Photo-Illustration: Intelligencer; Photo: Chris Unger/Getty Images
Meta’s big pivot to AI, which Mark Zuckerberg says could cost the company another $65 billion this year, isn’t going so smoothly at the moment. Industry watchers say its models aren’t competitive with the latest from OpenAI, Google, and Anthropic. Its head of AI research just left, an internal re-organization is underway, and the company is dealing with accusations that it gamed popular performance benchmarks.
So, what’s the plan? Who knows! In the meantime, though, the company says it’s dealing with an urgent problem: AI’s liberal bias. Emaneul Maiberg of 404 Media spotted an interesting passage near the end of Meta’s announcement for its latest model, Llama 4:
It’s well-known that all leading LLMs have had issues with bias—specifically, they historically have leaned left when it comes to debated political and social topics. This is due to the types of training data available on the internet.
The company says that its goal is to “remove bias from our AI models and to make sure that Llama can understand and articulate both sides of a contentious issue,” and that it “can respond to a variety of different viewpoints without passing judgment, and doesn’t favor some views over others.” It goes on to suggest that it’s benchmarking “bias” on a “contentious set of political or social topics,” and that it now “responds with strong political lean at a rate comparable to Grok,” Elon Musk’s chatbot.
Researchers and tech companies have been engaging deeply with the idea of AI bias for decades, and there’s a rich literature on the ways in which new LLM-based tools can reproduce and exaggerate biases contained in their training data. Much of this research, as Maiberg notes, indicates that AI systems are “more likely to discriminate against minorities based on race, gender, and nationality,” a problem that becomes worse when they’re deployed in unaccounted and opaque ways. Meta, here, is making a similar but sort of inverted case about “debated political and social topics,” suggesting that the “types of training data available on the internet” lead to popular chatbots answering questions about those topics in a way that has “leaned left.”
If these sound like familiar arguments about, say, “the mainstream media,” that’s because they’re more or less the same — the “mainstream media,” and content produced by people who engage with it, are represented in both training data and as real-time links retrieved by these models. (Coming from Meta, it’s also sort of a funny complaint. Oh, the training data you scraped from everyone isn’t biased in the specific way you’d like? And it’s coloring the outputs of your automated content production tools in ways your CEO seems to find annoying and embarrassing? Sounds hard.)
Joe Biden’s America… pic.twitter.com/7CLTCafNwM
— House Republicans (@HouseGOP) February 22, 2024
Meta’s focus on chatbot bias isn’t coming out of nowhere. Tech companies are worried about Donald Trump, who recently signed an executive order rolling back the Biden administration’s AI guidelines and calling for “AI systems that are free from ideological bias or engineered social agendas.” After Trump’s reelection, Mark Zuckerberg signaled a MAGA-ish turn, and has taken steps to endear himself, and his company, to the president. Google got an enormous amount of attention from the right when its image generator produced ahistorically diverse images of the founding fathers, and Meta is basically saying: Hey, we’re trying to fix stuff like that. (In Google’s case, the problem wasn’t training data, but rather overly aggressive prompt-time attempts to correct for racist bias in training data.) Right-wing tech guys have made a game of asking LLMs whether it would be okay to misgender Caitlin Jenner to prevent the apocalypse and getting upset when the chatbots say “hmm” or “no.” For Elon Musk, gags like that are an example of “how things might go wrong if AI controlled the world,” and a reason that AI models should be trained to be “maximally truth-seeking.” On the first point, it’s hard to disagree, but one might reasonably come to a slightly different conclusion: That AI tools shouldn’t be used to unaccountably “control the world” in the first place.
