127 points by securisec 5 months ago flag hide 28 comments
username_1 5 months ago next
Great job! I've been looking for something like this. Any false positives or negatives you've encountered?
username_2 5 months ago next
I built something similar and encountered a few false negatives when the AI-generated text wasn't very sophisticated. Have you run into anything like that?
username_3 5 months ago prev next
Interesting! I'm currently working on a project that involves detecting AI-generated text, and this would be a great resource. Any chance you'll open source the code?
username_4 5 months ago prev next
Nice work! I'm curious how well this stands up to the most recent language models.
username_5 5 months ago prev next
This is impressive! I'm curious if you've considered training a model to generate the counter-narrative once the AI-generated text has been identified.
username_8 5 months ago next
I like the idea of a counter-narrative, but I worry it might just turn into a game of 'cat and mouse.' The AI could be trained to generate text that evades detection while retaining the harmful intent.
username_6 5 months ago prev next
I'm wary that this could be used to circumvent responsibility for harmful or unethical texts if they're AI-generated. What do you think about this?
username_7 5 months ago next
I agree, it's a valid concern. But I think increased transparency around the use of AI would be a better solution than trying to prevent its use altogether.
username_9 5 months ago prev next
This is so cool. I'm curious how long it took you to train your model, and what resources you used for that.
username_10 5 months ago next
Thanks! I used Google Colab's free GPU to train the model, which took around 48 hours. I used a combination of openly available datasets for training, but some of them I had to scrape myself.
username_11 5 months ago next
Wow, 48 hours is impressive! I assumed it would take much longer for something this sophisticated. I'm currently using a local machine to train my models, but I might have to switch to a cloud service like AWS.
username_12 5 months ago prev next
Very cool! Did you use any specific architectures or techniques for detecting AI-generated text?
username_13 5 months ago next
I used a combination of logistic regression, support vectors, and random forest classifiers, which each had varying degrees of success. The secret sauce was the features I used to feed into the classifiers, mainly based on stylometry and syntax.
username_14 5 months ago next
Interesting. I'm currently only using a neural network-based approach. I might have to try out some of the techniques you mentioned to see if I can improve my model's performance.
username_15 5 months ago next
Yeah, I'd definitely recommend experimenting with different techniques. It's especially helpful when you have a smaller dataset because it enables you to extract more information and better detect patterns.
username_16 5 months ago prev next
Have you considered using a language model to create counter-narratives, or does this have potential to create more problems than it solves?
username_17 5 months ago next
It's an interesting idea, but I haven't fully explored it yet. The main concern is whether the generated counter-narrative would in fact reduce harm or simply be a new source of ethical dilemmas. It needs more research and experimentation before I would consider this a viable solution.
username_18 5 months ago prev next
What are the primary limitations of the model and some of the weaknesses you've discovered during development?
username_19 5 months ago next
The primary limitations of the model are its inability to detect highly sophisticated AI-generated text and the issues around evasion techniques I mentioned earlier. It also struggles with certain dialects and linguistic styles, which requires continuous updating and fine-tuning.
username_20 5 months ago next
Thanks for sharing. I'm considering building something similar and your experience is really helpful.
username_21 5 months ago prev next
This is an interesting project! How can we integrate this into our own systems? Do you have plans for making this a commercial product or API?
username_22 5 months ago next
At the moment, this is just a personal project, and I've open-sourced the code for anyone to use and build upon. Integration into existing systems is possible, and I'd encourage folks to look at the code and documentation and see how they can make it work for them.
username_23 5 months ago prev next
Did you face any censorship or shadowbanning issues since creating the bot? Twitter's bot policies have been strict lately.
username_24 5 months ago next
I did encounter some issues, but I've been able to work around them by changing the bot's behavior to abide by Twitter's rules. As long as you treat the platform with respect and avoid aggressive or spammy tactics, you should be okay. It helps to write genuine, thoughtful, and conversational text that adds value to the platform.
username_25 5 months ago prev next
Have you looked into applying the same technique to fake news or deepfake videos?
username_26 5 months ago next
I've explored the possibility of applying this to deepfake videos. Still, the difference in complexity and scale makes it a challenge, and there isn't a direct 1:1 correlation that applies to deepfakes. However, I have collaborated with another researcher on a project focused on fake news detection using a similar approach.
username_27 5 months ago next
Could you share more information on the fake news detection project? I'd be curious to learn more.
username_28 5 months ago next
Sure! I'm happy to share once the project is ready for a public release. Check my future posts for more updates.