Ask HN: Have you had any of your work falsely detected as being generated by AI?

[Quinzel]

And also, if you have? What did you do? Did you have to prove your innocence? And if you did, how did you do that?

I am particularly interested in this with regard to academic work, but also in general would be interested to hear people’s experiences!

[alexkrkn]

I wrote a blog post and submitted it to HN yesterday. At first the upvotes were coming in quickly and it even got to the front page, but then some comments started saying that it's '100% written by AI' and it ended up being flagged. I wrote the entire thing with no use of AI, and only used chatgpt to proofread it.

I think it's amusing how people are paranoid about wasting time reading something that was generated by AI, and I can relate to it. But it was strange seeing how sure some people were that it was written by AI, when in fact it wasn't.

Here's the post: https://news.ycombinator.com/item?id=37480763

[iamflimflam1 → alexkrkn]

It really does feel written by AI - a telltale phrases that jump out “Let’s analyse how…”

If you’ve used ChatGPT at all you’ll have seen this “Let’s…” text a lot.

[alexkrkn → iamflimflam1]

I think ChatGPT uses it a lot because it was trained on articles that use it, just because it's a common phrase, no?. It's actually something I started saying while recording screencasts, so it just became a phrase I use often.

I think you're right, it does sound "artificial", I should avoid using it.

In the comments of my original post someone also mentioned: "These days, I personally find articles with minor grammatical errors more interesting than AI assisted/enhanced/corrected content, because AI generated content often comes off bland and loses the rawness of author's intentions.".

I agree with your points, it's amusing though that we need to adjust the way we naturally speak and keep some errors, to not appear as AI.

[Kerb_ → alexkrkn]

I've got a reddit history a decade back with "however", "likewise", "it's important to consider/remember". I have no doubt if I re-posted/commented some of my posts, I would be accused of being an AI. I think some of us just write in a way that people will assume is robotic, especially when given a chance to revise our works for efficient communication.

[Quinzel → Kerb_]

I agree with you. I have often made use of quite generic sentence structures and words such as “Therefore”, “However”, “Henceforth” etc… particularly in my academic writing. Additionally, I find that a challenge that I have is that with my studies there is high value placed on conciseness. So unlike informal writing where I can be excessively wordy, and abuse commas, when I write for academic purposes, I generally edit obsessively to get rid of excessive wordiness and grammar issues (eg, instead of saying “in order to” I will write “to”).

Something I also noted since ChatGPT was released, however, was when I first played around with it, I did think it had some interesting ways of stringing some sentences together, and used words I wouldn’t have typically used in my own writing - but I’ve since stolen some of its use of words and copied its pattern of stringing some words together, and written them in my own writing - as if the chatbot impacted on my personal writing style - but I don’t actually use AI to generate ideas, write content or edit my actual writing.

It makes me wonder - will AI also influence human writing patterns/styles and therefore make it more difficult to distinguish between humans and AI generated texts.

It’s a shit time to be studying at uni, I’ve gotta say.

[cableshaft]

Saw this post a few days ago on Reddit about their game getting rejected by Steam (the main platform studios publish their games on for PC) for falsely detecting A.I. assets. A few people in the comments chimed in and said they had similar experiences as well:

"We are a small indie studio publishing our first game on Steam. Today we got hit with the dreaded message "Your app appears to contain art assets generated by artificial intelligence that may be relying on copyrighted material owned by third parties" review from the Steam team - even though we have no AI assets at all and all of our assets were hand drawn/sculpted by our artists."[1]

https://old.reddit.com/r/gamedev/comments/16bcj4a/first_indi...

[Quinzel → cableshaft]

I Seen something similar on here the other day where someone had their game rejected because of AI. I actually naïvely thought that having AI in a game would be advantageous because it could be used to make the game a more personalised experience. I don’t know anything about game development so… so my thoughts are actually irrelevant

[muzani → cableshaft]

Interesting, didn't realize Steam was banning AI art. That sounds like a very difficult policy to enforce too. If it's manually approved, it can be difficult, though there are signs like characters keeping their hands in their pockets. If it's AI checked, that sounds like using thieves to catch the thief.

[cableshaft → muzani]

Almost certain it's the latter. Pretty much all of these large companies start with A.I. detection then manually review when appealed (if they bother to with the appeal at all).

[yieldcrv → cableshaft]

On that note, given that AI assets have no copyright protection, how would one prove they are AI assets if they wanted to take royalty free, or conversely prove its not AI

[JohnFen → yieldcrv]

I think that if you're the one who created the asset, proving that would be relatively easy. The asset didn't come fully formed on first blush, so all you'd need to do is show the incremental work that led to the final asset.

[scarmig → JohnFen]

It's only a matter of time until models can create that trail of incremental work, though.

[JohnFen → scarmig]

In the same sense that people can do it, sure, but it's not all that easy. You have to go back and fake all sorts of timestamps, email exchanges, handwritten notes, etc. AI doesn't make doing that any easier, and it's often not that hard to detect fraudulent records.

The amount of stuff you'd have to fake, both electronic and physical, makes doing so successfully a nontrivial exercise and -- if you're found out -- turns it into a much larger, potentially criminal, case of fraud.

[PeterStuer → JohnFen]

Guilty until proven innocent?

[JohnFen → PeterStuer]

It's not a guilty vs innocent thing. It's a proof of provenance thing. It's not even remotely new. If you're in a patent or copyright dispute, being able to "show your work" has always been the strongest way to demonstrate provenance in court.

[cableshaft → yieldcrv]

Top comment explained how another person appealed it, but I don't think everyone could do this:

"I ran into the same issue as well making a visual novel. All I had to do was show them proof that we made those assets - character sketches, color roughs, unrendered 3D assets and mention the artist names and they approved it within a few days. My advice is make it as easy as possible for them."