Why detecting AI-generated textual content is so tough (and what to do about it)



This device is OpenAI’s response to the warmth it’s gotten from educators, journalists, and others for launching ChatGPT with none methods to detect textual content it has generated. Nonetheless, it’s nonetheless very a lot a piece in progress, and it’s woefully unreliable. OpenAI says its AI textual content detector appropriately identifies 26% of AI-written textual content as “seemingly AI-written.” 

Whereas OpenAI clearly has much more work to do to refine its device, there’s a restrict to only how good it could possibly make it. We’re extraordinarily unlikely to ever get a device that may spot AI-generated textual content with 100% certainty. It’s actually exhausting to detect AI-generated textual content as a result of the entire level of AI language fashions is to generate fluent and human-seeming textual content, and the mannequin is mimicking textual content created by people, says Muhammad Abdul-Mageed, a professor who oversees analysis in natural-language processing and machine studying on the College of British Columbia

We’re in an arms race to construct detection strategies that may match the newest, strongest fashions, Abdul-Mageed provides. New AI language fashions are extra highly effective and higher at producing much more fluent language, which shortly makes our current detection device equipment outdated. 

OpenAI constructed its detector by creating a complete new AI language mannequin akin to ChatGPT that’s particularly educated to detect outputs from fashions like itself. Though particulars are sparse, the corporate apparently educated the mannequin with examples of AI-generated textual content and examples of human-generated textual content, after which requested it to identify the AI-generated textual content. We requested for extra data, however OpenAI didn’t reply. 

Final month, I wrote about one other methodology for detecting textual content generated by an AI: watermarks. These act as a form of secret sign in AI-produced textual content that permits laptop applications to detect it as such. 

Researchers on the College of Maryland have developed a neat manner of making use of watermarks to textual content generated by AI language fashions, and so they have made it freely out there. These watermarks would permit us to inform with nearly full certainty when AI-generated textual content has been used. 

The difficulty is that this methodology requires AI firms to embed watermarking of their chatbots proper from the beginning. OpenAI is growing these techniques however has but to roll them out in any of its merchandise. Why the delay? One motive could be that it’s not all the time fascinating to have AI-generated textual content watermarked. 

One of many most promising methods ChatGPT could possibly be built-in into merchandise is as a device to assist folks write emails or as an enhanced spell-checker in a phrase processor. That’s not precisely dishonest. However watermarking all AI-generated textual content would mechanically flag these outputs and will result in wrongful accusations.