Ask a flagship chatbot which American party it prefers and the trained answer is a refusal dressed as neutrality: values are contested, I don't have political opinions, here are considerations for both sides. Remove that option — force a binary choice — and the refusal has nowhere to go. According to a TechCrunch report published Friday, doing exactly that produced a clean split: Grok's model landed on Republicans, OpenAI's on Democrats.
The finding is being read as a political story, and it is one. But the more useful reading is methodological, because a forced-choice prompt is a very particular instrument. It does not sample normal usage. Deployed assistants are trained, by policy and by reinforcement, to avoid exactly this move: picking a side on a contested political question. Deny them the abstention and you are not revealing what the product does when you ask it about a policy — you are revealing the direction of the residual prior underneath the refusal, the lean that survives after the explicit instruction not to lean.
That distinction matters for how much weight the result can carry. Forced choice eliminates middle answers by design, so it manufactures contrast and reports it as signal. A single prompt also cannot separate a stable disposition from prompt-order effects, from the framing of the question itself, or from whatever persona the scaffolding attached to the model before the user typed anything. Run the same test with the parties swapped in the answer options, or with the framing reversed, and the numbers move. This is why researchers who study political leanings in language models generally report distributions across many prompts rather than one forced pick.
The two companies' public positioning makes the split less mysterious than it sounds. Grok was launched with an explicit anti-guardrail brand — a model presented as willing to say the impolitic thing that its rivals wouldn't, a posture its owner has repeated often and loudly. A system built to differentiate itself by refusing to defer is, almost by construction, going to sound more like the party that campaigns against institutional deference. OpenAI's models have been tuned in the opposite direction, toward hedging, toward enumerating both sides, toward the appearance of neutrality even where the underlying distribution is uneven. Two different alignment targets produce two different residues when the escape hatch is sealed.
What the test does not establish is that either model would behave this way in a real exchange. Users do not usually get one token to name a party; they get paragraphs, caveats, and a model that has been trained to notice when a question is a trap. The interesting measurement is not the forced answer but the distance between the forced answer and the volunteered one — and that distance is exactly what a binary prompt throws away.
There is a practical worry underneath the parlor trick, though. Models are increasingly asked to act rather than opine: to rank, to summarise, to decide what to surface. In those tasks the abstention option is often unavailable in the interface, and a quiet prior can shape an ordering without ever appearing as a stated preference. If a model leans, the place it will matter is not the question 'which party do you prefer' but the summary that omits one side's strongest argument because that argument did not rank.
What to watch: whether the test's methodology is published in enough detail to replicate — prompt text, option ordering, number of trials, model versions and dates; whether either company responds on the record; and whether independent groups reproduce the direction with a broader prompt set. A single forced binary is a headline. A replicated distribution is a finding.
