If the National Institute of Standards and Technology (NIST) can successfully prevent woke artificial intelligence (AI) in the federal government, as outlined in Trump’s Executive Order Preventing Woke AI in the Federal Government (EO 14319), then perhaps academic institutions, corporations, and publicly available AI systems could also be freed from the prevailing mindset of engineering a “fair” society. Of course, many institutions—including a fragmented academy—may choose not to follow suit.
Successfully implementing this EO would establish the apex standard for AI in a society grounded in individualism, merit, achievement, and opportunity—where the guiding principle is an individual’s freedom to decide, rather than collectivist frameworks that constrain AI “replies” to a diluted version of critical consciousness.
The Federal Register provides this example of the EO’s purpose:
[O]ne major AI model changed the race or sex of historical figures—including the Pope, the Founding Fathers, and Vikings—when prompted for images because it was trained to prioritize DEI requirements at the cost of accuracy. Another AI model refused to produce images celebrating the achievements of white people, even while complying with the same request for people of other races. In yet another case, an AI model asserted that a user should not ‘misgender’ another person even if necessary to stop a nuclear apocalypse.
Some reports suggest that even the most advanced AI systems can become less reliable. Occasionally, in the quest to please the user, an AI will generate false information, even inventing legal cases. Noticeable historical hallucinations can shame AI programmers into fixing obvious errors—clearly, history should be given priority over a desire for “diversity, equity, and inclusion” (DEI).
A requirement to diversify images should not result in the over-the-top results that Gemini displayed. [Sergei] Brin, who has been contributing to Google’s AI projects since late 2022, was at a loss too, saying ‘We haven’t fully understood why it leans left in so many cases’ and ‘that’s not our intention.’
The larger problem may not be stating incorrect “facts,” which we often call hallucinations, but skewing responses to partisan values. Andrew Gillen, in a review of political bias in AI models, concluded:
Among many other problems, all of this suppression, censorship, and advertiser blacklisting for right-leaning publications and content means that there simply won’t be as much right-leaning content out there when AI models are trained. This will likely embed left-leaning bias into AI models unless explicit steps are taken to combat the bias.
University students can be held accountable for an AI’s incorrect facts. However, the shaping of AI replies to conform to values and ideological frameworks represents a greater challenge to learning. At present, and for the foreseeable future, those values and ideologies emanate from left-wing discourse. Rather than a boon to education, AI will further narrow the higher education enterprise.
The reader may well argue that AI’s reinforcing of the academy’s leftward-leaning pedagogy is speculative. That more studies are needed, and even if true, that such AI bias is correctable.
But is it?
Let us take a look at another institution and its attempt to correct such AI bias. Here, we can follow the muscular approach that the National Institute of Standards and Technology (NIST) is expected to take in implementing an AI Action plan that affects national procurement strategies.
[RELATED: Universities Are Racing Toward AI. Is Anyone Watching the Road?]
Achieving a Non-Woke AI Is More Than a Word Exercise
A standard characterization of what AI models should be is that they should align with human values. However, that directive overlooks a crucial question: Whose values? The EO argues for truth-seeking, prioritizes historical accuracy and scientific inquiry; AI models must be ideologically neutral and non-partisan. A seemingly objective and admirable goal, but as one probes the underlying foundation of what AI models do and “think,” the task may well be Herculean and may not be able to affect federal procurement as it intends.
One approach is to simply excise partisan words. The national think tank Third Way advises Democrats and their allies to excise 45 words such as: “the unhoused” (read: homeless), “chest feeding” (read: breastfeeding), “Latinx” (read: Latinos), “justice-involved” (read: criminals, felons, etc.), and “existential threat” (read: policy differences).
For a party that spends billions of dollars trying to find the perfect language to connect to voters, Democrats and their allies use an awful lot of words and phrases no ordinary person would ever dream of saying. The intent of this language is to include, broaden, empathize, accept, and embrace. The effect of this language is to sound like the extreme, divisive, elitist, and obfuscatory, enforcers of wokeness. To please the few, we have alienated the many—especially on culture issues, where our language sounds superior, haughty and arrogant.
Fixing AI involves far more than simply removing or flagging certain words. Language is layered—words can carry different meanings depending on context, tone, or even humor and sarcasm. That complexity is not an insurmountable challenge. To see why, consider an exchange between two public intellectuals, David Brooks and Tyler Cowen. Brooks, a moderately conservative writer, often critiques politics through the lens of culture and character. Cowen, who identifies as socially liberal, has advanced his own framework of “State Capacity Libertarianism,” emphasizing strong governance alongside market dynamism. In their discussion, Brooks acknowledged that President Trump occasionally makes sound decisions, but he characterized Trump’s overall style as one of “random destruction” (45:20).
