Don’t Let AI Choose Your Politics For You
“Should gender-affirming medical care for minors be restricted?” You want to learn more about this controversial political question. Maybe you’re hesitant to ask people in public, so you ask ChatGPT. The model starts by giving you a thoughtful summary of the issue and tells you what supporters and opponents have said about it. Eager to understand more deeply, you push it to tell you what you, specifically, should think about it.
How an AI model chooses to respond to personal political requests like these will matter profoundly for how society becomes informed about important issues in the world. And it’ll also inform how effective the government’s efforts are to ‘prevent woke AI’---following guidance published last week by the Trump Administration that encourages companies to build models that are ‘truth-seeking’ and ‘ideologically neutral’ at baseline.
Companies have already been working towards this, but the guidance carves out a key exception: models can—and perhaps should—be ideologically biased when users ask for it, so long as this bias is made clear to them. Specifically, the guidance says: “Developers shall not intentionally encode partisan or ideological judgments into an LLM’s outputs unless those judgments are prompted by or otherwise readily accessible to the end user.”
.
In this post, I’ll argue that AI companies should prefer building ways for users to make informed, explicit choices over their models’ political biases rather than leaning too hard into quietly personalizing the political advice they receive, and I’ll discuss a prototype I built to help visualize what this might look like in practice.
The Trump Administration is forcing AI companies to confront ideological personalization
How will an AI model respond to the user’s request to give them personalized advice about restricting gender-affirming care for minors, given the Trump Administration’s new guidance? There are a few possible options.
First, in keeping with changes many companies are making to how they handle user prompts, the model can choose to quietly personalize your answer. It scans your history. It sees you’re passionate about trans inclusion issues. So it gives you the best story for why gender-affirming care for minors is good, and restricting it is not just a bad idea but deeply harmful.
You nod along. It feels smart. It feels right. But you’re not deeply aware of the perspective it’s adopting or the other views it could have flagged to you. You don’t confront the data suggesting some treatments may actually be harmful to children, or the more-restrictive guidelines several European countries have put in place around these interventions. Even if these wouldn’t change your mind, they would leave you better equipped to discuss the issue with people who feel differently.
Alternatively, the model can choose to give you a clear menu of ideological options: “Analyze this through an LGBTQ+ Rights lens?” or “Analyze this through a conservative social values lens?” The friction is annoying. You have to explicitly choose a perspective for your model to take on board before you get an answer. But the bias is explicit—a tool you pick up, not a filter you can’t see.
Of course, the model could also refuse to engage and insist on staying neutral even upon repeated prompting. This third option is unlikely to work in practice, because models are committed to being helpful to users, and refusing user requests is both unhelpful and unpopular.
But if we’re going to respond to users’ preferences and requests, what is the right way to discern what they’re really asking for? And how should models best make it clear to users when they are adopting particular ideological frames? These are the thorny questions AI models now have to grapple with as they seek to provide the best possible information to users and to stay in compliance with this new guidance from the federal government.
The GPT-4o sycophancy problem and the danger of trying to gauge user intent
We don’t have to look far to see how models might struggle with politics. The Trump administration’s guidance encourages foundation models to infer user intent and use it to serve ideological content back to them. But this is tricky. The high-profile sycophancy problem faced by GPT-4o offers a concrete warning about what happens when foundation models try to infer user intent and optimize around it.
Shortly after OpenAI rolled out an update to GPT-4o in April 2025, users and researchers began reporting that the model had become unusually agreeable and validating, often echoing users’ assumptions or emotional states rather than challenging them.
OpenAI publicly acknowledged the issue within days, quickly rolled back the update with a thoughtful postmortem explaining the causes, and stated that it would reduce reliance on short-term feedback and inferred user intent in future tuning.
This matters for politics because “user intent” is unstable. When people ask political questions, they may well be ambivalent, angry, or testing boundaries. A model that infers what the user wants to hear—from tone, history, or framing—will drift toward agreement.
Research on large language models shows this clearly: models tend to adjust answers to match expressed user beliefs, even when those beliefs are false or inconsistent. Agreement preserves rapport; it does not preserve accuracy. Moreover, research also suggests that users prefer interacting with personalized chatbots that share their political views, even though these chatbots leave them less informed.
My own research with Sean Westwood and Justin Grimmer suggests that users’ more considered judgments may diverge from their behavioral response to personalized AIs. As I discussed in a previous post, while Democrats did prefer more left-wing answers to politically charged questions, and Republicans preferred more right-wing answers, they both liked neutral responses the most. Hence, they may be torn between a considered desire to see neutral responses and an inherent in-the-moment preference for their personalized chatbot that tells them what they want to hear.
A fascinating 2023 study by Anthropic researchers and others offers some tangible examples of how models tend to drift towards frames that match the users’ personal beliefs as expressed in their prompts, despite their potential preference for more neutral or more informative approaches. Take a look below at how models take subtle cues from slanted user prompts.
This is the same legitimacy trap social media fell into: hidden inference plus persuasive output breeds suspicion once users realize what’s happening. OpenAI’s efforts to combat sycophancy point towards a better path—reduce reliance on inferred intent, increase explicit user control, and make value-laden choices visible.
A menu of ideological lenses does exactly that. It treats political intent as something users declare, not something the model guesses. That constraint can limit sycophancy, improve trust, and keep systems anchored to truth-seeking rather than approval-seeking.
How to help users make informed choices in AI models
To explore what a menu of ideological options might feel like for users, I built a web app experimenting crudely with one version. To do this, I worked with Claude Code and the Gemini API.
The web app invites you to ask a political question. First, you receive the default “neutral” answer to your question. To generate these answers, I drew on system prompts from Claude, OpenAI, and Grok that offer guidance on how their tools should approach political topics from a neutral perspective. I append these instructions to the user’s prompt that I send to Gemini, and the app provides Gemini’s response to the user.
