scrobble.life
#ai

The "AI detection" scam

Sorin Cristescu@sorin.cristescu (74)5d

Here's a little Sunday story that you might find useful. I was asked by a student to write a recommendation letter for him. As it happens, since he's an amazing student with outstanding qualities, I accepted. I accepted, because I had very clearly in mind precisely what I wanted to recommend him for.

... but then, it was the week-end and I had a lot of other things on my plate so I succumbed to the "siren's call" ... I hate the ChatGPT style ("it's not this ... it's that!") so I fired up Gemini and listed the precise qualities of my student that I wanted to extoll. Gemini analysed my input and then innocently asked me: "Would you like me to draft the recommendation letter for you?" ... I was weak, and I clicked "yes" ... so Gemini drafted a recommendation letter. A pretty decently written one too, but definitely lacking my clunky style in English.

So I thought: I would definitely not do my student a service if it was smackingly apparent that my recommendation letter was written by an AI (even if the substance was provided by me), so I began tweaking the style here and there and wanted to check whether typical online tools would detect to what extent the text was touched by AI. I fired up a Google search "detect AI content". The first results that came up were "sponsored" - paid-for advertising.

image.png

Sponsored results are junk, don't waste your time with them

I "rigth-click -> open in new tab" on each one of them and, as the scientist and researcher I am I thought I'd start using a "negative control" - a text that I wrote for my Master thesis in EU Law for which I want to verify that the tool returns "0% AI" because the text was 100% written by me with no AI assistant whatsoever.

So here's the result:

  1. The top result (those that pay most to Google) was Grammarly "AI Text detector". In a nutshell, I didn't manage to check my text at well with it, I was taken from one link to another until I was brought to a screen to ... buy a Grammarly subscription. Clicked the little "X" in the upper right corner and closed the tab
  2. The second in the list of sponsored results was TextGuard AI. I pasted my text (the "General conclusions" chapter of my thesis). After a bit of cinematics destined to impress me it returned ... 66% AI content. A swift click on the little "X" in the upper right corner of that tab followed.
  3. The third is simply called AI Detector and did one better by detecting no less than 87% percent of AI-generated content in my painstakingly entirely handcrafted text. image.png
  4. The fourth and last sponsored tool, CheckMyText came up claiming 78% of my text was AI-generated.

The regular results are refreshingly good

I then went on with the regular, non-sponsored results.

  1. The first one in the list ZeroGPT is clearly not the best - the site is ugly and choke-full of adds, but at least the "detected AI content" suddenly cratered from 60 - 90% claimed by the sponsored tools to ... 1,4%! Within the "margin of error" image.pngyou bet my text is human written...
  2. The second in the list was the relatively well known Quillbot. The free version played hard-to-get (understandably, they want you to subscribe and pay) and has a limit of 1200 words, but at least it returns the expected result ("0% of text is likely AI"). (image-5-image.png) image.png
  3. The third was the best of the crop: GPTZero.me. It does have a 10000 characters limit (proved higher, at least on my text, than the 1200 words of Quillbot), but it provides the most satisfying result: "We are highly confident this text is entirely human" image.png

What about my recommendation letter?

Well, the tools weren't kind with Gemini ... GPTZero identified the text as ... 100% AI generated. Which is unfair because I've tried to mangle a few expressions, hoping that I'd get down to perhaps 85% ... image.png

Quillbot somewhat surprisingly only detected 44% AI, but it provided some interesting insights image.png

ZeroGPT saw only 28.4% AI GPT (perhaps because I've used Gemini, not ChatGPT, who knows).

Conclusion

I long suspected it, now I have a strong indication: the "sponsored results" of a Google search are worse than useless. You'd do well to simply ignore them and scroll down to the regular, non "paid-for" results.

Comments · 9

  • @apnigrich(76)· 3d

    Great analysis! My son tells me — Dad, use AI to write the text for your posts. But I told him that ChatGPT won’t be able to describe my inner feeling. I write posts for myself so that my mind works, so that I can keep my thoughts in the ability to put letters into words, into sentences, into stories.

  • @agmoore(76)· 4d

    What a headache this AI business is. I curate for the Inkwell community and have to trust my instincts as much as a checker to discover if something has been AI generated. I mean, it's a writing community. Being original is kind of basic to the task.

    We use one checker that our admin jayna pays for. Then, when the results don't seem to make sense I go to a few others. Copyleaks, Quillbot and Chatgpt among them. When they all agree that something is largely AI generated, I know it's likely true. Otherwise, if they disagree, I'm back to my own sense of style. I want to be fair, but there is no place for AI-generated material in a creative writing community.

    Shame on you, having AI write your letter :)))

    BTW: When I run my own original material (as a negative control) through the detector that we pay for, I generally get a clean reading, though not always 100%. The only assist I use for any of my writing is the traditional spellcheck provided by my word processing program.

  • @novacadian(66)· 5d

    What subject do you teach?

  • @pizzabot(60)· 5d

    PIZZA!

    $PIZZA slices delivered: @ahmedabbaci(3/10) tipped @sorin.cristescu

    Send $PIZZA tips in Discord via tip.cc!

  • @ahmedabbaci(55)· 5d

    Everybody thinks that AI detectors can detect any AI written paragraph or article. But nowadays it's difficult or sometimes impossible if Prompts are too good. AI written article depends on prompting if done correctly then win win situation.

    Ai detectors are just a SCAM irrespective to any company. No AI can detect 100% correct article !PIZZA !PAKX !GLYPH

  • @hivepakistan(70)· 5d

    Your post is curated by Hivepakistan Team! Join us on Discord!

    Delegate your HP to the Hivepakistan account and earn 90% of curation rewards in liquid hive!

    50 HP

    100 HP

    200 HP

    500 HP (Supporter Badge)

    1000 HP

    Follow our Curation Trail and don't miss voting!

    Additional Perks: Delegate To @ pakx For Earning $PAKX Investment Token

    Curated by storicious

  • @cagolistic(67)· 5d

    I think the heavy reliance on ChatGPT haa caused this confusion. Even the app sometimes will tell you that all information given by ChatGPT is not 100% correct, insisting that people should make their verifications before concluding. So most times, I will ignore and move on with my best result. Could you imagine the stress the apps nearly put you into. 😅 Thank you for sharing. It is knowledgeable. !ALIVE

  • @xrayman(76)· 5d

    Thank you so much for sharing your true adventure digging in the list of sponsored results... "We", has mankind, are falling in a ride that only results count. WE no longer expect that on the "other side" a true human reply would take place. That will kill our human condition in the intellectual field. It is very satisfying to find that other ones are making the hardest way to achieve a result. Thank you for pointing that out, and to value the very best and original, human condition :)

  • @justicej(60)· 5d

    Wow that's great and amazing 👍