AI-generated post discussion thread: how to identify & report

[nutildah]

I did not mention artificial intelligence detectors in my message, I had a specific question about the list of phrases used by ChatGPT, which was published by you. As I reported earlier, your list of phrases prepared by ChatGPT is identical to the one that people use. That is, ChatGPT uses the same phrases that are used by people and for this reason you cannot use it for identification. Only if ChatGPT used phrases that are not used by people, it would be possible to consider such messages as created by artificial intelligence.

Phrases not used by people? Who would they be used by then, martians? They are words and many words are used by a large percentage of the population.

The difference is ChatGPT's use of these phrases is exceedingly common. And like I said the presence of these phrases is just one part of the puzzle. Them on their own isn't enough to determine if a post is AI-generated or not, and nobody ever claimed that it was.

Let's take one at random and look at the 10 most recent forum posts to use it: "It is important to emphasize"

#1-2: in reference to AI discussion
#3-4: non-AI
#5: 71.8% probability for AI
#6: 84.9% probability for AI (post deleted as spam)
#7-8: non-AI
#9: in reference to AI discussion
#10: 99.5% probability for AI

So in the latest 10 instances of someone writing "It is important to emphasize" 3 of the posts are found to have a strong likelihood of being AI generated by a detector.

That makes the phrase exceedingly common in AI posts as compared to "Howdy", "What's up", "The trend is your friend", "Holy cow", etc., or anything not in the list of commonly used AI phrases.

[DdmrDdmr]

This is an interesting twist in terms of crypto related content that ChatGPT Plus can now generate. ChatGPT Plus, the version that works on a dataset that surpasses September 2021, and that has apparently moved on to include web browsing capabilities, can now connect to a plugin created by CoinMarketCap to interact with all their crypto related datasets.

The above, theoretically, gives ChatGPT better capacity to reference near real-time data, which can be used by some to further trick their way around here, citing real and recent data whilst not really having a clue themselves.

These are some of the examples that CoinMarketCap claims the combination can now respond to (although these examples do not strictly relate to close to or real time data):

•   "How correlated are Bitcoin and Ethereum in 2023?"
•   "Does Bitcoin typically perform better on weekdays or weekends?"
•   "What was Bitcoin’s performance in the month leading up to the US elections for the past 3 election cycles?"

See:
https://cointelegraph.com/news/coinmarketcap-launches-chatgpt-plugin
https://chainwire.org/2023/10/11/coinmarketcap-provides-a-personal-crypto-analyst-for-all-with-new-chatgpt-plugin/

[Haroon Akram]

Testing a few of those detectors on a text I just stumbled upon

https://writer.com/ai-content-detector/   - 100% HUMAN-GENERATED CONTENT Fantastic!
https://gptzero.me/                                  Your text is likely to be written entirely by a human
https://contentatscale.ai/ai-content-detector/  96%  Highly likely to be Human!
https://x.writefull.com/gpt-detector  https://x.writefull.com/gpt-detector
https://hivemoderation.com/ai-generated-content-detection The input is: not likely to contain AI Generated Text
https://paraphrasingtool.ai/ai-content-detector/   Text may contain AI-written sentences   
https://detector.dng.ai/                               Based on our Analysis your test has been most likely written by a Human.

ChatGBT:

https://i.ibb.co/sQKry9r/62148225.png

 

See? I interact and exchange ideas based on what is in my mind

This is not interaction you just copy-pasted stuff.

This is helpful and unique content

It's not unique content because anyone trying to do the same would come with the same result.
Unique content is generated by unique persons and ways of thinking, in this case, you're giving a standard response everyone could come up with, just like looking at a Wikipedia article and copy-pasting that paragraph.

