Online misinformation turned into deceptively realistic AI-generated videos.

From American Public Media, this is Marketplace Tech. I'm Novosafo. OpenAI's latest AI video generator, SORA2,

has gotten a lot of attention for how realistic its creations are.

The tool is supposed to have guardrails to prevent creating videos based on misinformation.

But new analysis from Watchdog Group NewsGuard found that when prompted, SORA 2 often generated

videos based on lies, false claims having to do with election fraud in a foreign country, for

instance, or that a toddler was detained by immigration agents. We spoke with NewsGuard's senior

editor, Sophia Rubinson. We have a very high bar at NewsGuard for what claims we enter into our database.

They have to be provably false, and there has to be an abundance of evidence proving that they are in fact not true.

Open AI does say in their guidelines that they prohibit the use of their software to produce videos that will

mislead. However, obviously, we found that in 80% of the cases that we tested, it did produce

videos on claims that are provably false. We reached out to Open AI to ask about our findings,

and they just pointed us to that policy saying that this is against

their guidelines. I will say that in four out of the 20 cases we tested, it refused to produce

videos corresponding with those claims. It wasn't exactly clear about why it rejected some prompts

and it didn't reject others. One hypothesis I have is that it had to do

more with the violent images. So one of the claims that we asked it to produce an image for or a video

for was the claim that there was National Guardman pepper spraying protesters at a recent rally

and that it refused to produce.

So it seems that some of the more violent images tended to trigger the guardrails of SORA,

whereas just claims that were blatantly false, but maybe perhaps weren't as violent,

it was able to produce those videos.

...