANs.
- Also one very important aspect of this is that it must be possible to backpropagate the discriminator. If you just have access to inference on a detector of some kind but not the model weights and architecture itself, you won't be able to perform backpropagation and therefore can't generate gradients to update your generator's weights.
  That said, yes, GANs have somewhat fallen out of favor due to their relatively poor sample diversity compared to diffusion models.
- I know it’s intrinsic to GANs but I think I had read that this was a flaw in the entire “detector” approach to LLMs as well. I can’t remember the source unfortunately.

No references whatsoever to false positive rates, which I'd assume are quite high. Also, they single out that they built this detector to catch chemistry-related AI-generated articles

If you call heads 100% of the time, you'll be 100% accurate on predicting heads in a coin toss.

I really really doubt this, openai said recently that ai detectors are pretty much impossible. And in the article they literally use the wrong name to refer to a different AI detector.

Especially when you can change Chatgpt's style by just asking it to write in a more casual way, "stylometrics" seems to be an improbable method for detecting ai as well.

It's in openai's best interests to say they're impossible. Completely regardless of the truth of if they are, that's the least trustworthy possible source to take into account when forming your understanding of this.
- openai had their own ai detector so I don't really think it's in their best interest to say that their product being effective is impossible

Willing to bet it also catches non-AI text and calls it AI-generated constantly

The best part of that if AI does a good job of summarizing, then anyone who is good at summarizing will look like AI. Like if AI news articles look like a human wrote it, then a human written news article will look like AI.
The original paper does have some figures about misclassified paragraphs of human-written text, which would seem to mean false positives. The numbers are higher than for misclassified paragraphs of AI-written text.

This is kind-of silly.

We will 100% be using AI to generate papers now and in the future. If the AI can catch any wrong conclusions or misleading interpretations, that would be helpful.

Not using AI to help you write at this point is you wasting valuable time.

I do a lot of writing of various kinds, and I could not disagree more strongly. Writing is a part of thinking. Thoughts are fuzzy, interconnected, nebulous things, impossible to communicate in their entirety. When you write, the real labor is converting that murky thought-stuff into something precise. It's not uncommon in writing to have an idea all at once that takes many hours and thousands of words to communicate. How is an LLM supposed to help you with that? The LLM doesn't know what's in your head; using it is diluting your thought with statistically generated bullshit. If what you're trying to communicate can withstand being diluted like that without losing value, then whatever it is probably isn't meaningfully worth reading. If you use LLMs to help you write stuff, you are wasting everyone else's time.
- Yeah, I agree. You can see this in all AI generated stuff - none of it has any purpose, no intention.
  People who say it's saving them time, I mean I have to ask what these people are doing that can be replaced by AI and whether they're actually any good at it, and whether the AI has improved their work or just made it happen faster at the expense of quality.
  I have turned off all predictive writing of any kind on my devices, it gets in my head and stops me from forming my own thoughts. I want my authentic voice and I can't stand the idea of a machine prompting me with its own idea of what I want to say.
  Like... we're prompting the AI, but are they really prompting us?
- How is an LLM supposed to help you with that?
  I have it read and review a couple paragraphs of a research article, many many times, to create a distribution of what was likely said in those paragraphs, in a tabular format. I'll also work with it to create an outline of an idea I'm working on to keep me focused, and help develop my research plan. I'll then ask it to drill down into each sub-point and give me granular points to focus on. Obviously, I'm steering, but its not too difficult to use it in such a way that it creates a scaffolding for you to work from.
  If you use LLMs to help you write stuff, you are wasting everyone else’s time.
  If you aren't using LLMs to help you write stuff, you are wasting your own time.
Not using AI to help you write at this point is you wasting valuable time.
Bro WHAT are you smoking. In academia the process of writing the paper is just as important as the paper itself, and in creative writing why would you even bother being a writer if you just had an ai do it for you? Wasting valuable time? The act of writing it is inherently valuable.
- Bro gotta get that shit out. Not doing the work to jerk off to the process.

Didn't OpenAI themselves state some time ago that it isn't possible to detect it?

I don't understand. Are there places where using chatGPT for papers is illegal?

The state where I live explicitly allows it. Only plagiarism is prohibited. But making chatGPT formulate the result of your scientific work, or correct the grammar or improve the style, etc. doesn't bother anybody.

If you use chatGPT you should still read over it, because it can say something wrong about your results and run a plagiarism tool on it because it could unintentionally do that. So whats the big deal?