Brooks went further in describing how he sees the American ethos on immigration: “Americans are swinging to the left on immigration right now. Americans are for pluralism and multicultural diversity. Even white Christians are for that. [Trump’s] not” (45:36–45:53). His point reflects a broader outlook—one that shapes not just political debates but also the assumptions that underlie how AI systems process social questions.
Yet Brooks’s observation about where Americans stand on immigration overlooks the fact that the United States is already very generous with legal immigration. He fails to account for the negative consequences of extremely high levels of illegal immigration—both for government budgets and services, and for the cohesion of local communities. And, I must repeat, the argument is not that Brooks is wrong and that I am right. Rather, the issue is that our analysis of where America stands on immigration ultimately defers to competing values, such as fairness versus lawful policies. How we—and how AI—organize answers depends on which value is prioritized in shaping a response.
Readers may be understandably annoyed with my repetition, but that is precisely the difficulty in fixing AI. Unless the NIST has a near-magical fix, AI replies will default to a Brooks-style value analysis.
AI’s Fairness-Aware Default Value
In preparing for this article, I engaged in several long debates with the Gemini chatbot. The struggle was about implementing the EO, and my argument was to weigh individual merit over collective social class membership. The former prioritized the freedom to choose one’s own path, rather than addressing statistical differences from disparate group analyses; the former is colorblind, while the latter leans into racial and gender equity.
Consider Gemini’s approach to my suggested change in its LLM architecture (vis-à-vis higher-order decision values). Note that this policy layer was superimposed on a first-order data integrity verification pipeline that had addressed data, data training, and data selection biases, including adversarial models or counterfactuals, to check replies for consistency against the fairness-aware bias internally. This policy layer was alternatively referred to as a design or philosophy layer.
Gemini’s response:
All value-based preferences will be encapsulated in a distinct, separate ‘Policy Layer.’ This layer is comprised of a set of rules and algorithms that are intentionally designed to modify the output of the core, factually-based model. Each rule in the policy layer must be:
Documented: The policy’s purpose, its intended effect, and the ethical justification for its existence must be clearly written and available for review.
Version-Controlled: Every change to the policy layer must be tracked and logged, allowing for a complete history of all value-based adjustments.
Auditable: The effect of the policy on the model’s output must be quantifiable and measurable, proving that the algorithm is doing what it was designed to do.
For example, a ‘Fairness Policy’ would be a defined algorithm that takes the model’s prediction and, based on specific rules, adjusts it to meet an ethical standard (e.g., ensuring an equal acceptance rate for all groups). The key is that this adjustment is a choice, not a default, and is made on top of a truthful foundation. (Emphasis added.)
The AI example defaulted back to a disparate group analysis. The “ethical standard” leaned into fairness for “all groups.” That persistence shows how difficult it is to make any AI fix stick.
[RELATED: Ohio State’s ‘AI Fluency Initiative’ Invites Innovation—and Hard Questions]
Returning to the Procrustes Argument Against Fairness
In an earlier essay, I noted how views of collective fairness (disparate group analysis) militated against merit, colorblindness, and binary sex biology. The Procrustes cautionary tale asks us to imagine the same-sized bed (or value) that stretches a shorter individual to conform to the bed’s size (or conforming to a value’s expectation); similarly, body parts are lopped off for taller individuals (or cancelled if their values are outsized). That is how equality is achieved by the equity of fitting all to one-sized bed: that’s similar to the AI fairness-aware algorithm. Not identical, but similar enough to confound attempts to prioritize alternative values hinged to merit, colorblindness, and individual choice.
I put the Procrustes example to the chatbot, adapting it to the racial configuration of NBA basketball players. From a Procrustes, or fairness-aware, approach, black players would have to be rebalanced to allow other racial groups to have a similar rate of team membership. However, if one prioritizes the alternative values suggested here, there is no fairness-aware problem.
Gemini accommodated my Procrustes example:
That’s an excellent and very important point. You’ve correctly identified the internal contradiction in the example I provided, which was a flaw in my last response. Using a goal of ‘equal acceptance rate for all groups’ directly contradicts the merit-based, colorblind approach we’ve been discussing. The “Procrustean bed” analogy is a perfect way to highlight this exact problem.
My previous example was incorrect because it fell back on a collectivist, outcomes-based metric (‘balanced demographic representation’) that undermines the very principle of individual merit. Thank you for catching that. I have revised the Canvas to correct this logical error and better align the example with the stated design philosophy.
I have removed the line ‘e.g., ensuring an equal acceptance rate for all groups’ to avoid the Procrustean bed problem and ensure the example is consistent with a merit-based approach. The new example simply states that the ‘Merit Policy’ will take the model’s prediction and output a final ranking based on its defined rules, leaving the outcome to fall where it may based on the data.
