« first day (3869 days earlier)      last day (480 days later) » 

1:12 AM
@Natty tp
@Natty tp
@Natty tp
@Natty fp
@Natty tp
 
 
1 hour later…
2:45 AM
@Pilot6 Thanks for the notification! Looks like it has been sorted at SO as well as here. This is the second time a user on Ask Ubuntu has utilised ChatGPT targeting Bountied questions. Definitely not a class strategy :(
There is some publicity in Australian press at the moment that has been mentioning the following AI detection:
Might be worthwhile as at least a backup?
 
 
1 hour later…
4:01 AM
@andrew.46 I tested one of the links to suspected AI-generated answers on SO that Pilot6 provided (all of which have been deleted at SO) in an online AI-generated content detection website and the test results were 95% positive for human written content. Now I will test multiple AI-generated content detection websites to see if I can find a better one.
 
@karel As our skills have developed in dealing with these posts I think that online detection is only one part of the picture. This may be more of an issue as users learn how to mask their posts from such detection :(
 
@andrew.46 The weird thing is that the deleted answers from SO all read like stilted AI-generated answers to me, but the test for AI-generated content website reported my test answer as written by a human anyway.
 
@karel You have tried Glorfindel's script or just the straight: openai-openai-detector.hf.space ? I am told that SO for the most part uses this but they also delete / suspend by the appearance of the answer as well as relevance to the actual question. The AI posts, as you know, often have a very loose focus on the original question...
 
5:13 AM
@NotTheDr01ds I have been using the search syntax you suggested for scanning for new answers, not deleted (and then look for ChatGPT). It picks up 508,000 answers by default. So perhaps this is better, it only looks for answers from December 2022 and ends December 2023. It only picks up 2,994 atm so should be a little faster? askubuntu.com/… December 2022 I chose as the beginning of ChatGPT on AU...
I could not find the syntax for 'use current date as an end point' so I just set it to the end of the year...
 
5:31 AM
@andrew.46 I used openai-openai-detector.hf.space and it gave me a 95% probability of being written by a human, which tells me that I still have to read everything.
 
@karel The whole thing can be a frustrating time soak :(
 
6:00 AM
[ SmokeDetector | MS ] Potentially bad keyword in answer (1): Anbox is not starting‭ by a deleted user on askubuntu.com
 
 
2 hours later…
7:48 AM
Autoflagging disabled until restart
 
 
2 hours later…
9:41 AM
@karel It was "human" before I removed the first and the last sentences.
 
10:00 AM
@Pilot6 That's a smart way of cheating on Stack Overflow, but I removed those two sentences before submitting the answer to the test for AI-generated content website and it returned a 95% probability of being written by a human result anyway. I was surprised at the result because I thought the answer was written by a chat bot myself.
 
 
2 hours later…
12:07 PM
@andrew.46 True, but I'm just visually scanning anyway, and the results are paged, so I only go back one or two days at most to view what I haven't seen already.
 
1:00 PM
@Natty tp
 
1:13 PM
@karel The real giveaway on that Stack Overflow answer that tested "real" was that it mentioned --debug-brk as the solution for a Node question. That option was deprecated in Node around 6 years ago ...
Here's the SO post since you can't see it any longer:
 
There is no config file similar to .gdbinit that can be used with the "node inspect" command. However, you can use the --debug-brk option with the node command to set breakpoints before the application starts. Additionally, you can use the debug command within the Node.js inspector to set breakpoints and perform other debugging tasks, Good luck
 
Comparison of AI Detectors:
* GPT2 detector - base model (the one we've been using for a while on HuggingFace): 2.5% chance of being fake
* GPT2 detector - large model (running locally): 5.88% chance of being fake
* demo.aicheatcheck.com: 24% chance of being fake
For the Stack Overflow answer above:
* GPT2 detector - base model: 10.42% fake
* GPT2 detector - large model: 18.06% fake
* demo.aicheatcheck.com: 10% fake
 
 
3 hours later…
4:58 PM
@Natty tp
 
 
1 hour later…
6:27 PM
We have to users with the same name who posted an hour ago 2 suspicious answers. What do you think?
Names are different, but it looks like one person
 
6:40 PM
@NotTheDr01ds
 
 
1 hour later…
8:34 PM
[ SmokeDetector | MS ] Bad keyword in answer, potentially bad ns for domain in answer (89): Touchscreen Input Doesn't Rotate: Lenovo Yoga 13 / Yoga 2 Pro ✏️‭ by Ch Sami‭ on askubuntu.com
 
8:45 PM
@Natty tp
 
 
1 hour later…
10:41 PM
NAA - more like a comment.. askubuntu.com/a/1452526/1066942
 
10:54 PM
@Pilot6 Hmm - Is there another hidden post from the first one? The only post I'm seeing is this, which looks real, since it has a screenshot.
 

« first day (3869 days earlier)      last day (480 days later) »