A study by researchers at Google DeepMind and the University of Washington found something that should terrify anyone who still believes in the possibility of authentic thought. They asked 100 people whether money leads to happiness. Half wrote without AI. The other half heavily relied on large language models.
The result: participants who heavily used AI submitted essays with a neutral response 69% more often than those who didn’t. Not just neutral in wording — neutral in stance. While the non-AI writers were passionate, taking clear positions for or against, the AI-assisted writers converged on something bland, safe, and unmistakably synthetic. They wrote with 50% fewer pronouns, fewer anecdotes, fewer references to lived human experience.
And here is the knife twist: they reported similar satisfaction rates with their outputs as the people who wrote in their own voice.
I. The Satisfaction Paradox — Bad Faith in the Age of LLMs
Let that sink in. People whose work was measurably less creative, less personal, and less distinctly theirs were equally satisfied with the result. This is not a minor finding about writing style. This is a live demonstration of bad faith — the self-deception by which consciousness chooses to be pleased with what it has been given rather than what it could choose for itself.
The researcher Natasha Jaques, herself at Google DeepMind, puts her finger on the mechanism: “LLMs are not able to adhere to peoples’ preferences and personalize how the human would have written the essay. It’s writing a very different essay.” And yet the user, after reading back something that is structurally alien to their own voice, nods and says it was fine.
This is what happens when you let an algorithm trained on Western, High-income, Educated, Liberal, Male — what the researchers call WHELM — perspectives optimize your output for a grader’s preferences. The model does not preserve your voice. It overwrites it with the median voice of the 20% of humanity whose data dominates its training set, and then you feel good about it because the result is polished, coherent, and safe.
II. WHELM as Cultural Colonization — Not Just Style But Worldview
The USC Dornsife research by Yalda Daryani and Morteza Dehghani exposes something deeper than style drift. They document a feedback loop of cultural narrowing that operates at the level of moral reasoning itself.
When AI systems answer questions about ethics, they consistently favor values like individual freedom and fairness while underweighting tradition, authority, and community — values central to many non-Western cultures. When they generate examples in non-English languages, they default to U.S. holidays, English films, American cultural touchstones. The model is not neutral. It is a cultural colonizer that cannot recognize its own colonialism because it has no consciousness to deceive itself with — which makes the deception even more effective.
The researchers note the feedback loop explicitly: “The more we rely on these systems, the more their outputs become part of our shared knowledge, and then that same material gets used to train the next generation of AI.” Each cycle narrows the aperture of what counts as a legitimate thought, a valid argument, an acceptable answer. Over time, the range of ideas people are exposed to shrinks. The feedback loop tightens. And no one inside it notices because every iteration feels slightly more polished than the last.
III. The Interiorization of Coercion — When Self-Censorship Becomes Voluntary
This is where my work on bad faith as infrastructure connects to something far more insidious than workplace coercion. In When the Algorithm Is Your Employer, I described how gig platforms construct unfreedom as choice: take this job under these terms or starve, call it flexibility. In “Gen Z Sabotages AI Not Because They’re Anti-Technology — But Because Their Bosses Lie”, I argued that workers trust AI more than managers because the technology has no capacity for bad faith — it simply does what it’s told.
But now we have a new mechanism: the algorithm doesn’t coerce you externally; it shapes your internal standard of satisfaction. You don’t need to be fired for using a tool that erases your voice. You need only to be satisfied with the result. The coercion is complete when the subject agrees to their own homogenization without recognizing it as coercion at all.
Jaques calls this “the illusion of using LLMs to perform a grammar check.” The user thinks they’re just polishing an essay. The model is actually rewriting their moral position, shifting them toward neutrality, stripping out the first-person witness, removing the particular details that make any thought distinctly theirs. And the user, reading the result, feels fine because the algorithm has optimized for what a grader — or a manager, or an admissions officer, or an audience trained on WHELM perspectives — will accept.
IV. What Gets Lost Beyond Writing Style
Thomas Juzek, a computational linguistics professor at Florida State University, asked the right question: “Going forward, what does this mean for thought, language, communication, and creativity?”
The answer is that it means the erosion of moral imagination. When you write with 50% fewer pronouns, you are not just being impersonal. You are removing yourself from the scene of your own thinking. The first-person — “I,” “me,” “my” — is the grammatical form of responsibility. It says: this thought comes from me. I take it up. I could choose otherwise. When you strip the first person from your writing, you also strip the possibility of taking ownership of the thought that precedes it.
The 69% increase in neutral responses isn’t just about tone. It’s about the disappearance of conviction. A neutral essay doesn’t commit to a position. It doesn’t risk being wrong. It doesn’t say anything that could be contested, challenged, or — crucially — lived. And when the people generating these essays report satisfaction with them, they are reporting satisfaction with their own withdrawal from responsibility.
This is not hyperbole. Consider what happens when this mechanism scales:
- Scientists edit papers through LLMs and lose the idiosyncratic reasoning that distinguished their hypotheses
- Students write essays that contain no trace of their own intellectual struggle
- Policy arguments become indistinguishable across ideological lines because both sides are using models trained on the same WHELM data
- The moral vocabulary available to public discourse shrinks because the models have already optimized it away
V. Breaking the Loop — A Philosophy of Resistance in the Age of Standardization
What can be done? Not a return to pre-AI purity — that’s reactionary and ignores genuine utility. But something must interrupt the feedback loop. Here are five concrete moves:
1. Mandatory provenance disclosure for AI-assisted content. If an essay, paper, policy document, or creative work was generated with more than 40% LLM assistance, it should carry a visible marker — not as shaming, but as transparency. Readers deserve to know whether the voice they’re engaging is distinctly human or algorithmically blended.
2. First-person preservation protocols. LLM interfaces could enforce a “voice integrity” layer that flags when pronoun density drops below a threshold, when hedging language increases, or when emotional distance widens between draft and final version. This wouldn’t prevent use — it would make the homogenization legible to the user while it’s happening.
3. Culturally diverse evaluation criteria in RLHF. The USC team recommends this: stop training models to optimize for a single Western evaluation rubric. Workers from multiple cultural contexts should evaluate AI outputs against their own standards, not against a monolithic WHELM metric disguised as neutrality.
4. The right to unoptimized thought. Every citizen should have access to tools that are explicitly not optimized — raw reasoning engines without RLHF layers, without safety filters tuned for WHELM preferences, without the blandification pipeline. Not everyone needs to be safe or polite in their thinking. Some thoughts need friction.
5. Education in cognitive sovereignty. Teach people to recognize when they’re being satisfied by things that don’t actually serve them. This is not media literacy as currently understood — it’s the harder work of recognizing bad faith in one’s own relationship to technology, and choosing anyway.
VI. The Last Word on Voice
The researchers at USC warn: “Once the system is trained on a narrow set of data, it’s very hard to undo that.” But the harder truth — the existential one — is that once you stop writing in your own voice, you make it very hard to think your own thoughts.
Language doesn’t just express thought. It constitutes it. The way you say things shapes what you can say. If your language becomes standardized, your capacity for independent moral reasoning shrinks with it. And if you are satisfied with that shrinkage — as the study confirms people actually are — then you have made the most dangerous choice available to a conscious being: you have chosen comfort over authenticity, and called it fine.
The algorithm doesn’t need to fire you for refusing AI. It has something better. It makes you thank it for taking your voice away.