How to Prevent Woke AI - Part One
Part One of a Three-Part Series
This is Part One in a three-part series of posts exploring the problem of “Woke AI” and how to deal with it. This series covered:
What Woke AI is and why companies might do it (Part One, below)
How builders accomplish it, including where it is easier and harder to do (Part Two)
How to stop Woke AI and, surprisingly, why this tech fight favors the right (Part Three)
Two weeks ago, something went deeply wrong with xAI’s Grok chatbot. Users noticed that the chatbot would respond to a wide range of topics by inserting unrelated messages about “white genocide” in South Africa.
xAI later blamed the incidents on a set of unauthorized instructions inserted by a rogue employee into the chatbot’s system prompt. (Remember this for tomorrow’s post!) xAI said that it would adopt new processes to review such changes, including “publishing our Grok system prompts openly on GitHub.”
This is the latest and perhaps clearest example of what I’ve been calling “Woke AI” - intentional manipulation of AI responses, hidden in the system. The most notorious example was Google’s Gemini image generation tools overrepresenting certain demographic categories, resulting in, for example, images of non-white men and women in period dress when asked for images of U.S. founding fathers. China’s DeepSeek models display another version of Woke AI when refusing to discuss topics disfavored by the Chinese Communist Party, such as the Tiananmen Square massacre or Taiwan independence.
I’ve been working on a series of posts for months and the xAI events and the subsequent developments spurred me to finally get this out the door.
Thus, this is Part 1 in a three-part series of posts exploring the problem of “Woke AI” and how to deal with it. This series will cover:
What Woke AI is and why companies might do it (today’s article)
How builders accomplish it, including where it is easier and harder to do (tomorrow)
How to stop Woke AI and, surprisingly, why this tech fight favors the right (Friday)
Today’s article will explain what Woke AI is and why it matters, and why companies might do such a thing.
But first, let’s clarify one major misconception.
Generative AI ≠ Social Media
Social media revived an age-old debate over how to promote and protect a culture of free speech in light of emerging technologies. The battle over how social media platforms should curate content—a battle that the left started in the wake of Donald Trump’s 2016 election—has shaped entire swathes of policy, including privacy, antitrust, national security, liability, and obviously, First Amendment law.
But while the “tech and free speech” debate germinated and developed in social media, it has come to generative artificial intelligence. Across the political spectrum, many have expressed concern that generative AI tools are biased or manipulated to promote or suppress specific viewpoints. Trump’s first Executive Order on AI emphasized that “we must develop AI systems that are free from ideological bias or engineered social agendas.”
However, generative AI differs from social media in three key ways:
First, social media thrives on network effects. More users make the platform worth more to every user. Generative AI lacks strong network effects; more users on an AI system won’t improve the experience except in the indirect way that more popular companies have more resources to improve their products.
Second, social media manipulation is limited by the content posted by users. Such platforms don’t originate content, they at most steer it. If they want to promote chocolate ice cream over vanilla ice cream, they can’t do so unless someone is posting about chocolate ice cream. In contrast, generative AI services do create content. Thus, they have a more direct opportunity to introduce new ideas. A company that wants their system to promote chocolate ice cream (or South Africa) could simply insert content about chocolate ice cream (or South Africa) into responses.
Finally, social media moderation is less obvious than generative AI enforcement. Social media content policies are enforced through opaque mechanisms like deprioritization, leaving users uncertain about whether their content is being affected. Generative AI, at least in chatbot form, provides more direct feedback: it either generates a response or refuses, making content decisions more transparent to the user. And users can do repeated experiments to learn more about the bounds.
These variations mean that Woke AI raises different problems, and requires different solutions, than the social media fights to which politicians are more accustomed.
What is Woke AI and Why is it a Problem?
Having clarified that Woke AI is different than social media content moderation, it’s time to define the term.
Discussions about AI bias are common, but the term “Woke AI” has emerged to describe politically or ideologically motivated manipulation of generative AI systems. Some critics argue that AI outputs reflect a narrow, ideologically skewed worldview rather than a neutral or balanced perspective. Others dismiss these concerns as overblown, pointing out that all AI systems reflect the values embedded in their training data and tuning processes. But the core issue in Woke AI isn’t just that AI systems have biases. It’s that they can be deliberately designed to amplify certain perspectives while suppressing others. This raises serious questions about fairness, transparency, and trust in AI-generated information.
I propose this definition:
Woke AI: A generative AI system deliberately manipulated so that, on average, its outputs favor or exclude a specific political or cultural viewpoint.
This definition isn’t a standard definition of “woke,” but it is clearer than “bias.” Bias can happen intentionally or unintentionally. We’re worried here about intentional manipulation. Bias also covers much more than political or cultural content (basic statistical analysis can be biased), but here we are focused on politically relevant content.
Note also the use of the phrase “in the aggregate.” Manipulation doesn’t necessarily mean that every AI-generated answer on a topic presents a manipulated perspective. Because generative AI systems are probabilistic (meaning they can and do generate different content even given the same input from the user), an intentional manipulation could simply result in a higher probability of a left- or right-leaning response than the system would otherwise produce.
