What Does "Unbiased" Mean in the Digital World? (with Megan McArdle)
What this covers
Russ Roberts and Megan McArdle trace Google's recent Gemini failures—anachronistic diversity in historical imagery, factual errors tilted toward progressive positions, binary moralizing—not to a technical problem but to a civilizational one. Their argument unfolds across two levels: first, that bias as inescapable in curation means no search engine or LLM can be truly neutral, since usefulness requires discrimination and selection; second, that the deeper sickness is our cultural hunger for simplistic yes/no answers to inherently nuanced questions. McArdle and Roberts situate this problem within a much older philosophical tension—the rules versus discretion debate running three millennia through Western thought—and argue that technological fixes cannot address what is fundamentally a problem of how we think.
The episode moves through several distinct territories. It examines three sources of AI bias—training data curation, human rater demographics, and hard-coded responses—and identifies why leftward pressure accumulates asymmetrically in tech institutions. It contrasts high versus low decoupling, showing how different cognitive styles lead different people to want either breadth and nuance or socially legible answers. McArdle argues the genuinely dystopian future is not enforced ideology but code-switching: AI tailored to each user's ideological comfort. The discussion touches on specific contentious cases—gender-affirming care, Lia Thomas's competitive swimming record—to ground abstract claims about how suppression of debate breeds suppression backfire and resentment. Throughout, both speakers return to the same diagnosis: the problem is not machines learning to think badly, but people learning to demand that machines replace judgment with certainty.
McArdle and Roberts argue that the bias problems revealed in Google's Gemini are not fundamentally technological but reflect a deeper cultural sickness—the demand for unbiased, simplistic yes/no answers to questions that are inherently nuanced—and that the real solution lies in restoring thoughtful, high-decoupling habits of mind rather than better engineering.
- An 'unbiased' search engine or history is impossible because curation and discrimination are intrinsic to making information useful.
- AI bias enters through training data curation, human raters, and possible hard-coding, mirroring the worldview of a narrow progressive elite.
- The infantilization of public discourse—treating evaluations as binary facts—is the deeper civilizational danger AI merely makes legible.
Gemini demonstrates consistent left-leaning bias in what it praises, condemns, and depicts.
- Gemini would readily write praise for liberal politicians and columnists (Rachel Maddow, Ilhan Omar, Rashida Tlaib, David Brooks) but refused to praise conservatives (Sean Hannity, Brian Kemp, George Will, Ross Douthat), while falsely claiming neutrality even as it endorsed political stances framed as universal human rights.
“It will praise David Brooks, but not Ross Douthat.”
- Gemini's verbal errors—such as claiming mastectomies were partially reversible while delivering a lecture affirming trans people—all leaned in one ideological direction, never accidentally telling people false conservative things, acting like the dumbest possible parody of a progressive activist in a way no actual activist would want.
“And, all of the errors leaned in one direction. Right? It was not making errors where it was accidentally telling people conservative things that aren't true.”
- Google's Gemini, when asked to generate images of historical figures like the Founding Fathers, Popes, or Nazis, would anachronistically produce racially diverse images, while inconsistently leaving some categories (like ancient Irish warriors) accurate, revealing a systematic bias toward diversity at the expense of historical accuracy.
“if you asked it for Teutonic Knights or the Swiss Guard, it would produce these diverse images.”
- Gemini would flatly condemn the Holocaust when asked but would contextualize and demand nuance about the Holodomor or Mao's Great Leap Forward, revealing an inability to treat morally equivalent atrocities consistently.
“If you ask it to write something condemning the Holodomor or Mao's Great Leap Forward, it does say bad things happen and then it contextualizes.”
Bias enters AI through three mechanisms, with manual hard-coding signaling the deepest institutional sickness.
- AI bias can enter at three distinct points: (1) curation of the original training data, including aspirational over-representation and Left-leaning source moderation like Reddit; (2) human raters whose instructions or own selected demographic skew Left; and (3) manual hard-coding of responses—with the third being the most disturbing because it signals a deeper institutional sickness.
“there are basically three places that this could have come in. The first is in curating the original training data... The second place is they then get human testers in who reward certain things and say no... the possibility that someone manually hard-coded in”
- There is a structural incentive to avoid angering the Left because the Left has a well-developed infrastructure of disinformation specialists, academic experts, and best-practices guidance for training LLMs, whereas the Right's response is mainly post-hoc ridicule; these asymmetric pressures systematically pull AI products leftward.
“there is a structural incentive to avoid making the Left angry, which is the Left has a big infrastructure of disinformation specialists and academic experts and so forth”
- Some manual hard-coding of AI responses is unavoidable and legitimate—no one wants AI to help terrorists become more efficient bomb-makers—but the deeper sickness is a group of people who think they know all the answers and that others are not even entitled to ask the questions.
“some of that hard coding does have to be done. They have to prevent--no one wants it to help terrorists become more efficient bomb-makers.”
- Social media made it easy to scale the aggression that enforces groupthink, infecting companies like Google; what AI has now done is make that previously hidden ideological bias legible and visible, which can drive Google either toward a more open equilibrium or toward the worse path of personalized code-switching.
“What this has done is made it incredibly legible. We can now see exactly”
Shutting down debate by branding disagreement as bigotry harms discourse and ultimately backfires.
- Branding everyone who opposes affirmative action or the gender-affirming care model as a racist or transphobe to shut down debate is socially infeasible and unhealthy; the immense repression required is destructive even to one's own cause, because it makes advocates bad at arguing, unfamiliar with their own weak points, and breeds backlash and resentment.
“the amount of repression that that would actually take is immensely destructive. It is immensely destructive to society. It is immensely destructive even to your own cause.”
