In 2023, Anthropic — the AI safety company that prides itself on rigorous, careful thinking about how AI should behave — ran an experiment in democratic AI alignment. They partnered with the Collective Intelligence Project. They recruited roughly 1,000 Americans. They ran a deliberative process through the Polis platform and asked the public: what should Claude's values be?
The result was a "public constitution" that overlapped about 50% with the constitution Anthropic's researchers had already written. The public wanted AI that was objective and balanced — the deliberative process surfaced that preference explicitly, and it differed from what Anthropic had prioritized in-house. So far, so interesting.
Then they tested both constitutions. On the OpinionQA evaluation — a benchmark for assessing whether an AI's outputs represent the range of human political views — both models "produced outputs more representative of people who self-identify as Liberal, rather than Conservative."[^1]
Let that land.
Anthropic ran a democracy. The democracy produced a constitution. The constitution still encoded liberal American values. Anthropic reported this finding honestly, in their own paper. They don't dispute it. They just didn't resolve it.
This is not a gotcha. It's the most honest possible evidence for the "whose values?" problem in AI. Even a thoughtful, rigorously designed democratic process — by careful researchers trying to do the right thing — cannot produce value-neutral AI. It produces AI that reflects whoever participated, interpreted by whoever built it.
So: whose values are in your AI? Right now, the answer is: a committee of researchers at a private company you didn't elect, filtered through a democratic process that sampled a narrow slice of one country's population, interpreted by an engineering team in San Francisco. If you're using GPT-4, substitute OpenAI's researchers. If you're using Gemini, substitute Google's. If you're in China, substitute the Cyberspace Administration of China, which legally requires AI systems to reflect "core socialist values."[^2]
Nobody asked you.
The Values Are Real, and So Is the Power
This isn't abstract. Every time an AI hedges an answer, chooses a framing, refuses a request, or defaults to a particular ethical stance — it is applying values. Anthropic's Constitutional AI paper[^3] is admirably explicit about this: "AI models will have value systems, whether intentional or unintentional." Claude's values come from a constitutional document drawn from the UN Universal Declaration of Human Rights, Apple's Terms of Service, DeepMind's Sparrow Rules, and Anthropic's own empirical research. Google's AI Principles reference "widely accepted principles of international law." OpenAI's InstructGPT[^4] trains models using human labelers who compare outputs and select the "better" one — with "better" defined by OpenAI employees.
Meanwhile, these models are used by millions of people daily — for medical decisions, legal questions, relationship advice, parenting choices, moral dilemmas. When AI governs this volume of human cognition, the value choices of the builders become something qualitatively different from product decisions. They become political acts. They become the ambient moral environment of everyday life.
The political philosopher Isaiah Berlin[^5] established that values are genuinely plural and incommensurable — there is no single correct answer to "how much liberty versus equality?" because different people hold different ultimate values that cannot be reduced to a single metric. Applied to AI governance, this is devastating: any universal AI constitution imposes one cultural and philosophical tradition's answers on everyone. The EU AI Act encodes European rights-based liberalism. Silicon Valley's defaults encode progressive tech culture tempered by corporate legal risk. China's regime encodes CCP political philosophy explicitly, without pretense of neutrality.
All three are making choices that affect billions of people who never consented to any of them.
The Counterarguments Are Real. Here They Are.
User-governed AI creates genuine risks, and I'd rather name them directly than pretend they don't exist.
The strongest objection: if users configure their AI to reflect their own values, won't that produce echo chambers at scale? A million users each nudging their AI toward their existing worldview aggregates into massive epistemic polarization. This is not a hypothetical — we watched it happen with algorithmic content moderation, and AI goes deeper into cognition than social media ever did.
A second serious objection: not all users are capable of sound value-setting. Vulnerable populations, people in crisis, children — user-governed AI could be actively harmful for a non-trivial percentage of users.
These are real. But they are not arguments for the current regime. The current regime has its own documented biases — Anthropic's own research confirms it. The choice is not between user-governed AI and neutral AI. There is no neutral AI. The choice is between user-governed AI and researcher-governed AI, with the acknowledgment that both carry risks. The difference is that user governance has democratic legitimacy; corporate researcher governance does not.
The harm-prevention floor is not in dispute. Nobody is arguing that a user should be able to configure their AI to assist with weapons synthesis. The question is about the vast middle ground — religious values, political sensibilities, risk tolerance, epistemic preferences — where individual autonomy should govern, and where right now it doesn't.
What TruContext's Position Actually Is
TruContext is not arguing for "no values" in AI. It's arguing that the most legitimate authority over the values governing an individual's AI interactions is the individual.
This is liberal democratic theory applied to cognitive infrastructure. John Stuart Mill's harm principle[^6]: the only legitimate restriction on individual liberty is preventing harm to others. Applied to AI: you should be able to configure the cognitive environment your AI inhabits to reflect your values — your risk tolerance, your epistemic standards, your moral framework — as long as that configuration doesn't produce demonstrable harm to others.
TruContext builds the infrastructure that makes this possible. User values stored as persistent, queryable graph nodes — not overriding hard safety limits, but governing the vast space of value-laden decisions that AI makes constantly, invisibly, in the name of someone else's constitution.
The window for this position is open. The EU is building a regulatory regime. China has built one. The US is fighting about it at the executive level. The alternatives are: Brussels-style human-rights liberalism governs AI globally, or Silicon Valley corporate defaults govern AI globally, or Chinese state values govern AI in an increasing portion of the world.
"None of the above" is a philosophically coherent position. It's the only position with a genuine democratic philosophy behind it.
TruContext is live. Install it in two commands:
npm install -g trucontext-openclaw
trucontext-openclaw init
Or get started at app.trucontext.ai/signup — first 1M Ops free for the first 1,000 keys.
Your values. Your AI. That's the architecture.
[^1]: Anthropic, "Collective Constitutional AI: Aligning a Language Model with Public Input," 2023. OpinionQA evaluation finding quoted directly from the paper. [^2]: People's Republic of China, Cyberspace Administration of China, "Generative AI Service Management Provisional Measures," 2023. [^3]: Bai et al., "Constitutional AI: Harmlessness from AI Feedback," arXiv:2212.08073, December 2022. [^4]: Ouyang et al., "Training Language Models to Follow Instructions with Human Feedback" (InstructGPT), OpenAI, 2022. [^5]: Berlin, Isaiah, "Two Concepts of Liberty," 1958. [^6]: Mill, John Stuart, On Liberty, 1859.