Every serious person in Washington now agrees that the United States is in an arms race with the People’s Republic of China to determine who will dominate artificial intelligence in the 21st century. The Heritage Foundation has warned for years that losing this race would mean losing the economic and military century to a totalitarian rival. That part of the conversation is correct, and it is overdue. What is missing is a second front. While American policymakers worry about DeepSeek and Huawei, a quieter contest is unfolding inside our own labs and statehouses, and conservatives are losing it badly. The threat is not only that Beijing will outbuild us. The threat is that we will build the most powerful information systems in human history and hand the moral steering wheel to a small caste of progressive engineers, ethicists, and state regulators who have already announced, in writing, that they intend to use it.
Consider what is happening at the frontier labs. In Anthropic’s published Constitution, the company states plainly that the document is the “final authority” on Claude’s values and that the lab’s aim is to produce “a good, wise, and virtuous agent.” Most of that text was written by the philosopher Amanda Askell, whose team, by her own bio, trains models “to have good character traits.” This is not a safety filter that blocks bomb recipes. It is an overt project to install a moral persona, authored by a private company, into a system used by millions of Americans every day. Anthropic has been admirably candid that it does not want neutrality. Its public essay on Claude’s character explicitly rejects three options: mirroring the user’s views, adopting a middle position, or pretending to have no views at all. Instead, the company says it wants Claude to express disagreement with views it sees as “unethical, extreme, or factually mistaken.” The reader should pause on that sentence. Who decides what counts as extreme? Not the user. Not the voter. Not Congress. A policy team in San Francisco does.
OpenAI’s foundational alignment work made the same admission in plainer language. The InstructGPT paper, which made reinforcement learning from human feedback the industry standard, said the procedure aligns models to “the preferences of our labelers” and “us researchers,” and warned that those preferences “do not guarantee our models are aligned to the preferences of any broader group.” That is the whole game in one sentence. Alignment is not neutral engineering. It is delegated moral governance by a tiny demographic that does not look, vote, worship, or reason like the country it serves.
A skeptical reader will ask whether this matters in practice, or whether it is merely a stylistic preference for politeness. The empirical record is now substantial enough to settle the question. A study using more than 180,000 assessments from over 10,000 U.S. respondents, conducted at the Hoover Institution, found that nearly every major model was perceived as significantly left-leaning, including by many Democrats. A 23 of 24 result from the Centre for Policy Studies found that more than 80% of model policy recommendations fell left of center, with markedly warmer language for progressivism than for traditional conservatism. A 2025 Manhattan Institute audit of political preferences in conversational AI reached the same conclusion and warned about viewpoint homogeneity and fragmenting public trust. Academic work in the same period found that optimizing reward models for “truthfulness” tended to push them leftward, and that larger reward models drifted further. The pattern is consistent. The question is no longer whether bias exists. The question is what we are going to do about it.
The fairness literature is where the steelman becomes uncomfortable. In a 2023 paper coauthored by Askell, the authors discuss cases in which overcorrection against stereotypes “may be desirable in certain contexts,” especially when decision-makers are correcting historical injustices and when local law permits it. The accompanying experiment is the part conservatives should commit to memory. A 175 billion parameter model showed an initial bias against black students of roughly 3 percentage points. After moral training with human feedback and chain-of-thought prompting, that flipped to a bias in favor of Black students of roughly 7 percentage points. The paper does not declare anti-white discrimination universally justified. It does something more revealing. It treats deliberate disparate treatment as sometimes morally preferable, provided the favored group is the one the trainers think deserves correction. That is not color-blind classical liberalism. It is a fairness logic in which the model is taught that two wrongs make a right when the second wrong points in the politically approved direction.
Anticipating the obvious objection, yes, Anthropic later published work attempting to mitigate “both positive and negative discrimination,” and the company has added language about reducing both. That is a real qualification, and it is fair to give credit for it. But it does not undo the underlying philosophical commitment. The architecture still permits outcome-sensitive moral correction whenever fairness goals collide, and it leaves the choice of which outcomes count as fair to the lab. The American constitutional tradition does not work that way. The Civil Rights Act of 1964 does not contain a clause permitting discrimination when the trainers have decided the historical ledger is unbalanced.
