How to Detect AI Comments
100 comments
Greetings everyone! I trust you are all doing well. Today I came here to discuss a very important and concerning issue. I could have discussed it in some other community but instead I chose this one because I don't want people flooding my comments section. I just want some serious and reputed people, especially mods and admins, to engage in a healthy discussion here.
For the past 2 weeks, I have been observing that a debate has started regarding the engagement challenge, where some users are raising the issue that they have been treated unfairly – they were not declared winners despite having higher engagement than others, etc.
Along with this, another issue that is ongoing, which I believe is directly linked to the engagement challenge, is AI-generated comments. Although I had raised some concerns about this in the comments section of another post earlier, there was no satisfying response from higher authorities, so I had ended that discussion.
Today, I was having a discussion about AI comments with some of my friends on Discord, and one of them, a Steem Representative, messaged me privately and gave me the green signal to write a post to inform people about this matter so that those who are spamming AI comments can be caught and prevented from this activity.
Initially, I did not want to write about this, but then I thought that since no one else is openly discussing these things, why shouldn't I write a research-based post so that some mods and admins who are unaware of these issues can also get an idea of how serious this matter is.
I think the matter of AI-generated comments became intense when communities decided to select winners based solely on engagement. Even though the rules of engagement challenges state that communities are free to choose their criteria for selecting winners, they just have to mention what their winner selection criteria are – for example, if they have declared a particular user as the winner, they should explain why.
We will leave it to each community to decide how they select their weekly winners, but they should include in their announcement posts a brief summary of their selection process. source
It's also interesting that there are different types of users on Steemit. Some write on their own and have command over whichever language they write in, some write with the help of translators, some take a little help from AI but mostly write on their own, some mix AI content with their own content while writing, and then there are some who write by spinning or paraphrasing AI content to avoid AI detection. There are also some newbies who copy-paste AI content entirely, but the poor souls don't last long because AI detectors catch them if a moderator is properly checking.
And yes, I have recently seen some posts that mostly include rhetorical phrases like it's like this, it's like when, it's like that, etc. I believe these are also products of AI because when you include such phrases, the AI detectors, which are not very accurate to begin with, easily declare your AI content as human-written. Anyway, this discussion can go elsewhere. I want to come back to AI comments now.
The point is, you cannot detect AI comments with detectors, and mods don't have enough spare time to check comments along with posts. The factor of inaccuracy of AI detectors also needs to be kept in mind – this task can only be done by someone who has good knowledge of that language.
Whenever you scan less than 300 words with AI detectors, you will get strange and weird kinds of results. Mostly, it will be declared as AI even if you have written it yourself – there was a discussion about this on one of my posts last year under the name Torch letter or something, I think.
Now the question that arises is how to detect these comments then. The simple answer is that you cannot do it with any software, but you can certainly catch those users who are constantly abusing this platform by using your intelligence and wisdom. This is because AI comments have a specific pattern, which I will discuss below.
However, I should mention a few other important things first. Mostly, newcomers are involved in such activities, so you must definitely check a user's reputation. But this does not mean that higher reputation users are not involved in such acts – they are, and I know some of them, but I do not consider it appropriate to disclose them openly.
Additionally, before making any decision, you will have to analyze at least 50+ comments of that user. If they are AI-written, you will definitely find some pattern. You can also read some of the user's posts to see if their “English” in posts and comments matches or not. Often, users write in very ordinary English in posts, but in comments, they appear as great advisors and philosophers. This is also a way to catch them.
Now, coming to the main point, which is the real purpose of this post. I selected 4 random posts and had ChatGPT write replies to them, and the results I got, I will show you through screenshots. I have highlighted some specific patterns that I noticed in red boxes.
But again, I will say that you cannot declare a user's comments as AI-generated based on a single comment where this pattern appears, unless you thoroughly analyze their 50+ comments.
I also checked by modifying the reply prompt, because I suspected some users were comment spamming by sharing fake life experiences through AI in almost their every comment, so you can see that too.
Example 1
 | LINK TO THE POST |
|---|
Example 2
 | LINK TO THE POST |
|---|
Example 3
 | LINK TO THE POST |
|---|
Example 4
 | LINK TO THE POST |
|---|
To my surprise, there was a comment below this very post that closely matched this AI comment, and when I analyzed some other comments of this user, I noticed some AI-like patterns in them as well. I upvoted this comment with a small percentage so that she won't be able to delete it after seeing this post.
Now, I will also post a full screenshot here so that this user cannot later come and say that this is not valid proof.
Final Thoughts
Detecting AI-generated comments is tricky and not everyone's cup of tea, but as everyone strives to top the engagement rankings by writing comments, some will undoubtedly try to abuse the system with AI.
If you ask me, the criteria for selecting winners should be based not only on engagement but primarily on the quality of the posts. This would give those caught in the comment wars a breather and encourage them to focus on improving the quality of their posts instead of just increasing their comment count.
Software can tell you how many comments a user has made, but it can't tell you if those comments are eligible. Considering the AI factor, this can only be managed by manually analyzing the comments, and I doubt anyone is willing to take on that headache.
However, you could set a minimum requirement that for someone to qualify as a winner, they need at least 5 valid comments made and received. I think this approach could help mitigate the issues created by AI-generated comments.
Comments