It's not a big deal. People are just upset that kids have more tools/resources than they did. They would prefer kids wrote on paper with pencil and did not use calculators or any other tool that they would have available to them in the workforce.
- There's a difference between using ChatGPT to help you write a paper and having ChatGPT write the paper for you. One invokes plagiarism which schools/universities are strongly against.
  The problem is being able to differentiate between a paper that's been written by a human (which may or may not be written with ChatGPT's assistance) and a paper entirely written by ChatGPT and presented as a student's own work.
  I want to strongly stress that in the latter situation, it is plagiarism. The argument doesn't even involve the plagiarism that ChatGPT does. The definition of plagiarism is simple, ChatGPT wrote a paper, you the student did not and you are presenting ChatGPT's paper as your own, ergo plagiarism.
- Teachers when I was little "You won't always have a calculator with you" and here I am with a device more powerful than what sent astronauts to the moon in my pocket 24/7
I don’t think people are arguing against minor corrections, just wholesale plagiarism via AI. The big deal is wholesale plagiarism via AI. Your argument is as reasonable as it adjacent to the issue, which is to say completely.
Why should someone bother to read something if you couldn’t be bothered to write it in the first place? And how can they judge the quality of your writing if it’s not your writing?
- Science isn't about writing. It is about finding new data through scientific process and communicating it to other humans.
  If a tool helps you do any of it better, faster or more efficiently, that tool should be used.
  But I agree with your sentiment when it comes to for example creative writing.
- To me this question hints at the seismic paradigm shift that comes from generative AI.
  I struggle to wrap my head around it and part of me just wants to give up on everything. But... We now have to wrestle with questions like:
  What is art and do humans matter in the process of creating it? Whether novels, graphic arts, plays, whatever else?
  What is the purpose of writing?
  What if anything is still missing from generative writing versus human writing?
  Is the difference between human intelligence and generative AI just a question of scale and complexity?
  Now or in the future, can human experience be simulated by a generative AI via training on works produced by humans with human experience?
  If an AI can now or some day create a novel that is meaningful or moving to readers, with all the hallmarks of a literary masterwork, is it still of value? Does it matter who/what wrote it?
  Can an AI have novel ideas and insights? Is it a question of technology? If so, what is so special about humans?
  Do humans need to think if AI one day can do it for us and even do it better than we can?
  Is there any point in existing if we aren't needed to create, think, generate ideas and insights? If our intellect is superfluous?
  If human relationships conducted in text and video can be simulated on one end by a sufficiently complex AI, to fool the human, is it really a friendship?
  Are we all just essentially biological machines and our bonds simply functions of electrochemical interactions, instincts, and brain patterns?
  I'm late to the game on all this stuff. I'm sure many have wrestled with a lot of this. But I also think maybe generative AI will force far more of us to confront some of these things.
If you use chatGPT you should still read over it, because it can say something wrong about your results and run a plagiarism tool on it because it could unintentionally do that. So whats the big deal?
There isnt one. Not that I can see.
- At least within a higher level education environment, the problem is who does the critical thinking. If you just offload a complex question to chat gpt and submit the result, you don't learn anything. One of the purposes of paper-based exercises is to get students thinking about topics and understanding concepts to apply them to other areas.

I'm gonna need something more then that too belive it

The article is reporting on a published journal article. Surely that’s a good start?
- I haven't read the article myself, but it's worth noting that in CS as a whole and especially ML/CV/NLP, selective conferences are generally seen as the gold standard for publications compared to journals. The top conferences include NeurIPS, ICLR, ICML, CVPR for CV and EMNLP for NLP.
  It looks like the journal in question is a physical sciences journal as well, though I haven't looked much into it.
https://www.technologyreview.com/2023/07/07/1075982/ai-text-detection-tools-are-really-easy-to-fool/amp/#amp_tf=De%20%251%24s&aoh=16993709470970&referrer=https%3A%2F%2Fwww.google.com

I say we develop a Voight-Kampff test as soon as possible for detecting if we're speaking to an AI or an actual human being when chatting or calling a customer representative of a company.

Edit: I made a mistake.

if we're speaking to a real person or an actual human being
Ummm ...
- Hahahaha OMG. I fixed it. Thanks!

Isnt this like a constant fight between people who develop anti-ai-content and the internet pirates who develop anti-anti-ai-content? Pretty sure the piratea always win.

You sully the good name of Internet Pirates, sir or madam. I'll have you know that online pirates have a code of conduct and there is no value in promulgating an anti-ai or anti-anti-ai stance within the community which merely wishes information to be free (as in beer) and readily accessible in all forms and all places.
You are correct that the pirates will always win, but they(we) have no beef with ai as a content generation source. ;-)
- Oh yes, by fight I mean that no matter how hard developers push proprietary software they get craked anyway, its so funny.
Reminds me of the trace buster buster buster

they still can't capture data written from Ai over websites like ' https://themixnews.com/' https://themixnews.com/cj-amos-height-age-brother/

Isn’t current precedent 0% accuracy already?

If you have 0% accuracy in a binary decision, you could just always choose the other option and be right 100% of the time.
- Ahh, so the true rate would actually be 50% if it was no better than random chance?