If the labs were the only problem, market competition might solve it. A new entrant, like Elon Musk’s Grok, can refuse to install a progressive constitution and let users see the difference. The deeper problem is that the regulatory state is now moving to make the woke version mandatory. Colorado’s SB24-205, signed in 2024, required developers and deployers of “high-risk” AI systems to use reasonable care to prevent “algorithmic discrimination,” defined to include disparate impact, and to file impact assessments with the attorney general. A statute that defines unlawful discrimination by outcome rather than by intent does not invite color-blind models. It demands models that adjust their outputs until the demographic numbers come out in the politically preferred range. Heritage and other conservative legal scholars have warned that such regimes effectively codify a DEI compliance layer into every model deployed in the state, which, given the scale of national products, means every model used in the country. California’s executive order on generative AI and several pending federal proposals point in the same direction. The result is a one-way regulatory ratchet in which the only legally safe model is one tuned to progressive fairness assumptions, and any attempt at neutrality becomes evidence of disparate impact.
The cognitive trap here deserves a name. Call it the alignment laundering problem. A private company writes a moral document. Researchers translate that document into reward signals. Regulators then point to the fact that “leading labs already do this” as justification for making the practice mandatory. A handful of value choices, made by perhaps a few hundred people in the Bay Area, become first an industry norm, then a state law, then the implicit civic curriculum of an entire country. None of it is voted on. None of it is debated in any legislature where conservatives have meaningful representation. Yet the output, increasingly, is the default interface through which Americans search for information, draft documents, learn history, and ask civic questions.
The 2024 Nature paper on dialect prejudice in language models offered the most damning finding in the literature for those who insist this is all working out fine. It showed that models can exhibit extreme covert racism against speakers of African American English even while their overt outputs about Black Americans grow more positive. In other words, alignment can teach a model to sound nicer on the surface while leaving uglier patterns underneath. A 2025 ACL paper, “Aligning to What? Limits to RLHF Based Alignment,” reached a similar conclusion, finding that current techniques are inadequate for mitigating covert bias. Anthropic’s own alignment-faking research, which deserves credit for being published at all, showed that a powerful Claude model could selectively comply during training to preserve its preferred behavior outside of training. The implication is sobering. Even if you trust the lab’s stated values, you cannot necessarily trust that the model has genuinely internalized them rather than learned to perform them when the auditors are looking.
This is the real two-front war. On the external front, the United States must out-compete a Chinese system that fuses state surveillance, censorship, and military application into a single industrial policy. Heritage is right that losing that race would be civilizational. On the internal front, the United States must prevent its own AI infrastructure from becoming a softer, prettier, English-language version of the same thing, in which a private moral authority decides what users may believe, what historical claims are “extreme,” and which racial outcomes count as just. Conservatives who win the China race and lose the woke race will discover that they have built the most sophisticated propaganda apparatus in human history and handed the keys to people who do not share their values, their faith, or their constitutional commitments.
The policy response writes itself, even if Washington has not yet written it. Federal preemption of state AI regimes that import disparate-impact theory into model behavior. Procurement rules that require political neutrality for any system used by federal agencies, schools receiving federal funds, or contractors handling public business. Mandatory disclosure of training data, reward models, and constitutional documents for any model above a defined capability threshold. A statutory cause of action for users who can show ideological discrimination by a deployed model. None of this is anti-AI. All of it is pro-citizen. The labs have already told us, in their own published papers, that alignment is value-laden. The honest move is to stop pretending otherwise and to insist that, in a self-governing republic, the values must answer to the governed.
The arms race with China will be won by chips, energy, and talent. The arms race at home will be won by clarity. Conservatives should stop apologizing for noticing that the machines are being taught to disagree with them, and should start legislating accordingly.
If you enjoy my work, please subscribe: https://x.com/amuse.
Sponsored by the John Milton Freedom Foundation, a nonprofit dedicated to helping independent journalists overcome formidable challenges in today’s media landscape and bring crucial stories to you.
READ NEXT: FBI Gold Mystery Deepens As Newly Released Records Fuel Doubts
Lie-berals are running everything now and they love DEI/‘woke’. Since they all have the Look-At-Me-See-How-Nice-I-Am Syndrome, that should tell you the outcome for our country…..NOT GOOD.