[LINK] OpenAI just admitted it can't identify AI-generated text.

[gerard]

Hi Jan


They have been trying a watermark:

https://docs.kanaries.net/articles/chatgpt-watermark


But there are already potential workarounds.


Gerard


On 29/7/23 09:32, Jan Whitaker wrote:
> Wouldn't it be simpler to embed a metadata element that identified the 
> source of an output from one of the generators? Including which one, 
> when, for whom (trickier if local installation) and possibly a unique 
> generation code for that item? I realise that wouldn't work for screen 
> shots or copy/past text (at least I can't see how that could be 
> stopped), but at least the original could be coded as such so if there 
> were any challenges to authenticity, there could be tracking for it. 
> Or maybe that won't scale.
>
> On 29/07/2023 12:23 am, Stephen Loosley wrote:
>>
>> Less than seven months later, the project was scrapped.
>>
>> "As of July 20, 2023, the AI classifier is no longer available due to 
>> its low rate of accuracy," OpenAI wrote in a recent blog. "We are 
>> working to incorporate feedback and are currently researching more 
>> effective provenance techniques for text."
> _______________________________________________
> Link mailing list
> Link at anu.edu.au
> https://mailman.anu.edu.au/mailman/listinfo/link