“Photos” of the Pope in a puffer jacket and recordings of superstars “singing” new tunes have alerted policymakers and the public to how artificial intelligence can be used to generate highly realistic but misleading content. In studies released Tuesday, the non-profit Partnership on AI (PAI) examined how organizations like OpenAI, TikTok and the CBC are contending with the rise of “synthetic media”—a term for content that has been artificially generated or modified. Here’s what you need to know:

The synthetic moment: AI has increased “the accessibility, the sophistication, the pace and the volume” of manipulated content designed to mislead, said Claire Leibowicz, PAI’s head of AI and media integrity, in an interview. A year ago, the group published guidelines for it, including urging those who build generative tools to be transparent and implement disclosure mechanisms for synthetic media. 

Dali or DALL·E?: OpenAI’s case suggests this is all still very much a work in progress. 

The San Francisco-based startup has been working for about a year on ways to show whether an image comes from one of its DALL·E models. It considered adding watermarks to the pictures, a visual signal of syntheticity; Canada’s voluntary generative AI code of conduct suggests this method. OpenAI also explored embedding the origin information in the metadata of the files its tool produces.  

But the firm concluded that bad actors could simply get around those safeguards by editing or screenshotting the picture to remove the DALL·E disclosure. Instead, OpenAI is building a classifier that it can use to determine whether a particular image came from its model. 

Even that comes with complications—OpenAI worries it could give policymakers and consumers a false sense that the problem of identifying synthetic media has been solved for good. And if the firm discloses how the tool works, bad actors will try to get around it. Still, OpenAI said it may start letting journalists or researchers try the classifier in the first half of the year. (The firm launched and then pulled a similar tool for text because it wasn’t very good.)

“There’s no silver bullet,” said Leibowicz. But generative firms need to figure out disclosure so other platforms and, ultimately, consumers can figure out what’s real and not. For example, OpenAI’s classifier could help Bumble spot AI-made photos to flag fake dating profiles, another PAI case study.

What comes next: Governments around the world are proposing new laws and codes for generative AI. Digital rights advocates have argued that tech firms are using voluntary commitments to evade or shape regulation. Leibowicz acknowledged that opt-in frameworks like PAI’s lack teeth, but said the organization and its members are filling a gap while policymakers catch up. “It is our hope that this work will feed into government regulation.”