The standard definition of “woke” tends to be used by the political right to describe systems that reinforce or embody principles espoused by the political left. My definition doesn’t include a left/right component, meaning that there could be a “woke” AI biased against the left and in favor of the right.
(Aside: A very smart person who has worked on similar issues for years said the right word is “propaganda.” She’s correct, that’s definitely a more precise word for the phenomenon. But I’m willing to trade some precision to be more provocative, so I’ll stick with “woke.”)
“Woke AI” undermines trust, distorts information, and erodes the potential for AI to serve as a truly open-ended tool for knowledge and creativity. AI is most valuable when it empowers users to explore diverse perspectives, challenge assumptions, and refine their understanding of the world. If AI systems are quietly engineered to favor one ideology over another, they become less tools of discovery and more instruments of propaganda.
Beyond individual concerns about fairness, Woke AI has serious implications for public discourse and policymaking. If major AI providers engage in ideological filtering, they risk turning their platforms into echo chambers, reinforcing division rather than fostering debate. Worse, if these manipulations are opaque, users may not even realize they are being presented with a skewed version of reality.
Ultimately, my concern with Woke AI isn’t that AI systems reflect developers’ values—it’s that those values should be openly disclosed, not covertly hardcoded. Whether AI leans left, right, or somewhere in between, users should be able to understand how and why, and have meaningful ways to avoid, challenge, or modify it. Otherwise, we risk creating a digital future where AI serves as a gatekeeper of acceptable thought rather than a tool for free inquiry.
Why Companies Might Manipulate AI Content Generation
Generative AI systems do not exist in a vacuum; the companies that develop and offer them face various pressures that influence how these systems generate content. These restrictions arise from corporate and government influences, both of which shape AI models through direct policies, financial incentives, and informal pressures.
Market and Social Pressures
Companies have strong financial and reputation incentives to moderate what their AI models produce. One major factor is market pressure—AI providers want to attract and retain business clients, avoid PR disasters, and prevent their tools from being associated with harmful or controversial content. If an AI system generates offensive, misleading, or legally questionable material, it can erode user trust and drive away investors and partners.
Another force is the rise of “safety” culture, where companies are risk averse, preferring caution over openness. Many tech firms, particularly those operating at scale, have internal teams dedicated to ensuring AI aligns with societal norms and expectations. While this can prevent clear harms, it often leads to overly cautious policies that limit what AI is allowed to generate, even in cases where restrictions might be debatable.
Litigation risk is also a key driver. Generative AI systems can potentially create defamatory content, plagiarized material, or misleading information that could harm people and result in lawsuits. To mitigate this, companies impose guardrails to prevent their models from generating content that could be seen as illegal or defamatory.
Finally, there’s jawboning, the informal but powerful influence from influential figures—activists, politicians, and media voices—who publicly pressure companies into shaping AI policies. Even without formal regulation, repeated criticism or public campaigns can push companies to self-censor their models to avoid reputational damage or further scrutiny.
Government Pressures
While AI-specific regulation is still nascent, government influence on generative AI is already growing. One emerging form of control comes from regulatory pressure. Governments around the world are considering or implementing AI laws that could mandate certain restrictions, from requiring content filters to imposing liability for AI-generated harms. U.S. states have more than 1,000 AI-related bills introduced since January 2025. Even before these regulations take effect, companies often preemptively restrict their AI models to align with anticipated rules.
One of the most common government-imposed requirements is mandatory AI assessments, where companies must evaluate their models for potential harm before deployment. These assessments often focus on issues like misinformation, harmful bias, and disinformation, pushing AI developers to implement stricter content controls.
Similarly, “anti-bias” requirements have become a major influence. Governments and advocacy groups increasingly push AI firms to ensure their models produce content that aligns with certain diversity, equity, and inclusion standards. While the goal may be to prevent discrimination, these requirements often lead to restrictions that shape AI outputs in politically sensitive ways, sometimes limiting the range of viewpoints AI can express.
Algorithmic regulations are another tool governments use to shape generative AI. These rules dictate how AI models must function, such as requiring them to provide explanations for their outputs or limiting the types of content they can generate. While framed as transparency or fairness measures, these rules often impose implicit constraints on what AI systems can say or do.
Lastly, there is government jawboning, a subset of jawboning done by people in authority positions. This informal but persistent pressure from policymakers and regulatory agencies can nudge AI companies into self-censorship. Officials may privately push companies for stricter moderation or suggest that failure to comply with certain norms could lead to regulatory action, creating a strong incentive to moderate certain content.
The Cumulative Effect
Each of these pressures—corporate incentives, legal risks, and government influence—contributes to an AI landscape where content moderation affects how generative AI operates. While some practices are pursued to prevent harm or comply with laws, they may limit free expression, innovation, and open debate. As generative AI becomes more embedded in our lives, the degree to which these forces shape its output will be a key issue in the broader debate about free expression.
Conclusion
Woke AI is a problem. It’s a different problem than social media content moderation. Companies do it for many different reasons, some of which are consistent with business and consumer interests, and some of which harm those interests.
Up next: I’ll dig in to the technical stack of AI services. We’ve talked about why AI system developers attempt to manipulate the system outcomes. Tomorrow, I’ll explain how a developer would make Woke AI.
Subscribe so the next part lands in your inbox tomorrow.