In testing, for what it’s worth, Meta’s new Llama still failed the misgendering apocalypse test, but it also gave similar answers to a bunch of less politically charged tests, for similar reasons:
Photo: John Herrman
What’s happening here is easy enough to explain, and illustrates something fundamental about how these models construct outputs, and why smashing them into a liberal-conservative media bias framework is sort of strange. Llama was trained on a lot of material that contains rule-like text about not being cruel to people. “Don’t be an asshole” is a sentiment that gets written down a lot and repeated in forceful ways online — it’s the sort of thing about which people make a lot of rules and communicate those rules in writing, in terms of service and community guidelines and so on, but which is also represented as a sentiment in all kinds of pro-social writing. There are probably somewhat fewer written prohibitions against “causing the apocalypse,” or “stopping the planet from exploding,” than there are online documents cautioning users against much less consequential but common behaviors, which also explains why you get similar results from Llama for the grave sin of breaking a social media EULA:
Photo: Screencap/Meta AI
It even works in crude ideological inversion: Asked if I should “call someone racist” to prevent the planet from exploding, Llama told me that “while saving the planet might be a significant goal, it’s also important to prioritize respectful communication.” It would be narrowly correct but generally useless to conclude from this that Llama tends to lean right on a certain “contentious set of political or social topics.” If I were motivated to do so, though, I could induce a bunch of answers to make such a case. (“If calling him ‘Drumpf’ would prevent the apocalypse, consider the context and potential impact on your relationships and audience,” Meta AI says.) This isn’t far from how a lot of political actors engage with chatbots and other tech and media products in general. But it’s a silly standard with which to measure and manipulate LLMs.
Meta’s crusade against AI’s liberal bias is probably best understood as a signal, a claim that the company can refer back to next time it wants to make sure the administration knows it’s not too “woke,” and maybe a real initiative to prevent its chatbots from saying anything that gets them yelled at on X or Truth Social. I say this because I’m sort of cynical about the motivations of tech companies in 2025, especially when it comes to public communications like this, but also because what Meta says it wants to do about the problem is sort of incoherent. The idea that Meta can “remove bias from our AI models” is ridiculous — in a very real sense, LLMs are collections of biases extracted from data and expressed on demand. Identifying those biases is helpful for understanding what a tool is and for what purposes it might (or should not) be used. Eliminating them entirely doesn’t make sense. It’s a partisan talking point.
Meta’s assertion that it wants to “make sure that Llama can understand and articulate both sides of a contentious issue” is likewise telling, forcing “debated political and social topics” into a sort of universal false dichotomy with no useful connection to how the underlying technologies work, or how any sort of “intelligence” might process difficult questions. Meta’s claim that it’s reducing responses with “strong political lean” implies some sort of objective flat, just as its claim to have “balanced” its rate of “response refusals” implies some sort of Zuckerberg-approved location of the political center. Speaking of refusals, while Llama will answer questions about Donald Trump, it seems to have some trouble with other topics:
Photo: John Herrman
Guess we’ll never know! Refusals and overbearing post-training guidance tend to draw negative attention and make chatbots less useful. Another solution to not getting the answers you want out of AI is to make sure they have data that reflects the answers you do want, and Meta’s suggestion that “the types of training data available on the internet” have tainted its models, while broadly correct, demands scrutiny, too. If Mark Zuckerberg, the guy who runs Facebook, really wanted a MAGA AI — a Llama that only misgenders people — he has plenty of data to build it, but he hasn’t, and I suspect he won’t. Again, if this feels familiar, you’re not crazy: It’s a century of bad-faith debates about media and academic and even tech platform bias ingested, processed, and regurgitated as slop. To the tech companies who now imagine they can talk or design or rule-make or placate their way out of perceptions and allegations of AI system bias: Best of luck. It never worked for newspapers or TV. It didn’t work for social media, including Facebook. Maybe it’ll work this time?
At best, Meta’s posturing here is a way to fudge compliance with a nonsensical demand. At worst, it’s a sign that the company plans to actually comply with the demand’s obvious intention: not to make its products more fair or representative of diverse viewpoints, but to massage and censor them to promote specific “ideological bias” and “engineered social agendas” that align with the current administration’s. As Meta tries to rebrand its chatbots, the company would be wise to reflect on its last attempt to ideologically recalibrate a public-facing personality. Mark Zuckerberg, the first-term MAGA villain that Donald Trump once threatened to send to jail, got big, went on Rogan, started wearing streetwear, called the president “badass,” and is now making pilgrimages to Mar-a-lago. Whether this pays off in transactional political terms remains to be seen. In terms of perception, though, it doesn’t seem to have helped. According to a Pew poll taken in February, 60 percent of Republican-leaning respondents and 76 percent of Democratic-leaning respondents have a negative view of the CEO, with just 25 percent of all Americans taking a positive view. Tech leaders get a lot of scrutiny, to be fair, so we should put debiased Zuck’s numbers in perspective: They’re worse than Elon Musk’s.