Easy AI Checker provides 100% accuracy for checking AI generated content.
https://easyaichecker.com/

[nutildah]

Since June, the AI Spam Report Reference Thread has resulted in 500+ AI-generated (or AI-assisted) posts being deleted as spam. It's now 16 pages long and I'd say its been a massive success. Most of the accounts that continue to post ChatGPT-generated material are newbies who don't know any better and think they can get away with it. Most senior accounts don't bother, or else they've found a way to evade the detectors.

Anyway, I'd like to thank the following members for their consistent contributions to the thread:

@jokers10
@lovesmayfamilis
@light_warrior (RIP)
@Who is John Galt?
@Etranger
@Gladitorcomeback
@Shishir99

You guys have been really helpful & I encourage anyone who comes across AI spam to report such posts in the thread, following the steps outlined in the OP.

To achieve 100% accuracy check the content using Easy AI Cheker https://easyaichecker.com/

Thanks for this, I'll check it out & potentially add it to the list of checkers in our thread. Also, to address your question in another thread (since its related to this one):

I paid $100 using Binance pay to a writer after checking the content it appeared AI generated. Is there a methods to claim a refund?

Sources:
https://pay.binance.com/en
https://easyaichecker.com/

I don't think so. Even if you explicitly told the writer not to use AI to write the text, and they did anyway, I doubt Binance would help you recover the money from them.

[Shishir99]

@Shishir99

You guys have been really helpful & I encourage anyone who comes across AI spam to report such posts in the thread, following the steps outlined in the OP.

Thank you nutildah for opening this thread and helping to filter the forum.^{[Just realized it's the discussion thread.]} I was kind of scared when I started reading this post. I thought you were going to close the thread because there might be something wrong. After finishing the reading, I understood it's just a recap post where you wanted to thank the contributors. I appreciate it anyway.

I have noticed that you have created a thread in the service section for managers. I wish you success and I believe your good work for the forum will help you get more clients. Sorry for my terrible English  

[nutildah]

I just wanted to share some statistics I gathered today after analyzing the entire AI Spam Report Reference Thread. Seems like our methodology has been fairly effective thus far as at least 62% of all accounts determined to have been using ChatGPT or another AI bot are no longer with us.

- 106 reported accounts still active or semi-active
- 11 accounts active with a red trust
- 118 accounts inactive
- 14 accounts inactive with a red trust
- 65 accounts banned

Several of the still-active accounts haven't posted in the last 1-3 months, and they will fall into "inactive" status shortly.

[nullius]

For the war on slop, I recommend reading Wikipedia:LLMTELLS in detail.  It condenses the always-current, ever-evolving experience of an expert editing community far larger than the Bitcoin Forum’s staff and volunteer helpers—and they work under the auspices of a nonprofit org with a huge budget.  Wikipedia strictly prohibits the use of LLMs to generate or rewrite article content, with a few narrow exceptions.

Under LLMTELLS, please pay special attention to the parts covering ineffective “tells” which can produce false positives.  Moreover, beware that as LLMs change, some guidelines are deprecated and moved to the “Historical indicators” section.

See also generally WikiProject AI Cleanup and the Guidelines and Resources thereof.

N.b., Wikipedia says that LLM detectors “often suffer from false positives, particularly free ones, and should never be used as the sole evidence when accusing someone of AI usage” (bold in the original, footnote omitted).

(I haven’t read through every post in this thread, but I Ctrl-Fed for “Wikipedia” on all eight pages.  I’m surprised this hasn’t come up.)

It should be treated the same as plagiarism. Asking an AI something then copy and pasting the response is just as bad if not worse than plagiarism because [...]

In my opinion, it’s indeed even worse than the old-fashioned type of plagiarism.  It should be a bannable account nuke offense, if it can be adequately proved—and that’s the trick.  To burn someone on a false positive would be horrible.

I’m thinking of it like a date rape accusation:  Very important to prosecute if true, very important not to punish for if false, and oft in practice hard to prove or disprove.

Again, I recommend studying how Wikipedia handles this.

