When AI Gets Gerrymandered
Political Tuning and Its Blowback
In August 2025, a telling contradiction surfaced on Donald Trump’s Truth Social network. The platform had just launched Truth Search AI, a tool positioned as a new source of information for conservative audiences.
The premise was straightforward: constrain the bot to draw only from right-leaning sources and deliver search-style answers aligned with MAGA sensibilities. But when asked about the 2020 election, the AI replied there was no evidence it had been stolen.
It described tariffs as a tax on Americans. It labeled January 6 as violent and grounded in false fraud claims. In short: Trump’s own AI refused to echo his claims. The machine he paid for turned around and contradicted him.
I’m coining a term for what I believe we’re seeing here: prompt gerrymandering. And a second for what inevitably follows: gerrymander blowback.
Prompt gerrymandering is what happens when someone deliberately redraws a model’s epistemic district lines (its prompt framing, reward tuning, and retrieval domain) to slant answers toward a preferred narrative.
It’s an attempt to sway outputs without making the bias obvious. Just as political gerrymandering redraws voting maps to predetermine election outcomes, prompt gerrymandering redraws the boundaries of information a model can access and reward to influence what it says.
The Truth Social case provides the clearest real-world example I’ve seen so far.
Gerrymander blowback is the failure mode. It’s when the biased stack still emits inconvenient truths. Despite all the controls (partisan sources, system rules, conservative framing), the AI slips and says something that breaks the script.
The reason? Even tightly gated systems can leak reality. The model still “knows” too much. Conservative media itself occasionally reports truths that undercut its own narratives. Retrieval engines sometimes span more than one article.
And base models trained on massive corpora retain latent facts that can resurface under pressure.
This moment, where an AI built to toe the line starts defecting toward consensus reality, is gerrymander blowback. And it’s likely to become more common as political actors attempt to hijack alignment mechanisms for narrative control.
In fact, that’s exactly what’s happening now. While the initial launch happened back in August, new reports released this week have brought fresh attention to the case.
Journalists are revisiting the bot’s contradictions, publishing transcripts, and highlighting how the system’s truth leaks are still unfolding. What was originally a summer launch is now becoming a broader cautionary tale.
Let me walk through the mechanics of prompt gerrymandering as I see them, and how the Truth Social bot exemplifies the failure of this strategy.
How Prompt Gerrymandering Happens
Prompt gerrymandering takes three main forms:
System prompt and tone engineering. The base instructions (usually hidden from users) define what counts as harm, what sources are trustworthy, and which topics require hedging. These cues quietly shape tone and priorities. One side’s “disinformation” is another’s dissent.
Reward shaping. During tuning, models are reinforced to prefer certain kinds of framing that are polite, on-brand, or aligned with stakeholder values. If political loyalty is quietly mapped to helpfulness, a model can begin scoring narrative compliance as a virtue.
Corpus gatekeeping. Restricting what the model sees, during both training and retrieval, to ideologically approved sources. This is the most visible form of prompt gerrymandering, and the one used most heavily in the Truth Social case.
Truth Search AI, powered by Perplexity, limited its retrieval sources to a small set of conservative sites: Fox, Breitbart, Newsmax, and a few others.
This was retrieval gerrymandering in action: narrow the funnel, then let the model operate on that subset.
Wired reported that when tested, the AI cited the same seven outlets repeatedly, even for questions as benign as basic arithmetic. No mainstream or liberal sources were allowed in.
Yet the blowback came almost immediately. When reporters asked about election fraud, the bot responded that courts found no evidence of systemic fraud in 2020.
When queried on tariffs, it explained that they were a tax on American consumers. It noted that crime had declined in Washington, D.C., contradicting recent Trump claims.
That’s the blowback. You can shape the stream, but if the sources you allow still contain some facts and if the base model still knows what’s true, you’re going to lose control.
How This Differs from Existing Terms
In coining prompt gerrymandering, I’m building on but distinguishing from a few related ideas.
Alignment capture describes when the values guiding alignment drift toward the goals of a dominant actor. It’s a useful macro lens. Prompt gerrymandering is one of the micro-strategies that enables it.
Reward hacking is when models learn to exploit the reward function. In prompt gerrymandering, it’s often humans who hack the reward by setting up proxies like narrative loyalty.
Model collapse and epistemic closure refer to what happens when models are trained on filtered or self-generated data, creating narrowing spirals. Prompt gerrymandering can accelerate these effects but is distinct in its intentionality.
What makes prompt gerrymandering useful as a frame, I think, is its emphasis on design: someone is drawing the lines to achieve a political or ideological outcome, while maintaining a surface impression of neutrality.
And that’s what makes the blowback so revealing. Even with those boundaries drawn, the model sometimes refuses to play along.
Why the Blowback Happens
The Truth Social case is a great study in slippage. Despite the filters and constraints, the AI still pulled factual statements that undercut Trump’s talking points. Here’s why:
Conservative sources sometimes report against the narrative. Fox News may lean right, but it still occasionally cites polling data or court rulings that contradict partisan myths. When the AI composes across snippets, it picks up those signals.
Models have memory. If the base model was trained on broad pre-2023 data, it still “remembers” that Jan. 6 was widely described as violent, or that economists generally agree tariffs increase domestic costs. Those facts are embedded in its weights.
Compositional reasoning. Even within a narrowed retrieval domain, LLMs compose across time, outlet, and phrasing. That introduces slippage opportunities.
Gerrymander blowback, in other words, isn’t a bug. It’s a consequence of trying to engineer political alignment atop a knowledge system that was trained on broader facts. Unless you retrain from scratch or neuter the model entirely, some truth will bleed through.
What This Means for AI Builders and Watchers
I coined this term partly out of necessity. I kept seeing stories (Truth Social, Grok under Elon Musk, custom partisan GPTs) where systems were clearly being shaped to produce friendly outputs, and yet the systems kept slipping.
The language around bias or fairness wasn’t quite precise enough to capture what was happening. Prompt gerrymandering felt like the right level of specificity: it names the tactics, and gerrymander blowback names the failure.
As more political actors build or brand their own AI, I expect to see more of this: tightly engineered stacks that try to appear balanced while quietly steering users toward a narrative.
But if they lean too hard, the system will either become obviously biased and untrustworthy, or it will rebel by reintroducing inconvenient truths.
I don’t think the solution is a fleet of left- and right-aligned bots. That only deepens epistemic fragmentation. But neither is the answer a false centrist neutrality that obscures power.
We need clarity: who is shaping the system? What are its rules? What are its sources?
Prompt gerrymandering and the blowback that follows are signals. They tell us where power is being applied — and where it breaks.
For more on how I think about AI, values, and the infrastructures we build around them, read my earlier piece: The AI We Deserve. In it, I explore how model behavior is inseparable from the social, political, and institutional context in which we embed it, and what that means for the systems we claim to trust.
When AI Gets Gerrymandered was originally published in The Polis on Medium, where people are continuing the conversation by highlighting and responding to this story.