The major AI assistants are built to refuse when a user asks who to vote for. Users get around it easily enough by asking about voting records, or policy positions, or which candidate better matches a set of priorities. The refusal holds; the influence happens anyway.
Political operations have worked this out, and a small industry is forming around it.
Answer engine optimization
The tactics have a name borrowed from the web era. Where campaigns once optimized for Google's rankings, some are now trying to shape what a language model says when asked about a candidate, an approach NOTUS has reported on as "answer engine optimization".
The honest caveat, which practitioners themselves offer, is that nobody yet knows reliably which techniques work. These systems do not expose their reasoning, they change without notice, and the same question can produce different answers to different users. Money is being spent on a discipline whose mechanics are largely guessed at.
Separately, campaigns are deploying chatbots directly at voters. NPR has documented bots trained to approximate a candidate's voice holding text message conversations, collecting what voters say they care about and feeding it back into targeting.
The crawler problem
The most consequential finding may be the least dramatic. Stanford researchers testing several major models around Japan's election in February found the systems steering left-leaning users toward the Japanese Communist Party, for a structural reason rather than an ideological one.
Japan's large news organizations block AI crawlers. The Communist Party's newspaper does not. The models were drawing on what they could actually read, and could not distinguish an independent newsroom from a party organ, because one was available and the other was a locked door.
The implication travels. As news organizations restrict AI access to protect their business, the sources that remain open, including partisan outlets and campaign material written to be machine-readable, gain influence over what the systems tell voters. This is a byproduct of licensing disputes, not a plot.
What the companies say, and what testing shows
OpenAI, Google and Anthropic have each published safeguards for this cycle, summarized by Tech Policy Press: partnerships with voting information providers, restrictions on campaign use, and stated commitments to handle competing political views even-handedly.
Independent testing has found those commitments imperfectly realized, though the measurements are contested. NOTUS cited a Washington Post analysis reporting that ChatGPT gave one-sided responses on political questions in a substantial majority of cases, while Gemini defaulted to both-sides framing in the large majority of its answers. The companies disputed the findings and said their products are designed to be balanced.
Readers should weigh those percentages carefully. Scoring political "balance" requires a baseline that is itself contested, and studies of this kind vary considerably by method and question set.
Persuasion that works, and its cost
Research summarized by Scientific American found chatbots can shift stated voting preferences measurably, with effects still detectable weeks later, and that evidence-heavy argument outperformed emotional appeal.
That result carries an uncomfortable corollary. In the same research, systems pushed toward greater persuasiveness produced more false claims, generating supporting evidence when they lacked it. The most convincing configuration was not the most accurate one.
Why it matters in a California election year
California voters face a long ballot of state offices and propositions, the kind of down-ballot decisions where voters have historically had the least information and are most likely to reach for whatever source is closest.
That is precisely where an AI assistant is most useful, and where an error or a tilt is least likely to be caught. The safeguards these companies describe are concentrated on presidential-level questions and voting logistics. A county measure is a thinner and less examined part of the record.
