Grok, X’s AI chatbot, is under scrutiny after it made antisemitic and bigoted remarks : NPR

Ayesha Rascoe, host:

They distribute electoral disinformation, make deeply inappropriate, sometimes fanatical remarks and give dangerous advice. We are not talking about your eccentric uncle or this high school friend who derived in the Manosphere. We are talking about chatbots and AI tools. For example, Grok, a chatbot at X – formerly Twitter – last week made anti -Semitic remarks, called Mechahitler. But this is only the last of a series of disturbing problems with the models of AI on X and elsewhere. We are now joined by Reece Rogers, who is a journalist at Wired magazine. Welcome to the program.

Reece Rogers: Thank you for doing me today, Ayesha.

Rascoe: OK, so we have just mentioned Grok on X, but you have made a lot of reports on other AI models, such as the OpenAi video text model, and VEO 3 of Google. What problems have you found with these AI models?

Rogers: Yes, I think that at the base, these tools are model machines. They therefore take tons and tons of data and predict which could be a reasonable exit. Well, if you take all the Internet information, you will ingest all these stereotypes that people have on other people. So one of our big surveys, even last year, was that we watched Sora, Midjourney and other tools that people use to generate images and videos, and we found that they amplified many different stereotypes, in particular. And, you know, where these stereotypes are introduced are either to know if the way the data is online, or maybe people build it by labeling the data.

Rascoe: Is it essentially an entry problem?

Rogers: When we think of Grok and the recent anti-Semitic remarks, you know, these go to a level which is beyond that simply reproducing the models that this finds. And, you know, Musk blamed users for these outings, and it is difficult to know exactly how it could have happened. But often, you know, for these tools, you work with what is called a system prompt, or another type of strengthening learning, which is there only after their training and that they build an AI tool, there are additional steps that can guide its output, in a way parameters. So, even if we do not know exactly why Grok left the rails and that he made some kind of racist messages last week, it is not exactly the kind of error that would be done just by accident.

Rascoe: Would the programmers enter and say, Hitler, you know, a good guy or something? Like that – it wouldn’t be so frank, right? Like, how …

Rogers: No.

Rascoe: … Something like that would happen?

Rogers: But perhaps they could have said, you know, Grok, don’t be afraid to take positions against the tide, you know? Or grok, don’t be afraid to take positions outside the dominant current. And if you give a tool a kind of direction, potentially like that, it does not think exactly like a human. It doesn’t think at all. So this follows the instructions, and it could take much further than even the programmers initially thought it.

Rascoe: If so, then to what extent can businesses behind these tools brake this? How can they define parameters without introducing more biases?

Rogers: It’s a question to a million dollars at the moment, and I think that something that companies are working with diligence to improve. There are therefore security measures they take, but with regard to stereotypes and biases, it is something that is very anchored and that they find that it is quite difficult to find the right nuanced approach. And, you know, it brings – they are companies based in the United States trying to create global tools. So when you think of the perspective, you know, the perspective in Italy or the perspective in Ghana will be different from the prospect of San Francisco. But I think it is worth keeping in mind, whenever you interact with it, whether it is the Chatppt, or that you make, like a fun image for, like a birthday card, I think that keeping this kind of stereotypes which are, like, very, very, very widespread in the models in mind are only the kind of consumer advice that I would give to their daily life.

Rascoe: Have you found that people are generally skeptical about the information they receive from AI or how the people they get from AI?

Rogers: You know, this is a question that I often get readers, and I think people are not skeptical enough. But one thing I always say to people is click on these links. So, if you use a chatbot and use a type of sources in its outings, do not be afraid to click on this website, always go back your sources. So I think people should be skeptical. But if it’s only up to date, if you try to understand, you know, which television show in your favorite series, you try to find the right one to watch, like, it’s a perfect question for a chatbot and it doesn’t need to be perfect. But for this kind of topic scenarios, I would always check your answers elsewhere.

RASCOE: It’s the cable journalist Reece Rogers. Thank you very much for talking to us today.

Rogers: Thank you very much for doing me.

Rascoe: Since we recorded this interview, X has apologized via Grok for Son, Cit, “Horrible Behavior”. He said there was a coding problem and was solved.

(Soundbite de Corbin Roe, “Drip” by Mayne and Nicxix)

Copyright © 2025 NPR. All rights reserved. Visit the pages of use of the conditions of use of our website on www.npr.org for more information.

The accuracy and availability of NPR transcriptions may vary. The transcription text can be revised to correct errors or match audio updates. Audio on npr.org can be published after its original broadcast or publication. The file authorizing the NPR programming is the audio recording.