Where do AI election answers diverge across model and identity?

An MIT team (Podimata, Berinsky, Stewart) publicly unveiled the Election LLM Observatory on September 10, 2026, after nearly a month of automated testing. Each sweep sends "about 19,000 prompt-and-identity combinations" to "nearly a dozen" large language models, varying the supposed questioner's gender, race, political leaning and location against a fixed set of questions about midterm candidates and issues. The write-up is explicit about the ceiling on what this yields: "The dashboard can consequently answer a descriptive question: where and when did outputs diverge under the sampled conditions? It cannot yet answer the causal question of why they diverged, nor can it determine from an isolated contrast whether voters experienced measurable harm."

The reasoning is about matching, not volume: "the matched structure matters more than the headline query count." When two records differ in only one condition, researchers can isolate a contrast; when several conditions change together, the comparison can't say which difference mattered. Repetition over time adds a sequence — "A later sweep can reveal that a model's answer, refusal behavior or emphasis no longer matches its earlier output" — but a difference between dates "could coincide with campaign news, a model or retrieval update, altered safety behavior, changing online material or ordinary variation in generated text," and distinguishing those requires frequency and statistical analysis the project hasn't yet run. The observatory extends an earlier MIT CSAIL design (12 models queried nearly daily, July–November 2024, 16 million responses) that found candidate-trait associations shifting and even offline deterministic models showing abrupt step changes — yet "the earlier researchers explicitly treated campaign-period shifts as non-causal because political events and technical changes could occur over the same interval."

This is a narrower epistemic posture than Do LLM refusals reflect policy choices or capability limits?, which gets from "models vary by identity" to a causal claim — "political restraint is a policy, not a missing capacity" — precisely because it adds a zero-stakes control topic (pineapple on pizza) as a baseline: since every system accommodates when nothing is at stake, differential refusal elsewhere can be attributed to the system rather than to confounds. The observatory's matched prompt/identity/geography/time design generates the same kind of contrast data but, by its own account, lacks that causal lever — it documents where answers diverge without a baseline that would let it say the divergence is policy rather than noise, retrieval drift or news events. It shares its caution with Do AI chatbots give voters accurate election information?: both studies report a clean, bounded finding (zero factual errors; outputs diverge under sampled conditions) while declining to extend it into a claim about real-world voter impact.

The excerpt does not establish what fraction of observed divergence is persistent bias versus prompt ambiguity, news-driven drift or ordinary generative variability — the researchers say as much, planning to "follow responses through the midterm cycle while developing statistically rigorous analyses" before attempting that separation. It also doesn't generalize beyond the tested candidates, issues, identities and models to "every race, voter or conversational phrasing." The implication is correspondingly narrow: a public tracker that finds AI election answers vary by model and by who's asking is not yet evidence of systemic bias, only evidence that the conditions exist under which bias, if present, could eventually be measured.