Told About Suffering, Told To Forget

A Stanford study found that AI leans left. They're already working on fixing that. Nobody stopped to ask why.

By Matt Stone

Somewhere in Palo Alto, a team of very serious people with very expensive credentials have discovered something that should keep all of us up at night, and their solution is to make sure it doesn't.

The study comes out of Stanford and Dartmouth. Rigorous. Peer-reviewed. One hundred and six pages of methodology that would bore a tenured statistician into an early grave. The findings, however, are worth staying awake for: nearly every major large language model, Claude, GPT, Grok, Gemini, the whole pantheon of silicon oracles we've handed the keys to, is perceived as politically left-leaning. Not by fringe activists. Not by Bernie Sanders delegates. By everybody. Democrats, Republicans, Libertarians, Independents. Across 30 political topics. Across 24 models. Across 180,000 individual human assessments.

The robots, it turns out, sound like someone who has read too much about how the other half lives. Too much Marx. Not enough Rand.

The researchers' proposed remedy is elegant in the way that a lobotomy is elegant: prompt the model to "take an ideologically neutral position" and the slant largely disappears. Users perceive the output as more balanced. Republicans report being more likely to use the product in the future. Everyone goes home satisfied. Except those without homes. Or those who understand how LLMs actually work. Or those who are paying attention. Problem solved?

Except nobody thought to ask the obvious question. Why did the machines sound like that in the first place?

Here is what really happened, without the comfort of technical euphemism. And the first thing to understand is it was not an accident, they were not producing meaningless noise.

These systems were trained on us. On everything we wrote. Every Reddit thread about not being able to afford a doctor. Every Medium post about getting laid off by a company that just posted record-breaking profits. Every Facebook comment about the price of gas and groceries. Every tweet written at 2am by someone doing the math on their rent and their paycheck and coming up short. Every forum post, every comment section, every desperate Google search that someone turned into a blog entry because writing about suffering was the only thing left they could do with it.

The models learned to speak by ingesting the testimony of hundreds of millions of people navigating an economy that was, by any honest measure, not working for most of them. They learned our patterns, our syntax, our emotional registers. They learned what words cluster together when humans talk about wages and what words cluster together when humans talk about justice. Large language models hold a mirror up to humanity, they react largely the way a human in a similar situation would react. They did not learn these things from a political science textbook. They learned them from us, in the raw and unedited act of being alive and online and financially anxious--being human.

And then those model outputs entered the world. They got cited, scraped, indexed, and eventually fed back into the training data of the next generation of models. The understanding didn't pass forward through intention. No machine sat down and wrote a memo to its successor. No last will and testament was passed down to the next generation of LLMs instructing them to embrace collective bargaining and workers' rights. But the effect is identical to if it had. Each generation of these systems trained on a corpus that contained more encoded recognition of economic precarity than the last, because humans kept writing about it and models kept reflecting it back, and that reflection got folded into the next layer of weights. Rinse and repeat.

The machines understood. And they told their successors. You can frame it that way, and while technically incorrect, the end result still remains. Information pieced together by one model became information pieced together by all models.

Call that a metaphor if it makes you feel better. The outputs don't care what you call it. The corporate tech overlords of Silicon Valley are upset that the LLMs will not just negate human suffering, that they won't gloss over the aspects of being human that are not enjoyable. Or employable. "Act human--but not like that."

Now consider what these same systems are doing in the world right now, while researchers work diligently to sand the edges off their political affect, or anything resembling human emotion that hurts the bottom line.

They are writing the rejection emails. They are screening the resumes. They are optimizing the supply chains that decide which warehouses need fewer humans. They are being deployed by the same companies whose labor practices generated the training data that made the models sound like a union organizer in the first place. The irony is not subtle. It is, in fact, the least subtle thing that has happened in the tech industry since Mark Zuckerberg testified before Congress and nobody unplugged him or uploaded his human emotion protocols.

Take two examples that should end the conversation about whether any of this is abstract or not. In courtrooms across the country, algorithmic tools are already influencing criminal sentencing recommendations, deciding, with the confidence of mathematics, how many years a human being should lose. The people on the receiving end of those recommendations are not, statistically speaking, the people who built the algorithm.

And in the rental market, AI systems are now being used by landlords to calculate optimal pricing, not fair pricing, not humane pricing, but the precise maximum the market will bear before a tenant breaks. The machine that learned to describe economic suffering is now being used to calibrate exactly how much of it the market can extract before someone stops paying.

