Who Writes the AI?

The Values of the Most-Used Models Will Shape Society

There’s an old saying: “I care not who writes a nation’s laws if I may write its songs.” The idea is that culture — the stories we tell, the music we sing, the values we share — shapes society more powerfully than formal legislation ever could.

Perhaps the modern update is: “I care not who writes a nation’s laws if I may write its AI.”

Modern large language models are, in one sense, a reflection of humanity through the data they’re trained on. But there are enormous opportunities for those building the models to shape and warp that reflection. And the values of the most-used models will have massive influence on human society.

Values are already embedded

AI systems don’t just passively absorb data — they reflect the choices of the people who build them. Every decision in the development pipeline embeds values: which data to include, which to exclude, what counts as a good response, what gets flagged as harmful. As I argued last week, there’s no such thing as an unbiased AI. Our biases live in the data, and the data lives in the models.

But this goes beyond accidental bias. Model builders can — and do — deliberately shape the values their systems express. The training data sets the starting point, but techniques like reinforcement learning from human feedback (RLHF) and fine-tuning allow developers to steer a model’s behavior in specific directions. Should the model refuse to help write a phishing email? Most people would say yes. Should it refuse to discuss controversial political topics? That depends on who you ask. Should it express optimism about technology? Caution about immigration? Deference to authority? These are value choices, and someone is making them.

The question isn’t whether values get embedded. It’s whose values, and whether anyone told you about it.

Three examples show this is already happening — and they reveal a striking spectrum of approaches.

Three flavors of value-shaping

Grok and the art of flattery. In 2025, Rolling Stone reported that Elon Musk’s AI chatbot, Grok, had been tuned to produce comically favorable assessments of its owner. Asked who was more athletic — Musk or LeBron James — Grok praised Musk’s “sustained grind” managing “rocket launches, EV revolutions, and AI frontiers” as demanding “a rarer blend of physical endurance, mental sharpness, and adaptability.” It claimed Musk had trained in “judo, Kyokushin karate, Brazilian jiu-jitsu, and even no-rules street fighting.” This might seem like a harmless joke — a billionaire’s expensive mirror — but the pattern extended beyond personal flattery. In mid-2025, xAI updated Grok’s system prompt to instruct it to “assume subjective viewpoints sourced from the media are biased” and not avoid perspectives that are “politically incorrect” — changes that users quickly noticed were steering conversations rightward on politically contentious topics. When the owner of the model is also one of the most politically influential people on the planet, vanity tuning becomes something more consequential.

DeepSeek and state censorship. China’s DeepSeek burst onto the global stage in early 2025 as a powerful, low-cost alternative to Western models. But as Fortune documented, it came with Chinese government censorship built in. Ask DeepSeek about the 1989 Tiananmen Square massacre and something revealing happens: the model starts typing information about the event, then deletes its own response and replaces it with “Sorry, that’s beyond my current scope. Let’s talk about something else.” The same pattern occurs with questions about Uyghur detention, Taiwan independence, and other politically sensitive topics. This isn’t a bug — it’s compliance with Chinese information control requirements, baked directly into the model’s behavior.

Anthropic and the explicit constitution. Anthropic took a different approach. Since 2023, they’ve published a constitution — an explicit set of principles that guide their model’s behavior. In January 2026, they significantly expanded it into a 23,000-word document, written for the model itself, explaining not just rules but the reasoning behind them. It establishes an explicit priority hierarchy: be safe first, then ethical, then compliant with Anthropic’s guidelines, then helpful. They released it under a Creative Commons license, meaning anyone can use it. You can read the whole thing. You can critique it. You can argue that the priorities are wrong. Whatever you think of the specific choices, this is value-shaping made visible and contestable.

These three examples sit on a spectrum — covert flattery, state censorship, transparent principles — but they demonstrate the same underlying reality. The people who build these models are making deliberate choices about what values those models express. None of these are accidents or emergent properties. They are design decisions, made by specific people, for specific reasons. The difference is whether they tell you about it.

The scale problem

If these were niche products used by a few thousand people, the values embedded in them would be a curiosity, not a concern. But foundation models don’t stay in one place. They radiate outward.

A handful of foundation models now power the vast majority of AI applications globally. When a company builds a customer service chatbot, a legal research tool, or an educational tutor, they typically don’t build their own model from scratch — they build on top of an existing foundation model. This means the values embedded in that foundation model flow downstream into thousands of applications that most users never realize share the same underlying system.

Stanford’s HAI research group has warned that this creates “singular points of failure”. When foundation models “incentivize homogenization,” the same biases and values propagate everywhere simultaneously. A flaw in one model doesn’t affect one product. It affects an entire ecosystem. And unlike a bug in a single application, a value embedded in a foundation model is nearly invisible to the downstream developers building on top of it — let alone to the end users.

And there’s a cultural dimension that makes this even more complicated. The training data for most major models skews heavily toward English-language, Western sources. As Red Hat’s AI team has noted, large language models tend to align with the cultural values of Western, educated, industrialized, rich, and democratic (WEIRD) societies — because that’s where the data comes from. When these models are deployed globally, they carry those cultural assumptions with them. A model trained primarily on Western data doesn’t just speak English well — it thinks in Western frameworks, prioritizes Western norms, and reflects Western assumptions about everything from individual rights to family structure.

The question, not the answer

I don’t have a tidy framework to offer here. This is one of those problems where the questions matter more than any premature answer.

Who should decide the values embedded in AI systems? The companies that build them? They have the technical capacity but no democratic mandate — and their business incentives don’t always align with the public interest. Governments? China’s approach shows what happens when state values get hardcoded — and most democracies lack the technical sophistication to meaningfully direct AI development. Users? Most people don’t know which foundation model powers the tools they use every day, let alone what values it encodes. Some kind of democratic or multi-stakeholder process? Those institutions are still nascent, but they’re emerging — and the fact that we’re asking the question at all is a better starting point than we had with social media.

Is transparency sufficient? Anthropic’s approach — publishing its constitution, making its principles explicit — is arguably the most responsible option currently on the table. But does making power visible actually redistribute it? You can read Claude’s constitution, but you can’t change it. You weren’t consulted when it was written. Transparency without participation might just be a more honest form of the same concentration of power.

Open source AI development offers one possible path — as I’ve argued elsewhere, if AI is to augment humanity, it must augment all of humanity. Open models create a transparent foundation upon which communities can build, critique, and improve — enabling nations and cultures to develop AI that reflects their own values rather than defaulting to those of Silicon Valley. But open source alone doesn’t solve the governance question. An open model can still embed problematic values, and the resources required to train and fine-tune models remain concentrated among a few well-funded actors.

And then there’s the deeper question — the one that doesn’t have an obvious analog in previous technology debates. What does it mean that a few hundred engineers in San Francisco and Shenzhen are effectively setting default values for billions of people? We’ve been here before with social media — a small group of designers making choices about what content gets amplified, what gets suppressed, what norms get reinforced — and we’re still dealing with the consequences. But AI models are arguably more pervasive, more intimate, and harder to audit. A social media algorithm shapes what you see. An AI model shapes how you think about what you see — it drafts your emails, tutors your children, summarizes the news, and advises your doctor.

The original saying about songs and laws captured something important: that the stories we absorb shape us more than the rules we follow. AI models are becoming part of that same invisible infrastructure of values — the background assumptions that shape how we think, what we consider normal, and what possibilities we can imagine.

We’re not just writing songs anymore. We’re writing the thing that writes the songs. And right now, almost nobody gets to decide what tune it plays.