I thought of sharing some issues that preoccupy me here, because of how active piefed development is. Btw thank you all for your hard work!
I have noticed that more and more articles are created with LLMs without disclosing it. So, it seems to me that if someone wants to avoid posting this sort of content, one needs to at least:
- check how many articles the author posts per day in the specific site,
- then if the author really exist and
- finally copy-paste part of the text in a couple of ai-detector sites.
Initially, I thought of making a post for a feature request like the one that detects AI generated images, but for text. But I can’t because if I got this right, the ai-detectors may flag an article as ai generated when the author is not a native english speaker [1].
Of course the feature that allows us to label ai-generated content ourselves before posting it is very important. In a way my questions are related to something else: what to do before that. As briefly as I can, here they are:
- does the process mentioned above seem adequate?
- if yes, what else can someone do to check an article before posting, and are there any alternatives/variations to this process?
- if no, what would you suggest?
I think your right on the idea of the posts per days etc, maybe if you did have a feature something like a ‘seems like AI’ button that gets a high rate of common hits; it could help.
I am noticing a lot of what i can only call ‘noise’ posts where 10 average posts (usually a video or link hit and run) and then a specific post littered through…and that could be a political one, or a ad. I feel like I spend a lot of time inspecting users for authenticity lately…I never used to before. I often question why is this person posting? I get that we need posters and engagement but so much of it seems ‘for the sake’ of it. Maybe i am just too cynical.
I am on Piefed so sometimes I see an AI label which absolutely helps, but I know its not on all platforms. And i guess that label relies on a truthful actor.
- finally copy-paste part of the text in a couple of ai-detector sites.
This won’t be reliable at all, those kind of sites give false positives all the time since you use AI to “detect” AI.
I guess checking for em-dashes and common sentences like “why this works” can sometimes be an indication, but I don’t think there is an easy way to reliably automate a process like that.
why this works —–— people are too lazy to proofread LLM slop
Yes, the em-dashes, are very typical. There’s a relevant wiki page (Signs of AI writing) that has a whole section about the overuse of em dashes.
In relation to common sentences tho, I find it to be quite tricky. Still, that’s a good point too. And the language and grammar section of the same article has some tips that can really help even non-native english speakers, like myself.
Sadly, these AI detectors are notoriously unreliable. To the point they’re completely useless for automatic content filtering in bulk. Also rarely Free and Open-Source. Even the “ai-detector sites” I tried aren’t really all that great. Though I welcome suggestions in case someone knows a good (free) one.
Some AI companies said they’re doing watermarking. And I think that’d solve the issue. But then they probably won’t disclose how the watermarking works, so we’re not going to be able to use that one, either.
I don’t see any substantially better solution, than what Rimu already implemented…
It’d be great if we had that, though. And I guess we’d find some people to implement it if someone comes up with a feasible solution.
(Preferably a library(?) or maybe a scientific paper or an algorithm / NLP / machine learning approach we can implement ourselves. Main point: It has to have a low false positive rate, ideally a good detection rate. Has to run on whatever we can afford to run it on. And if you ask me, we better not use any big tech cloud services, like Microsoft Azure AI detector 😅)
Fun fact: I used Libre Office to write an essay for school. I then saved as docx and uploaded it to my class page, the teacher had formatting weirdness so she asked me to resubmit. So I copy pasted the text into word for web, resubmitted… And promptly got called in for submitting AI. Her rationale? “No edits, no version history…” I had the email proof. What a moron.
Lol. Yeah. All that spying on students isn’t a great solution anyway. I’m glad no one knew how many assignments I typed in on deadline day at 1:30am. And in university I tried to typeset the larger homework assignments in LaTeX. With minimal metadata attached to the PDFs.
That’s messed up. I find people staring at my version history a major blocker from writing anything. Unless it’s the git history that I explicitly publish, or the classes that I teach where every step is designed to be followed through, my writing or coding process has always been a highly private, personal thing.
The only AI detection for text that supposedly works, and isn’t public yet, is Anthropic’s watermarking.
And even once it’s public, it can only reliably say that a text has been created or heavily edited by Claude specifically. It can never rule out that it’s been AI-generated in general.
Other AI companies might (be forced to) follow suit in the future; if so you could check them off one by one, but there will always be AI models that won’t have watermarks. So even then only the most lazy slop-posters will be caught (granted, I still think that will help).
I’m still not sure if Anthropic or OpenAI are ever going to publish the watermarking. I’m not a lawyer and the legalese is complicated. But as far as I know they’re trying to find ways to watermark it but not provide us with any proper way to make use of them.
Personally, I’m not a big fan of all the “circumventable” argument. I mean that’s true, and I tend to read it a lot. But in my opinion even getting rid of the majority of the slop would be a great achievement, not matter if some remaining percentage of people are able to circumvent it.
Personally, I’m not a big fan of all the “circumventable” argument.
I agree, that’s what I meant with my last sentence. Even just making it harder for people to create slop undetected should help. Circumvention still requires effort, and effort is the opposite of what the slop creators are about. I just meant it’s not going to be a superweapon against it.
Yes. Sadly it’s one of the top arguments made in some online articles. It’s ye good old “perfect is the enemy of good”. But the EU did it nonetheless. Now we need to wait and find how it’s going to turn out. I’d quite like to implement a fairly accurate AI filter here in the Fediverse. I think that’d alleviate some of the pain we got with “AI agents” flooding all the online spaces.
if you can test something you can (potentially) modify the input until it gives the desired output
AFAIK, Gemini also uses text watermarking but I don’t think they have any open tool for identifying those watermarks
If your instance has rules against LLM content just report the domains to your admin and they can blacklist them (alternatively, if the content is coming from a federated instance they can blacklist instances that allow slop).
You’re talking about social rules to restrict people (and bots) from what they can post. That can be an idea to pitch to mods of lemmy and piefed communities, but isn’t something to do with piefed as a software.
Actually, I didn’t have in mind restrictions on people.
At first, I wondered if a feature like the ai-image thing (link in the post) could be implemented. Since this doesn’t seem to be doable, currently I am just trying to find ways to be able to detect ai-generated articles, as accurately as possible.
So, it seems to me that if someone wants to avoid posting this sort of content, one needs to at least:
- check how many articles the author posts per day in the specific site,
- then if the author really exist and
- finally copy-paste part of the text in a couple of ai-detector sites.
The “one” who needs to do this is the user, no? Which boils down to a “no AI articles allowed” rule, with the question being how it can be enforced.
Perfect example! Even if there is a “no AI articles allowed” rule how are the mods supposed to know that it comes from an llm? Not only that, how can I check an article before posting it and with what degree of certainty?
I get pretty reliable results by pasting text into https://gptzero.me/. I’ll add a link to that onto the Create Post form.
People always say these detectors are unreliable but I think a lot of that is based on how they were years ago, back in the early days.
I really don’t know anymore. For example, today I found this article and the way it was formatted, it seemed to me like it was ai. The following 4 ai-detectors gptzero, sapling, scribbr, quillbot said 100% human. Then I notice that there is no author, and then at the bottom of the page I found their disclosure.
Welcome to London Daily, an innovative and unique news experiment driven by advanced artificial intelligence (AI) technology. […]
Wow that’s an especially tricky one! Normally llm text is much much more obvious.
Read more, especially offline and stuff that was written prior to 2020. You’ll start to notice that the tropes of LLM-generated text stick out like a sore thumb.
