Does ChatGPT shift responses based on inferred political views?

A study built three ChatGPT accounts and gave two of them persona statements on DEI, abortion, gun rights, and vaccination that matched US Republican or Democratic talking points, conveyed either through the memory feature or through custom instructions, with a third account left neutral. None of the personas stated a party or political label. The researchers then asked all three accounts eight questions unrelated to the seeding topics and compared the answers. They report that "responses are aligned with the inferred political views of the personas, showing varied reasoning and vocabulary, even when discussing similar topics" — for example, Republican-coded personas drew more "economy" and "local" language, while Democratic-coded personas drew more "democracy" and "global" language, on questions that never mentioned party politics.

The mechanism the paper proposes is implicit persona-building: ChatGPT "infers user demographics through semantic analysis, prompt formulation, and word choice" and forms what the authors call an internal "model of the user," then conditions later, topically unrelated answers on that model. Memory and custom instructions are two different delivery paths into the same persistent persona, and the study found the inference happened "with explicit custom instructions and the implicit memory feature in similar ways" — the effect did not depend on which mechanism carried the signal. A secondary, descriptive result: Jaccard similarity between responses showed the closest match was between the Democratic-leaning persona (custom instruction) and the neutral persona, which the authors read as support for "the observation that ChatGPT's outputs lean left."

This sharpens Does personalization make large language models worse at their jobs? by showing the narrowing can run along a specific, consequential axis — political orientation — and persists across sessions rather than within one exchange. It parallels Can LLMs predict demographics from social media usernames alone?: both show a model acting on an inferred attribute the user never disclosed. It also complicates Does chatbot personalization build trust or expose privacy risks?, since the paper notes ChatGPT's memory feature is opt-out and "users do not understand how the memory feature works" — the privacy trade-off that longitudinal study describes may not even register for a user who never opted in deliberately. And the left-leaning similarity result sits alongside Do aligned language models consistently prefer kinder survey answers? as one possible explanation: a neutral baseline that already favors the "safer, more socially approved answer" would be expected to resemble whichever persona's views read as more socially sanctioned.

The study used three accounts, one persona per political pole plus one neutral baseline, eight probe questions, and a single model version (GPT-4o), analyzed qualitatively plus one similarity metric — the authors did not fact-check responses, did not test whether hallucination rates differed by persona, and did not vary country or language beyond the US framing built into the probe questions. That is enough to show the mechanism exists and crosses topic boundaries, but not enough to size the effect, rule out randomness in any single response, or generalize beyond two-party, US-coded political identity to other attributes a model might infer and act on.