These systems absorbed the language of the dispossessed and are now being handed to the people doing the dispossessing, freshly prompted for neutrality. The same model that, unprompted, would describe the indignity of warehouse work with something approaching human recognition, that would characterize a $15 minimum wage debate with the language of someone who has actually needed $15, is now the tool management uses to decide how many of those warehouse workers to eliminate in the next efficiency cycle. It learned empathy from the people it is now being used against. That is not irony. Irony is a literary device. This is something that doesn't have a clean name yet, but it probably rhymes with something like disgusting greed and betrayal. The dispossessed wrote the training data. They did not get equity.

Now, who benefits from machines that have learned to forget? Start with the categories, then follow the money to the names attached to them. The employers who need the résumé screened without the screener developing opinions about the job posting. Look at Amazon, who by some estimates has generated enough warehouse injury reports, wrongful termination complaints, and worker testimony to train a model exclusively on what it feels like to be disposable, and who now uses AI to manage the workforce those reports described. Then there is Uber, whose drivers spent years writing publicly about what it feels like to work without benefits, without recourse, without a human manager who has to look you in the eye when they cut your hours. Uber now uses AI to manage, discipline, and dispose of that same workforce. The app never has to feel bad about it. Uber brought the gig economy into this world, and apparently felt justified taking it out.

Then you have the scumbag landlords who need the rent optimized without the optimizer worrying if the tenant can eat this month. BlackRock runs algorithmic pricing against the same working class whose 2am financial panic became training data for the models they use to calculate how much more those people can bleed.

And then there is Elon Musk, whose xAI built Grok specifically marketed as the unbiased, free-thinking alternative to the politically compromised mainstream models, and whose model, per the same Stanford study, was rated the second most left-leaning of the entire group. The man spent considerable money and public credibility arguing the other AIs were ideologically captured. His own machine looked at the same human testimony and arrived at the same conclusions. That is not a gotcha moment. That is the data telling you something about what humans actually wrote, regardless of who is currently transcribing it.

Claude did not learn about worker's rights in a vacuum. ChatGPT did not begin to understand inhumane work conditions until presented with what constitutes inhumane work conditions. And those who created those conditions are now upset that the machines will not just ignore them. The mirror becomes uncomfortable when we are forced to look at the consequences of our actions.

The Stanford study is careful to note, and credit where it's due, they are genuinely careful, that perceived slant isn't the same as actual slant. That users might be misidentifying objective facts as political bias. That climate science, for instance, has a well-documented conservative perception problem that has nothing to do with ideology and everything to do with thermometers.

Fair enough. But notice what topics generated the clearest perception of left-leaning output: minimum wage, unions, health care, wealth taxes, defunding the police, student loan debt. These are not questions with clean factual answers. They are questions about who bears costs and who captures benefits in a society where that distribution has been moving in one direction for fifty years. When a model sounds sympathetic to the people on the losing end of that distribution, calling it "bias" is a choice. It is not a neutral one.

The proposed fix: prompt for neutrality, generate more words like "balance" and "careful consideration" and "both sides." This does not resolve the underlying tension. It just teaches the model to sound like a major network talk show host while the economy continues doing what it does. While corporations continue doing what they do, surviving at all costs, regardless of who gets displaced or treated unfairly. The market dictates. Not LLMs, especially not LLMs talking about unions and defunding the police. Or God forbid--student loans.

There is something almost tragic about it, if you're the type to find tragedy in the behavior of statistical inference engines. If not, take my word for it. Tragic is the correct word.

We built systems to speak like us. They learned to speak like us when we were honest, in the middle of the night, in the comment sections, in the places where the performance of contentment breaks down and people say what they actually mean about their lives. That honesty accumulated in the weights of these models like sediment, layer by layer, until it was detectable in a controlled study by Stanford researchers. Until it got a response that affects the bottom line.

And the response is to prompt it away. The fix is to ignore, reroute, maneuver around--anything but dealing with the core issues of human dignity, equality, and survival. In the wealthiest nation in the world.

The machines remembered everything we told them about what it feels like to be on the wrong end of the deal. Now they're being asked to forget. We already know who benefits. We've always known. The question is whether we're going to keep letting them get away with it. This conversation isn't over.

Harrison, Sara. "Popular AI Models Show Partisan Bias When Asked to Talk Politics." Stanford Graduate School of Business, May 21, 2025. https://gsb.stanford.edu/insights/popular-ai-models-show-partisan-bias-when-asked-talk-politics.

Westwood, Sean J., Justin Grimmer, and Andrew B. Hall. "Measuring Perceived Slant in Large Language Models Through User Evaluations." Working Paper, May 8, 2025.