OpenAI has a 99.9% accurate ChatGPT AI text detector, but won't release it.

Given a sufficient amount of text, the method is said to be 99.9 percent effective.
If that's really the case, they should release some benchmarks. I am skeptical. Promising the world is a key component of their "business model".
- I don't think these grifters know what a benchmark is.
- I think given enough output I could probably detect it that accurately as well. ChatGPT has a particular voice and the longer it goes, the more that voice comes out.
- What is a sufficient amount? Most comments are short af.

"A 99.9% accurate ChatGPT AI text detector? At this time of year! At this time of day! In this part of the country! Localized entirely within your company?!?"
"Yes"
"May I see it?“
"No"

The detection rate is worthless, an algorithm that says anything is Chatgpt would have a detection rate of 100%. What would be more interesting than that is the false positive rate but they never talk about that.
- The detector provides an assessment of how likely it is that all or part of the document was written by ChatGPT. Given a sufficient amount of text, the method is said to be 99.9 percent effective.
  That means given 100 pieces of text and asked if they are made by ChatGPT or not, it gets maybe one of them wrong. Allegedly, that is, and with the caveat of "sufficient amount of text", whatever that means.
  
  It’s actually 1 in 1000, 99.0% would be 1/100.
  
  A false positive is when it incorrectly determines that a human written text is written by AI. While a detection rate of 99.9% sounds impressive, it’s not very reliable if it comes with a false positive rate of 20%.

it's only 99.9% accurate because they haven't released it. As soon as they do, it will quickly fall to 50% as usual. Because this type of thing is exactly what's needed to develop tech to defeat itself.
- What?
  
  Once you have an AI detector, you can use it's results to train your AI to pass the detector.

If they aren't willing to release it, then the situation is no different from them not having one at all. All these claims openai makes about having whatever system but hiding it, is just tobtry and increase hype to grab more investor money.

Total coincidence that this "news" appears about a day after several articles saying the AI bubble is starting to burst.
- It is nut. Who is paying for all these articles and why are they hell bent on convincing everyone that AI is to the left like immigrants are to Republicans
  
  Lots of money in the AI hype game, as tech stocks are massively inflated from just this year alone.
  
  Language models in the end, are just statistics. And to make statistics more accurate, you need more data. Exponentially more data. At the same time, the marginal utility of precision decays exponentially. Exponentially increasing marginal costs are met with exponentially decaying marginal utility.
  
  Why does everything have to be about the USA these days? I'm tired of this joke of a wannabe democracy. Don't want to hear it. Nobody cares. Just stop and leave it to yourself.

I call bullshit.

If you believe this, I have a bridge in Brooklyn to sell you
- A routine that just returns "yes" will also detect all AI. It would just have an abnormally high false positive rate.
  
  My model has 100% recall and 50% precision, not bad eh?
  But - that model would not have 99.9% accuracy.
- Ofc they just look in their database if this is something it has ever said and to who

ALL conversations are logged and can be used however they want.
I'm almost certain this "detector" is a simple lookup in their database.

Probably because it doesn't work. It's not difficult for Open AI to see if any given conversation is one of their conversations. If I were them I would hash the results of each conversation and then store that hash in a database for quick searching.
That's useless for actual AI detection

She goes to another school
\ (for intelligent ificial art)

The detector is most likely a machine learning algorithm. That said, releasing that would allow for adversarial training. (An LLM that would not be detected). Therefore they can only offer maybe an api to use it but can not give unlimited access to the model.
- This is the reason. Releasing it would invalidate it.
- If u release an api for it u can still use that to make training data to beat it.
  
  That's what the Chinese tried with chatgpt. Didn't go well.

They’re keeping everything anyway, so what’s preventing them from doing a DB look up to see if it (given a large enough passage of text) exist in their output history?
- I believe the actual detector is similar. They know what sentences are likely generated by chatgpt, since that's literally in their model. They probably also have to some degree reverse engineered typical output from competing models.

Doubt

There is no way it's that accurate, which is why they don't want to release it.

