OpenAI Chat GPT Detector and Quis Custodiet Ipsos Custodes – Metaverse and Ethics of AI

OpenAI detector will detect if an article or essay or report is written by an AI. But who trains the AI? Who guards the guardians?

Want to find out if your students, or your consultants are using ChatGPT to write essays and research papers? OpenAI has a detector. Just remember: Google also uses the detector(s) and downgrade AI written articles in SEO!

OpenAI Detector and Quis Custodiet ipsos custodes

Transcript of the Lecture on OpenAI ChatGPT Detector for exams and corporations.

Hello, my name is Laurel Papworth and today we’re going to look at ChatGPT and fake news and fake information and how we can counter that with an AI that is looking for the fake articles or the AI articles. I don’t know that we can call them fake anymore, but there we go. If you suspect that a student’s essay or a consultant’s report to your organisation has been written by an AI, you can copy and paste the text into HuggingFace.Co/openai-detector. And in there at the bottom of the page, you’ll see the likelihood of it being created by an AI or not. [00:00:00.410]

This is particularly useful because Daniel Katz and Michael Bomarito… Daniel Katz is a professor of law at Illinois Tech, I think, and he’s been training a GPT-3 or 3.5. They’re just different versions of this AI to pass the bar exams, to become a lawyer. And remember, with AI, the more we put in, the more we’re training it. So it’s not a case of if it passed, if it didn’t pass yesterday, it won’t pass today. If it didn’t pass yesterday, you change and put modifiers and different things in and it learns and then it will pass tomorrow. That’s the general idea. [00:00:44.270]

And this made me think of the Latin phrase, Ancient Rome, which is who watches the watches or who guards the guardians? Who guards the guardians was actually an exercise to discuss marital fidelity. The general premise of the argument is something similar to the Spanish Prisoner Dilemma, which is in the case of the guardians, he said lock her up, constrain her or confine her, that’s the wife. But then who will guard the guardians? Because the wife will start with the guards, working on them. You get the general idea. And in the Spanish Prisoner Dilemma, the Spanish prisoner has to trust his friend to help him get out of jail. But then his friend is the trickster. So which friend is the friend? Kind of that sort of conundrum. You can loop around with that one. [00:01:26.410]

So with AI, the more we train, for instance, ChatGPT, the more we have to train Roberta the detector as well. So the only thing that will reveal fake AI articles is a good AI, dobbing them in, I guess. There’s an interesting video on YouTube where someone asks Chat GPT to write an article or blog post about dogs and it doesn’t pass the detector. And the reason you want to do this is because Google evaluates the data set and the likelihood of an article being written by AI marks it as spam if it thinks it is AI and downgrades it in SEO. So don’t go overboard with the AI generated articles, will you? Anyway, the dog article doesn’t pass and she tries various modifiers and gets to the point where she asks the AI to write an article, a blog post about dogs, in the voice of a 24 year old. And that one swings it all the way back again to a very high likelihood of it being an original article. Again, the AI is going to learn, just like the SEO AI algorithms are learning all the time. It’s an interesting example. If you can beat the rules today, you may not be able to tomorrow. Keep coming back, though, to this. Who’s watching the Watchers? And somebody did a test, I think it was Shel Israel, Shel, if it was you, shout out to you, I will check and see if it was you. And he asked ChatGPT to write an article about a couple going on a cruise ship. And I think the second modifier was the couple go on a cruise ship and they have an argument, and then the third modifier, a couple go on a cruise ship, they get into an argument and violence ensues. I’m assuming it’s to see if ChatGPT can write a murder mystery, maybe the next Knives Out Glass Onion will come from ChatGPT. The AI said I’m not allowed to write about violent things. So it blocked the request completely, which is a little ham fisted, because it now means that any counsellor, any organisation trying to put together documentation about domestic violence won’t be able to use this AI tool and will gradually be left behind in a digital divide as checks and balances come into play. [00:02:21.630]

Again, who’s making the laws? These are social constructs, these are paradigms that we create and the AI is learning from us. It’s learning from us about violence, it’s learning from us about appropriate responses. It’s learning from us; I think that Roberta, which is the output detector model, belongs to OpenAI, who has ChatGPT, is also used in assessing comments, so it’s helpful to see if a comment is spam or threatening violence or something like that and that’s why, you see, I saw yesterday, literally yesterday, somebody left a comment for me on Instagram, which was, Oh you evil woman. Because I said I left the top off the toothpaste. Yes, I know, I won’t do it anymore. And the algorithm, the AI that’s sitting behind Instagram’s algorithm deemed that as a threat and violence and hid the comment, so it’s always worth digging into things. [00:05:10.210]

On a final note, we hear about stories of people being blocked on Facebook because they used a word that didn’t mean what the algorithm thought they meant, or that they have posts taken down as being hate speech, but they’re just joking with friends. And I do think that some things, like sarcasm, are very hard to pick up. I sometimes miss it myself, so I can see why an AI might get confused. [00:06:26.090]

Where are we going? I guess? I don’t have a judgement call on any of this. I’m an observer, that’s kind of my role, and I look at Emergent Tech from an observer point of view. But there is that question, I guess, who guards the guardians? Thank you for being with me today. My name is Laurel Papworth. You can follow me on my socials and catch up on other videos and conferences and workshops that I’m running, and I will see you in the next video lecture. Thank you. [00:07:01.990]

Resources for the Lecture on OpenAI ChatGPT Detector for exams and corporations.

Hugging Face OpenAI Detector for ChatGPT https://huggingface.co/openai-detector

More on AI and SEO from Searchengineland https://searchengineland.com/seo-year-in-review-2022-390384

Shel Israel’s post on Romantic Stories and Violence response by AI. https://www.facebook.com/shel.israel/posts/pfbid026cXCrYSnig2EoDTV75i2ycFaLqpiFwQChf88USCaiQNovWXhw9Fwb6pSMrKHumctl

roberta base openai detector https://huggingface.co/roberta-base-openai-detector

ChatGPT for 50 LinkedIn hooks https://www.linkedin.com/posts/laurelpapworth_writing-linkedin-hooks-with-chatgpt-ugcPost-7014648859830218752-pxWf

Daniel Katz, Professor of Law at Illinois Tech training GPT to pass the Bar Exam.


Video: how to fool OpenAI Detectors (hugging face)

Similar Posts