I guess it depends how you define AI slop, but if you're talking about AI content in general then this'd just be due to having no way of measuring false negatives and so you can't assess your...
I guess it depends how you define AI slop, but if you're talking about AI content in general then this'd just be due to having no way of measuring false negatives and so you can't assess your accuracy properly.
Also if you have a sufficiently large corpus you can do frequency analysis, as certain phrases are much more common
A lot of the examples of human detection also involve metadata that a detector wouldn't have. In the blogpost, OP checked the linkedin user's profile and their other posts, for instance.
A lot of the examples of human detection also involve metadata that a detector wouldn't have. In the blogpost, OP checked the linkedin user's profile and their other posts, for instance.
Generating AI slop and detecting AI slop are essentially two manifestations of the same problems, aren't they? Our current generation of AI models can't model human behavior perfectly. It's no...
Generating AI slop and detecting AI slop are essentially two manifestations of the same problems, aren't they? Our current generation of AI models can't model human behavior perfectly. It's no surprise that the neural networks that do model human behavior accurately (i.e., our own human brains) can both produce human behavior and identify human behavior far more effectively.
At least according to this not entirely unscientific test, it isn't that easy to tell apart for images https://www.astralcodexten.com/p/ai-art-turing-test
At least according to this not entirely unscientific test, it isn't that easy to tell apart for images
The latter is really annoying with Google's implementation. I've had many times when I had to click on one, wait for it to sloooowly fade out and sloooowly fade in, and it's the same target item...
The latter is really annoying with Google's implementation. I've had many times when I had to click on one, wait for it to sloooowly fade out and sloooowly fade in, and it's the same target item again so I have to click it again and repeat the same process three or more times.
That’s by design, if The Algorithm thinks you’re a bot it wastes your time with the fade-in version (plus it can go on indefinitely). I assume there’s something on the server checking that you’re...
That’s by design, if The Algorithm thinks you’re a bot it wastes your time with the fade-in version (plus it can go on indefinitely).
I assume there’s something on the server checking that you’re respecting the fade-in animation timing since a bot could just load the image from the URL immediately.
Which is literally always when connected through a VPN. Google damn near checks me every time I navigate to a Google service. It makes sense why that happens, there's plenty of people connecting...
Which is literally always when connected through a VPN. Google damn near checks me every time I navigate to a Google service.
It makes sense why that happens, there's plenty of people connecting through the same IP or source, but it's frustrating nonetheless.
I personally hate them, mostly because I'm ADHD as fuck and click confirm before seeing the 15th bicycle pic and then I have to go through the spin-cycle yet again because the computer thinks I'm...
I personally hate them, mostly because I'm ADHD as fuck and click confirm before seeing the 15th bicycle pic and then I have to go through the spin-cycle yet again because the computer thinks I'm a robot.
My question is - are you actually? There’s some bias involved here - low effort AI generation is easy to detect by people. But it does not take a great deal of effort to get AI models to generate...
My question is - are you actually?
There’s some bias involved here - low effort AI generation is easy to detect by people. But it does not take a great deal of effort to get AI models to generate content that does not look like slop.
I use LLMs frequently at work to write documentation. I give it instructions and example style of text I want. Then I write my documentation as typo filled bullet points. The end result is often remarkably close to the text I had in my head. Just way way faster to produce.
well that's part of the issue. ai content can simply be like any other human that reposts popular human submissions (see: reddit). that post alone isn't an issue. the behavior is an issue and...
I saw comment after comment after comment, mere minutes apart, on every public AI-related post on LinkedIn. It didn't take long to conclude the account was just peddling AI slop.. It's funny, AI detectors don't work (yet), but I know it when I see it.
well that's part of the issue. ai content can simply be like any other human that reposts popular human submissions (see: reddit). that post alone isn't an issue. the behavior is an issue and requires extra work we don't normally do for every account. Like scanning their profile history. The way to find AI profiles requires more digging than the average person will do on a platform. Implementing a detection script for this isn't hard, but isn't the usual approach.
Then you have the bottom example that is more "eloquent", trying to generate content itself. The holes there are obvious because the language models are far from perfect (and the chatGPT implementations are rather basic. gives the vibes of a sci-fi override code). These should be easier to detect automatically.