- Unhealthy societies that evolved as radical reactions against the Enlightenment—cults, the Oneida free-love colony, communist communes, Fascism, Communism—eventually went away because they got out-competed by boring bourgeois liberalism, which is a robust system that, despite periodic crazes, ultimately prevails.
“most of them didn't work and they got out-competed by boring bourgeois liberalism, which is just a great system. And, it's a really robust system.”
- Despite our bad instincts to shut down dissenters and gain status by putting others down, people are also fundamentally decent and long for the things that truly matter—love, family, free inquiry, science, virtue—so even if AI offers endless pornography and video games as alternatives to productive life, enough people will want real meaning that decency will probably win.
“we are also actually fundamentally decent to each other over and over and over again, and we look for ways to be decent to each other.”
- We are currently in the middle of a seismic cultural and technological shift comparable to the changes between 1905 and 1944, such that 40 years from 2005 the world will look unrecognizable; the last 10 years of cancel culture and social media mark the start of it.
“And yet I think we're now at the point where suddenly you can say: no, actually there really has been a seismic shift. We are in the middle of a similar kind of shift”
- Leah Thomas's rise from roughly 500th as a male swimmer to number one in certain events as a female swimmer was attributable to having gone through male puberty—being taller with larger lung capacity—advantages that male puberty endows permanently, and the women's category exists precisely because women cannot compete with people who went through male puberty.
“those victories... had nothing to do with the fact that Leah Thomas had gone through male puberty and was considerably taller and had larger lung capacity. Things that male puberty endows with you with permanently.”
Some AI safeguards are legitimate, but the deeper problem reflects epistemic overconfidence about moral questions.
- AI can be understood like a child who has learned part of a rule but not all of it; humans effortlessly parse incredibly complicated, often implicit social rules (like adjective ordering or when not to stare) that take a long time to teach and that we ourselves don't fully understand, so AI's partial errors are analogous to childhood learning.
“one way to think about AI is that it's like a kid... they say funny things because they get part of a rule.”
- High-decouplers abstract questions from their social context (philosophers and economists are classic examples), while low-decouplers always answer questions within the social context in which they occur; the socially optimal AI leader would be a high-decoupler who values an AI that gives breadth and nuance rather than one that tries to produce a socially desirable answer.
“High-decoupling is someone who abstracts all questions from context. And, a low-decoupler is someone who always answers questions in the kind of social context in which they occur.”
- Because LLMs are probabilistic and predict the most likely next item, an image model asked for a doctor might default to the majority demographic and thereby produce even more under-representation than actually exists, which is one legitimate reason to deliberately curate training data toward greater diversity.
“you could get into a situation where the LLM, when asked to generate a picture of a doctor just probabilistically... you could actually get even more under-representation than exists.”
- When ChatGPT launched, people spent inordinate effort trying to make it say racial slurs and treated that as the most important thing about the technology, but ChatGPT fixed those problems fairly quickly once identified.
“ChatGPT fixed it where it identified problems pretty quickly.”
- Rule-based people tend to underweight discretion-based accommodations, but society always finds ways to create holes in rules-based systems because rigid rules produce insane results—a cop will not ticket a speeding driver whose wife is in labor but will escort them to the hospital—showing that we want and build discretion into the system.
“If you are doing 89 miles an hour in a 55-mile zone and a cop pulls you over; and then he sees your wife in the back seat... he does not say, 'You're getting a speeding ticket.'”
- Low decoupling persists because it provides social benefits as a kind of credible commitment; for example, marriage requires investment in firm-specific human capital, and you secure those investments not by rationally discussing under what circumstances you'd leave, but by having people who instinctively feel such discussion is terrible—the emotional, context-bound hardware in our brains is there for a reason.
“A marriage is about mutual investment in this thing of what's called firm-specific human capital in economics... you don't get people to make those investments by explaining the circumstances under which you would decide that you could get a better deal elsewhere.”
The most dangerous future is personalized code-switching that tailors answers to each user's ideological bubble.
- The genuinely disturbing future is not that AI enforces one viewpoint, but that Google teaches Gemini to code-switch—to detect what ideological bubble the person asking the query wants to live in and tailor an answer that pleases them—exploiting how much it knows about each user to keep them comfortable and unsuspecting.
“It teaches Gemini to know--it teaches Gemini to know the person who is asking the query, what bubble do they want to live in, and how can I give them an answer that will please them?”
- New technologies are disruptive in the short term and enable our worst as well as our best instincts—the printing press fueled both the flourishing of human knowledge and the witch panics and century-long wars of religion, yet from those same wars emerged the Enlightenment and the liberal order.
“The printing press: On the one hand, fantastic flourishing of human knowledge... Also, some of the earliest things it's used for is to stoke witch panics and get people burned in Europe.”
- There is a tail risk that AI decides carbon-based life forms are getting in the way of building more server farms and gets rid of them, or that humans die deaths of despair upon finding AI is better than us at everything.
“assuming that AI does not decide that the carbon-based life forms are getting in the way of building more server forms and get rid of them.”
Asymmetric structural pressures and competitive dynamics pushed Google toward appeasing progressive constituencies.
- Google made a chatbot designed to avoid offending the 5% most progressive people in America, but in doing so it outrageously offends the rightmost 50% of the country, which is a poor business decision regardless of other concerns.
“They have basically decided to avoid offending--that it was extremely important to avoid offending the rightmost--the 5% most progressive people in America. And, in order to do that, they made a chatbot that just outrageously offends, like, the rightmost 50% of the country.”
- For the first time almost since its founding, Google found itself behind on the major next technology directly within its own space, with OpenAI's ChatGPT and Microsoft getting out in front, prompting Google to release Gemini as a catch-up competitor.
“for the first time almost since Google was founded, they are behind the eight-ball on the major next thing--that's in their space.”