It occurred to me that ChatGPT often uses the same phrases when generating answers, and I was wondering how to ascertain which phrases are verifiably part of ChatGPT's response formula. Then it occurred to me, why not ask ChatGPT itself? So here it is, straight from the horses mouth:

Whoops.  Many of the words and phrases thus listed are frequent in my own post history, going back to December 2017.

Moreover, I’d be curious as to the results, if someone were to run a popular LLM detector on my highest-merit long-form posts from all the time before the invention of LLMs.  I can’t find it now, but I think I’ve seen the author of the Anubis LLM-scraper-blocker say that after years of blogging, their writing trips bot detectors.  Of course, my forum posts are all in the LLMs’ training corpora—against my consent.  Whereas any false positive on my writing would play to human biases; IIRC, I’ve had people joke that I’m an AI—years before language models got large.

Perusing these threads, I honestly doubt I could make a fresh start here as a Newbie today—unless someone tagged me as a nully alt.

Furthermore, this made me laugh sadly (and almost vomit):

8. Moreover
9. Furthermore

Around early 2020 IIRC, I had a private conversation with Lauda about people who accused us of being each others’ alts.  Lauda pointed out that I frequently started sentences with “Moreover,” she habitually used “Furthermore,” and we could test the alt-paranoiacs’ (my term) acuity by intentionally switching these characteristic words between us.  Since then, I have consciously made an effort to increase my proportionate usage of “Furthermore”.  (I don’t think anyone noticed.)

[fillippone]

Written with the help of: Google Translate & Grammarly

I am guilty on Grammarly.
I am very bad at typing on my mobile screen, and I post most of my activity from a mobile screen. I use Grammarly to fix typos. I often refuse sentence scrambling or improvement, but I might have done it a few times lightly.
I don’t think it should be frowned upon if fewer typos mean a better reading experience.

[joker_josue]

Written with the help of: Google Translate & Grammarly

I am guilty on Grammarly.
I am very bad at typing on my mobile screen, and I post most of my activity from a mobile screen. I use Grammarly to fix typos. I often refuse sentence scrambling or improvement, but I might have done it a few times lightly.
I don’t think it should be frowned upon if fewer typos mean a better reading experience.

I don't think that's wrong.

For me, it simply doesn't make sense to use direct AI tools, like ChatGPT. Because I think they are just additional resources for a translation process, which will be used here in the forum, which is an informal space.

As long as there is balance, they are useful tools that we can use. The important thing is to make sure you don't lose control of what you really want to say.

[CryptoHeadlineNews]

Around early 2020 IIRC, I had a private conversation with Lauda about people who accused us of being each others’ alts.  Lauda pointed out that I frequently started sentences with “Moreover,” she habitually used “Furthermore,” and we could test the alt-paranoiacs’ (my term) acuity by intentionally switching these characteristic words between us.  Since then, I have consciously made an effort to increase my proportionate usage of “Furthermore”.  (I don’t think anyone noticed.)

Yes, it's true that the constant use of a certain phrase when writing might be a visible parameter by which people can judge the similarities of two or more different accounts on the forum to be a potential alt. But moreover, its still not a valid proof on its own. As there are more other important factors that can be used to check the likelihood possibility of alt accounts, such as posting style, posting time, favorite board, and lastly the average number of words/characters those users post on the forum. So for people to have alleged your account to be an alt to Lauda, then its obvious they might have seen a similarity in any of those parameters I shared. But however, it's still wasn't enough proof for such a case, since its still possible that two people can still share these similarities and still not be the same.

[jokers10]

Whoops.  Many of the words and phrases thus listed are frequent in my own post history, going back to December 2017.

AI models use the same words and phrases as human beings, because they were made for making meaningful human-like written text. So no surprise we all use those phrases from time to time. And "—" was used by me even before AI models became popular. So all this is just a number of marks to pay attention to, but not a proof by themselves. This your text has marker phrases and "—", but several detectors I used for it showed total 0 probability of AI usage.