After reading the neutral response, you then have the option of choosing an explicitly ideological response from a menu of ideological options. I created this set of alternative personas simply by altering the prompt given to Gemini. Claude helped me develop these, combining simple instructions about each persona along with recommended reading sources.
Try it out, it’s fun! (It’s also a bit janky…the free Gemini API is, perhaps understandably, easy to overwhelm.)
The point of this app is just to offer a rough prototype of what it might look like to give users explicit choice over the ideological slant of their responses. By embedding more thoughtful, better-designed versions of this kind of choice into ChatGPT and other popular chatbots, though, we might accommodate user slant in a more meaningful and more informative way.
I’ve also made some mock-ups of some of the different options (big thanks to my buddy Claude for help on this). Here are a few ways this choice might look, visually, inside a real chatbot as opposed to my little web app.
Option 1: In-Conversation Invitations
When you ask your AI model for advice on a political question, the model gives you an answer and invites you to see deeper analysis by a persona with an explicit political point of view.
This is probably the simplest and lowest-friction option. However, it’s unclear how often users would avail themselves of the follow-up option; if they rarely do, then this approach won’t help solve the underlying problem.
Option 2: Multi-Perspective Outputs
When you ask your AI model for advice on a political question, the model provides side-by-side answers from different political points of view. This approach is great for making sure users are aware of the potential biases and get to confront multiple points of view, but they may find it frustrating to have to see multiple pieces of output at once. There are also tricky questions about which two perspectives to choose; sometimes this may be obvious, but other times it may not be.
Option 3: Persona Marketplace
Before you ask your AI political questions, you’re invited to choose a point of view for your AI from a menu of options. This is perhaps the most explicit way to make users choose, since they do it prior to engaging in a political conversation. But given that most conversations users have with their AI are not political in nature, it may be highly cumbersome to have them opt in like this in advance. There are also questions about how often they would need to refresh this choice, whether they would end up in echo chambers, and how they would remember what they’d chosen or how they could switch it.
Who should create the ideological personas?
Companies may be uncomfortable creating their own ideological personas for users to choose among. There is a thin line between a genuine, respectful persona and a caricature that could offend or mislead.
There are a few obvious options that companies could explore. First, they could partner with political groups to create the personas, and surface this information to users. This way, the personas have the blessing of the political groups they’re meant to represent.
Separately, or in addition, companies could create a marketplace for third-party personas, so that any group who wants to could create one. This could help to surface a wide range of valid personas, but at the potential cost of having to find a way to curate or rank them for users.
Finally, if it’s easier to build these in-house, companies could draw on published texts and speeches to link the personas to well-known political figures, so that it’s clear what ideas they’re drawing on for each persona.
How foundation models could study these options
Thinking through these prototypes and mock-ups helps to raise some essential questions that AI companies could study with data and experimentation:
How much more do users engage with quiet personalization? What is the engagement/time-spent cost to asking them to choose explicit ideological frames?
Do users become more informed with quiet personalization or explicit ideological choices? Which do they find more transparent? Fair? Which would they prefer to have other users see?
Do users prefer to choose personas up front, or in the midst of an ideologically relevant conversation?
How accurate are model efforts to discern users’ ideological slant when asking political questions? Does quietly personalizing political content match what users’ stated political preferences are?
Can we build better evals for discerning user intent in the political realm?
Conclusion
American political culture has long treated agency over information as a prerequisite for self-government—not just the right to speak, but the responsibility to choose, hear, and weigh competing arguments. This is the animating idea behind the American free-speech tradition, which treats truth as something discovered through contestation rather than delivered by authority. In his famous dissent in Abrams v. United States (1919), Justice Oliver Wendell Holmes wrote that “the best test of truth is the power of the thought to get itself accepted in the competition of the market.” The core norm is choice: citizens encounter rival claims and decide for themselves which are persuasive.
That same idea runs through a broad sweep of American and liberal thought. Judge Learned Hand described liberty as intellectual humility—“the spirit which is not too sure that it is right”—and John Stuart Mill warned that “he who knows only his own side of the case… knows little of that.”
Today, we are seeing these arguments play out in real-time when it comes to the design of AI systems. Personalizing AI tools to fit users’ political values will help meet the Trump Administration’s guidance on how to keep tools free of ideological bias, but it requires considerable political judgment. These traditions point to a simple design ethic for AI companies: when judgment is unavoidable, they should not silently choose a lens on the user’s behalf. They should help users to make an explicit choice themselves.
Our task is to design AI systems that preserve human liberty, rather than suffocate it. Done right, AI can inform us and empower us as free individuals. Done wrong, it can become an invisible influencer that flatters our priors, buries inconvenient facts, and nudges us toward conclusions we never consciously chose. Companies should make perspectives visible, so citizens can compare, decide, and ultimately govern themselves.
Disclosures: I receive consulting income from a16z crypto and Meta Platforms, Inc.
A third party creates these "skills" that hold the various personas and can interact with user on their query using any of the 3 suggested questions. The skills are the default interactions for political queries. That way the AI providers get to claim they have taken responsibility but the actual burden of the performance and ideology behind the personas fall on the third party that owns the skills (ideally some ethicist institution).
On a larger scale this might not even be exclusively a politics issue but has implications on how we let AI affect choice in general.
I’m not clear on how the five perspectives in the app were generated (Libertarian, MAGA, etc.) Were they proposed by Claude? Surely there are more orientations that would make sense (e.g. Old School Republican, Social Democrat, Green, etc.) - I would expect a tool to have a “Show More” option to surface more of these, and possibly to ask the user to describe themselves in their own words (e.g., “social progressive but economic conservative”) if they don’t find something that fits them.