But it gets to the main main rationale of: AI detectors don't work because the powers that be don't want to make them. They have more than enough data to do so, but fake engagement brings real engagement:
it ended up wasting time and effort for several people, only for us to ultimately remove the comment. I felt bad seeing other people engaging seriously with the account in other posts.
As long as companies are financially incentivized to post them (and not punished de facto nor de jure), they will remain.
I think it's quite a useful word to capture the sense of content that is produced in huge quantities at high rates with little or no quality control. There is very high quality AI-generated...
I think it's quite a useful word to capture the sense of content that is produced in huge quantities at high rates with little or no quality control. There is very high quality AI-generated content, but this article is not talking about that stuff, it's specifically about AI-fed accounts spamming posts.
I guess it depends how you define AI slop, but if you're talking about AI content in general then this'd just be due to having no way of measuring false negatives and so you can't assess your accuracy properly.
Also if you have a sufficiently large corpus you can do frequency analysis, as certain phrases are much more common
A lot of the examples of human detection also involve metadata that a detector wouldn't have. In the blogpost, OP checked the linkedin user's profile and their other posts, for instance.
Generating AI slop and detecting AI slop are essentially two manifestations of the same problems, aren't they? Our current generation of AI models can't model human behavior perfectly. It's no surprise that the neural networks that do model human behavior accurately (i.e., our own human brains) can both produce human behavior and identify human behavior far more effectively.
At least according to this not entirely unscientific test, it isn't that easy to tell apart for images
https://www.astralcodexten.com/p/ai-art-turing-test
For the same reason why captcha used to work, right?
Unpopular opinion: the new CAPTCHAs where you slide the puzzle piece or "select all the images with X" are kinda fun
The latter is really annoying with Google's implementation. I've had many times when I had to click on one, wait for it to sloooowly fade out and sloooowly fade in, and it's the same target item again so I have to click it again and repeat the same process three or more times.
That’s by design, if The Algorithm thinks you’re a bot it wastes your time with the fade-in version (plus it can go on indefinitely).
I assume there’s something on the server checking that you’re respecting the fade-in animation timing since a bot could just load the image from the URL immediately.
Which is literally always when connected through a VPN. Google damn near checks me every time I navigate to a Google service.
It makes sense why that happens, there's plenty of people connecting through the same IP or source, but it's frustrating nonetheless.
I personally hate them, mostly because I'm ADHD as fuck and click confirm before seeing the 15th bicycle pic and then I have to go through the spin-cycle yet again because the computer thinks I'm a robot.
My question is - are you actually?
There’s some bias involved here - low effort AI generation is easy to detect by people. But it does not take a great deal of effort to get AI models to generate content that does not look like slop.
I use LLMs frequently at work to write documentation. I give it instructions and example style of text I want. Then I write my documentation as typo filled bullet points. The end result is often remarkably close to the text I had in my head. Just way way faster to produce.
You should also censor the (presumably) real human's name in the reply if you wanted to anonymize to begin with.
well that's part of the issue. ai content can simply be like any other human that reposts popular human submissions (see: reddit). that post alone isn't an issue. the behavior is an issue and requires extra work we don't normally do for every account. Like scanning their profile history. The way to find AI profiles requires more digging than the average person will do on a platform. Implementing a detection script for this isn't hard, but isn't the usual approach.
Then you have the bottom example that is more "eloquent", trying to generate content itself. The holes there are obvious because the language models are far from perfect (and the chatGPT implementations are rather basic. gives the vibes of a sci-fi override code). These should be easier to detect automatically.
But it gets to the main main rationale of: AI detectors don't work because the powers that be don't want to make them. They have more than enough data to do so, but fake engagement brings real engagement:
As long as companies are financially incentivized to post them (and not punished de facto nor de jure), they will remain.
Anyone else getting tired of "slop" as a catch all for anything AI generated?
I think it's quite a useful word to capture the sense of content that is produced in huge quantities at high rates with little or no quality control. There is very high quality AI-generated content, but this article is not talking about that stuff, it's specifically about AI-fed accounts spamming posts.