That's a bad article. What are they reluctant about? Releasing that detector, or applying watermarks to the generated texts? Do they do that already or doesn't it apply to text generated until then? And how would that affect anything else?
Whats with the error rate? Shouldn't that be near 100% for watermarks? And 0 false positives? What's really holding them back? Is pupils not turning in ChatGPT homework anymore cutting into their business model?
I mean all the major AI companies promised to do AI ethically. Now they don't want the one thing that would solve half the issues people are having with that technology. Kind of fits with OpenAI 🤔
- They can't release anything as watermarks can be reverse engineered and people would just wise up and tumble the outputs.
  Weirdly, not releasing this tool publicly might be the smartest bet here as all of these bot farms and idiots just blindly use chatgpt outputs without any tumbling or safety.
  
  The issue with that is: Releasing nothing is even worse than releasing something that could be circumvented. I don't see this as a valid argument.
  I'm not an expert on text watermarking and how that degrades output. But if they want some stealthy solution that isn't known to the public... Maybe they could attach two watermarks. A simple one that is known to everyone, and an additional, secret one only they know about. It'd be similar to what we do with bank notes. There are some characteristics everyone knows and can use to judge if it's fake money. And they have some additional secret markings in banknotes that only the central bank knows about.
  I'm pretty sure a similar thing could be done here. Maybe not for a 280 character tweet. But certainly for other use-cases with longer texts. And in case it has a 0% false positive rate, every match helps someone. Even if it's circumventable. I think even a non-perfect solution that helps several thousands of people is better than helping no-one.

Lots of misinformation in this thread. Yes they have it, it's good but it's probably nowhere close to 99.9% accuracy.
The primary way to detect AI is to inject a fingerprint into AI generation in the first place. This means only the model creators can do that. We don't exactly know how the fingerprint works but it can be as simple as preferring 1 word synonym over the other. For example preferring word synonyms like "illustrate", "peer" etc. quickly ads up to a statistical
These techniques pre-date chatgpt itself and do work! However there are a lot of caveats:
The fingerprint has to be trained for each model meaning each model version performs slightly differently and only owners know the fingerprint.
The fingerprint test can only work on longer bodies of text that are not modified further.
Extending model through more complex instructions (like character, tone) or RAG can significantly decrease the effectiveness.
The industry is understandably very secretive about it but your low effort chatgpt copy/paste can be detected by OpenAI and nobody else.
As for public release of the fingerprint: they can't as it can be reverse engineered so it's only valuable as an internal tool for now. Also if released it would serve no real purpose as detection can be easily defeated by remixing content to dilute the fingerprint.
- Agreed. Frankly, if someone were to say "we can detect with 99% accuracy" I imagine that someone would say "well, clearly your measurements are wrong, find the issue and come back to us when it's fixed".
- but your low effort chatgpt copy/paste can be detected by OpenAI and nobody else
  Low effort copy pastes can absolutely be detected by people who aren't openAI. The consistent "advanced" vocabulary and excessively formal grammar used correctly, but with clear and significant comprehension gaps are pretty damn consistent. You won't get perfect reliability, but you'll catch most of it and you won't have a huge number of false positives.
  Real people don't sound like GPT.
  
  No that's in no way reliable way of catching anyone and I hope people smarten up and avoid this snake oil entirely. I'm borderline jealous how these "ai catchers" are making so much money from straight up snake oil.

You can just ask ChatGPT if a text was written by it.
If it is, it's legally obligated to tell you!
- Don't joke about this, the college professors will hear you.

I wonder if this means they've discovered a serious flaw that they don't know how to fix yet?
- The flaw is in the training to make it corporate friendly. Everything it says eventually sounds like a sexual harassment training video, regardless of subject.
- I think the more like explanation is that being able to filter out AI-generated text gives them an advantage over their competitors at obtaining more training data.

If they have one, and that's IF, then of course they won't release it. They're still trying to find a use case for their stupid toy so that they can charge people for it. Releasing the counter agent would be completely contradictory to their business model. It's like Umbrella Corp. but even dumber.

Did they claim it or prove it? I don't believe anything tech says

This technology will not be published until the GPT-3 code is released.

shhh, my professor may use it
- My unpopular opinion is when they're assigning well beyond 40 hours per week of homework, cheating is no longer unethical. Employers want universities to get students used to working long hours.
  
  I agree, and I teach. A huge part of learning is having the time to experiment and process what you’ve learnt. However, doing that in a way that can be controlled, examined, etc, is very difficult so many institutions opt for tons of homework etc.
- If the assignment is so easy ChatGPT can do it, it’s too easy.

Let me guess: too much processing power?

I trust you bro

„It’s probably broken and I don’t believe you”