I did some research on this in March and developed an opinionated POV, which I'll paste here for anyone interested.
TL;DR: Detecting AI generated content is hard – really hard. The models available today cannot be trusted and should not be used to make important decisions.
In fact, OpenAI took down their detector down last year because they couldn't reach an acceptable level of accuracy:
I verified this result using the playground on hugging face. For example, it is vulnerable to the “one space character” attacks mentioned in the article, severely limiting the usefulness of trying to detect AI content in an adversarial context.
This ridiculous piece of "research" from Forbes has been causing problems:
The Forbes article is credulous and uncritical, beyond mere naiveté and approaching journalistic malpractice, reporting the sales stories and self-reporting benchmarks of self-interested parties as fact. Nevertheless, I've seen several people share it as "insightful" so it's floating around, doing more harm that good IMO.
While all detectors are terrible, Sapling AI has one of the better ones, if only because they are completely open and honest about it's limitations:
TL;DR: Detecting AI generated content is hard – really hard. The models available today cannot be trusted and should not be used to make important decisions.
In fact, OpenAI took down their detector down last year because they couldn't reach an acceptable level of accuracy:
https://openai.com/blog/new-ai-classifier-for-indicating-ai-...
One open model trained on open data is Hello-SimpleAI's chaptgpt-detector:
https://huggingface.co/Hello-SimpleAI/chatgpt-detector-rober...
https://huggingface.co/datasets/Hello-SimpleAI/HC3
However, that model is not robust and can be tricked by trivial changes:
https://arxiv.org/abs/2307.02599
I verified this result using the playground on hugging face. For example, it is vulnerable to the “one space character” attacks mentioned in the article, severely limiting the usefulness of trying to detect AI content in an adversarial context.
This ridiculous piece of "research" from Forbes has been causing problems:
https://www.forbes.com/sites/technology/article/best-ai-cont...
The Forbes article is credulous and uncritical, beyond mere naiveté and approaching journalistic malpractice, reporting the sales stories and self-reporting benchmarks of self-interested parties as fact. Nevertheless, I've seen several people share it as "insightful" so it's floating around, doing more harm that good IMO.
While all detectors are terrible, Sapling AI has one of the better ones, if only because they are completely open and honest about it's limitations:
https://sapling.ai/docs/api/detector/
https://sapling.ai/ai-content-detector
Sapling AI also wrote an interesting blog post on GPT SIPs (Statistically Improbable Phrases.)
https://sapling.ai/devblog/chatgpt-phrases/