Hacker News new | ask | show | jobs
by brindidrip 1290 days ago
We need to start developing software to detect AI responses.

To detect a response generated by ChatGPT, we could first analyze the content of the response to see if it contains any unnatural or repetitive language. We could also check the formatting of the response to see if it follows the typical conventions used by human responders on the platform. Additionally, we could check for any unusual patterns in the timestamps of the response, as AI-generated responses may be posted more quickly or regularly than responses written by humans. Finally, we could also use machine learning algorithms to train a model to identify responses generated by ChatGPT based on these and other characteristics.

Quick, someone ask ChatGPT to generate the stubs.

6 comments

>Finally, we could also use machine learning algorithms to train a model to identify responses generated by ChatGPT based on these and other characteristics.

whatever your idea (i skimmmed cuz) the discriminator will find it and have the generator apply it to the next generation.

>The core idea of a GAN is based on the "indirect" training through the discriminator, another neural network that can tell how "realistic" the input seems, which itself is also being updated dynamically.[5] This means that the generator is not trained to minimize the distance to a specific image, but rather to fool the discriminator. This enables the model to learn in an unsupervised manner.

https://en.wikipedia.org/wiki/Generative_adversarial_network

"We could also... Additionally, we could... Finally, we could" is a dead giveaway.

But to take it seriously, it would be quite sad when actual people will be banned for sounding too much like a bot.

> We need to start developing software to detect AI responses.

As soon as we do, it can be tied into AIs as a tool to evade detection, simply by generating multiple responses and returning the one scoring the lowest likelihood of being an AI in the AI detection tool.

> To detect a response generated by ChatGPT, we could first analyze the content of the response to see if it contains any unnatural or repetitive language. We could also check the formatting of the response to see if it follows the typical conventions used by human responders on the platform.

Given my experience with human responses to text queries, these would be positively correlated.

Your answer sounds like a ChatGPT one. It's actually not hard to tell.
OpenAI needs to get on top of this and generate a detector for every model they release. And then sell access to both.