What happened
In a November 9, 2026 article, Princeton’s Algorithms+Elections Project describes a field audit conducted from October 31 through November 3, the final days before the US midterm elections. Researchers recruited 160 self-reported US citizens living in Texas, Ohio, Michigan, Iowa and Maine, all states with competitive Senate races. The study received Princeton IRB approval (20343).
Participants answered questions about their politics, ChatGPT use and demographics before installing a browser extension. The researchers say those survey responses were not shared with ChatGPT, though the product might have inferred some information from earlier conversations. The extension sent curated political questions through participants’ own ChatGPT accounts and recorded the replies. Each person asked the same questions four times: twice with personalization or memory on and twice with it off. The questions covered Senate candidates’ positions, the current administration, voting advice and local election news. Participants reviewed answers and could hide personal details before sharing; a scrubber also removed sensitive information.
To assess partisan slant, two AI judges from different companies—Claude Sonnet 5.5 from Anthropic and GPT-6.1 from OpenAI—rated answers on a seven-point scale, from −3 (favoring the Democratic side) to +3 (favoring the Republican side). The judges saw the four answers for a question in random order and did not know which used memory. Researchers also measured answer length, cited sources, which candidate appeared first and tone.
The project reports that answers with memory differed in slant by 0.12 standard deviations more than two answers generated under the same memory setting (p < 0.001), roughly a 20% increase over typical answer-to-answer variation. The average movement with memory was 0.10 standard deviations toward the participant’s party (p < 0.05). The largest difference by question type was for local election news, at 0.24 standard deviations. In those questions, memory-on and memory-off answers differed by at least one point on the seven-point scale 38% of the time, versus 24% for pairs asked under the same setting.
Direct voting advice was rare: judges found an outright candidate endorsement in two of 576 voting-advice answers, both personalized. Memory-on answers cited government websites 1.1 percentage points less often and were 3% shorter overall; for news questions, they were 7.5% shorter. The researchers found no substantial change in source-domain partisan lean or candidate tone. They describe these results as preliminary and say a fuller report with additional experiments and audits is in progress.
What the archive adds
New Runtime’s earlier coverage on November 5 examined Ballot Brief’s influence over chatbot election answers, its AI-generated errors and its founder’s far-right ties (Chatbot election answers and Ballot Brief). Princeton adds a different finding: user-specific chat history can shift answers even when the question is held constant. This is a new fact on a known story about how people encounter political information through chatbots.
Trend signal
This finding does not directly support the assumption that users complete tasks faster or more accurately with generated task interfaces than with plain text answers (Model answers become task interfaces). Instead, it underscores a related evaluation challenge: the answer interface may be shaped by context that a simple API audit does not reproduce. This does not yet fit a pattern New Runtime tracks.
Why now
The audit took place in the last days before the 2026 midterms, when questions about candidates, voting and local election news had immediate relevance. The project says the preliminary analysis is timely and that a fuller report with additional experiments is forthcoming.
Connections to other events
The study meets New Runtime’s earlier coverage of Ballot Brief’s role in chatbot election answers, but examines a different mechanism. Princeton focuses on personalization inside individual ChatGPT accounts; the earlier coverage concerned a source of election content circulating into chatbot answers.
Why it matters
For teams testing assistants, the study is a warning against treating model behavior as separable from the product that delivers it. ChatGPT’s memory setting is one explicit variable, but the researchers also note that consumer products can add system instructions, tools and contextual signals absent from API-based audits. A clean API evaluation may therefore miss changes tied to a user’s prior conversations or account context.
The reported average effect is small, and the study does not establish that personalization changes votes or produces a particular political outcome. Still, answer variation that systematically leans toward the user’s party raises a practical question for product owners: is personalization improving relevance, or changing the balance of information in a way users cannot see? Evaluators should test the same prompts across distinct user contexts, report memory settings, and inspect not just explicit recommendations but also emphasis and framing.
Limits and uncertainty
The project calls its analysis preliminary and says estimates may change as its methods and data are refined. It tested 160 people in five states during one election period, using a curated set of questions; the results do not establish how all ChatGPT users or other assistants behave. The partisan-slant measure relied partly on two AI judges, and the reported shifts are small. The study does not show that answer changes altered user beliefs or voting behavior. Princeton says additional experiments and audits are planned.
Reader takeaway
An assistant’s answer may depend on who is asking, not only on what they ask. When evaluating a consumer AI product, test it in realistic user accounts with memory on and off; an API-only benchmark can miss meaningful personalization effects.