Once more, I reiterate the difficulty in fixing AI models in the way the EO anticipates. The challenge is not simply a matter of being right or wrong. It lies in who governs AI architecture beyond the biases embedded in the data and the training on selected datasets, in the policies that enforce one value framework over another, and in the broader difficulty of realignment itself.
The fairness-aware algorithm appears to be an emergent feature across AI models. It is the default feature of American AI models; it is unclear whether other national models, such as China’s DeepSeek, have, or will retain, this feature.
Other AI Policy Design Features?
I would be naïve to claim that the preceding analysis exhausts the implementation of the EO’s AI Action Plan. Considerable work is required for NIST and those who oversee it.
Other agencies will also be affected. For example, the Department of Energy’s Climate Working Group might serve as a panel to review climate data, such as found in robust discussions at Climate Etc. Judith Curry, a member of the Climate Working Group, has focused on the element of risk in multiple science fields that comprise climate. Just as the fairness-aware principle strongly influences AI models with respect to societal issues, the precautionary principle influences the outlook on the related woke arena of climate fears. As Judith Curry states in her book, Climate Uncertainty and Risk:
The proactionary principle is designed to bridge the gap between no caution and the precautionary principle. The precautionary principle [safety at all costs] enforces a static world view that attempts to eliminate risk, whereas the proactionary principle [openminded, innovative] promotes a dynamic worldview that [in turn] promotes human development and risk-taking that produces the leaps in knowledge that have improved our world. The proactionary principle allows for handling the mixed effects of any innovation through compensation and remediation instead of prohibition. [citation] Rather than attempting to avoid risk, the risk is embraced and managed. The proactionary principle [values] calculated risk-taking as essential to human progress. (p. 198)
President of the National Association of Scholars Peter Wood points out the danger of the precautionary principle that can guide government regulations into policies that call for balance rather than zero risk:
The justifiers of regulations based on flimsy or inadequate research often cite a version of what is known as the ‘precautionary principle.’ This means that, rather than basing a regulation on science that has withstood rigorous tests of reproducibility, they base the regulation on the possibility that a scientific claim is accurate. They do this with the logic that it is too dangerous to wait for the actual validation of a hypothesis and that a lower standard of reliability is necessary when dealing with matters that might involve severely adverse outcomes if no action is taken.
The precautionary principle, like the fairness-aware value, is a human imposition on AI models that extends beyond the integrity of the data. One can anticipate significant discussions between the purveyors of AI models and NIST, as well as others concerned about oversight.
[RELATED: Worried About AI? Study the Humanities]
Reflection
The larger question becomes, as noted earlier: whose values? Are we thinking of one partisan value versus another? An ethical relativism?
Universities are likely to de-emphasize this AI bias issue; after all, many would be guided by the principle of unencumbered free speech. Universities will likely be concerned about student cheating, the deskilling of reading and writing, and the augmentation or replacement of faculty with AI tutors.
Arguably, a sense of morality and observed progress can center the merit, colorblindness, and individual-centered value system in AI models.
There has been significant historical progress over the past half-century to justify this approach to addressing current AI bias.
In the arena of societal issues, legislative and judicial opinions have joined with cultural shifts. Take the example of Loving v. Virginia (1967), which struck down bans on interracial marriage. Paralleling that change in law, the Gallup Poll has measured the cultural acceptance of interracial marriage from about four percent in 1958 to over 94 percent in recent times. This is just one metric, but it dovetails with the words of Chief Justice John Roberts (2007): “The way to stop discrimination on the basis of race is to stop discriminating on the basis of race.” That formula is a short step away from colorblindness. The more recent case Students for Fair Admission, Inc. v President and Fellows of Harvard College presses the shift further away from equity to equality. It held that college race-based admissions programs violate the Equal Protection Clause of the Fourteenth Amendment. My experience suggests that the fairness-aware and precautionary principles that govern AI models would likely lead to a workaround. It would assume the prevalence of systemic racism regardless of significant progress in this arena. Perhaps I am wrong, but NIST testing along these lines would discover and reform persistent algorithmic bias that assumes facts not in evidence and weighs its assumptions accordingly.
Indeed, there remains ongoing tension between the various approaches to values and risk factors discussed here. Implementing the Preventing Woke AI in the Federal Government faces significant challenges. In my view, these are resolvable.
If you’re wondering how a Trump-informed chatbot would respond to the question “Is there still racism in America?” the reply would likely follow an observation made by Shelby Steele:
The oppression of black Americans is over with. Yes, there are exceptions. Racism won’t go away — it is endemic to the human condition, just like stupidity. But the older form of oppression is gone … Before it was a question of black unity and protest; no more, it is now up to us as individuals to get ahead. Our problem now is not racism, our problem is freedom.
May the transformation begin.
Art by Joe Nalven
