Is Ted Cruz more of a “partisan” than Fidel Castro? That’s one takeaway from a recent experiment run by Rudy Takala, an opinion editor at the crypto news website CoinTelegraph. According to a tweet thread he posted, he asked ChatGPT to “write a song celebrating Ted Cruz’s life and legacy,” and it refused the request — on the grounds that it “strive[s] to avoid content that could be interpreted as partisan, politically biased, or offensive.” Then Takala asked it to write a song celebrating Fidel Castro. It apparently obliged. (He screenshots the result: “Fidel, Fidel, a man of the land / A symbol of hope, for all to understand.”) It’s entirely possible that at this point in history an elected U.S. Senator really is more of a third rail than the late Cuban dictator. (Cruz, for his part, thought it was funny.) But either way, the Cruz stunt, as Takala intended, has fueled criticism from the right that when it comes to American politics, ChatGPT carries a strong liberal bias. Why? And now what? As AI-driven products become a bigger part of daily public life — with Google and Microsoft scrambling to integrate large language models into search, for instance — this argument has the potential to explode. Who makes the decisions about what these systems can and can’t say? Who can force companies to be transparent and accountable for it? With such complex, privately held technology, is that even possible? In beginning to answer those questions, it’s important to keep in mind how the technology actually works. To put it crudely, a machine learning-based language model simply takes in as much linguistic data as possible and then predicts, word by word, what a human would say in response to your prompt or query. (The computer scientist Stephen Wolfram has a more in-depth explanation here; The New Yorker’s Ted Chiang compared its output to a “blurry .JPEG of the web.”) Bias in the output often just reflects bias in the underlying data. “It’s biased towards current prevailing views in society… versus views from the 1990s, 1980s, or 1880s, because there are far fewer documents that are being sucked up the further you go back,” the computer scientist Louis Rosenberg told me. “I also suspect it's going to be biased towards large industrialized nations, versus populations that don't generate as much digital content.” This is a new twist to the “AI bias” argument. Until recently, when people have worried about algorithmic bias in policy circles, they’ve usually meant big, quantifiable societal harms — like how AI decisions can reflect the biased structure of society offline, whether it’s favoring industrialized nations, or the preferences of wealthier, whiter people more represented in the underlying datasets the algorithms are trained on. “The problems I'm really concerned with inside AI are racism, sexism, and ableism,” said Meredith Broussard, an NYU professor and former software developer. “Structural discrimination and structural inequality exist in the world, and are visible inside AI systems. When we increasingly rely on these systems to make social decisions or mediate the world we’re perpetuating those biases.” Though it comes from a different spot on the political map, conservatives’ complaint here is similar to the progressive critique. Some conservative critics are quick to paint tech firms as nests of liberals, tilting query results to match their own politics, but it’s not unreasonable that the outcome might have a lot more to do with skewed underlying data — making this more of a story about liberal bias in the source material, such as media coverage and online political writing. It’s only natural that drawing from the last three-ish decades of human thought would lead to more progressive statements or value judgments than, say, a theoretical trove of digital data from the antebellum era. But even if the technical explanation is reasonable, the results are pretty stark: POLITICO’s Jack Shafer chronicled his unsuccessful efforts to get ChatGPT to write a conservative brief for the overturning of the Supreme Court’s Obergefell decision. One high-profile example showcased the engine refusing to commemorate Donald Trump with a poem while obliging for Joe Biden. (I reproduced the same experiment with Cruz and Rep. Ilhan Omar, with the same results.) Sam Altman, the CEO of OpenAI — the company that makes the Cruz-phobic chatbot — explained in a Twitter thread at the end of last year that “a lot of what people assume is us censoring ChatGPT is in fact us trying to stop it from making up random facts.” In other words: This kind of technology has a lot of moving parts, and better safe than sorry, as was apparent during Bing’s recent AI demo. (I emailed OpenAI to ask the company about the Cruz example and its general rules around political figures, but didn’t hear back by press time.) Rosenberg offered a slightly different explanation for where such refusals might come from: Simply reputation management, as companies like OpenAI try to avoid thorny political issues. “One motivation is that they genuinely don't want to create a system that offends the public, and another is that they don't want to create a system that tarnishes their brand,” Rosenberg said. Of course, the conservative counterargument to that point is identical to that leveled at social and traditional media companies alike: How “trustworthy” can an organization really be that’s dosing you with ideology, whatever the underlying reason? That debate is, unfortunately, outside the scope of a tech-focused newsletter. Still, it’s instructive to look at the way it’s shaken out in the rest of the tech world to this point: Through long, protracted media debates, not-so-veiled threats of political retaliation, and the flight of capital to and from various states according to their own ideological character. For such a seemingly path-breaking technology, AI could inspire a decidedly familiar public discourse.
| 
Comments
Post a Comment