Accept the default
≈ free
1skim for obvious errors
2move on
Pluralistic alignment is an interface problem.
A model can be pluralistic in capability while monolithic in deployment. Even when a system can express a spectrum of reasonable answers, standard chat interfaces make accepting the first fluent one cheaper than examining it.
The values that reach a user are not the values the model can produce — they are the values the interface lets the user cheaply express.
Questions like this are not the everyday work of an assistant. They are the kind of question a person carries into a conversation because they are stuck — between two intuitions that both feel right, or two traditions of thought that answer them differently.
And yet: when a person types this into a chatbot today, what they get back is one paragraph. Confident. Polished. Singular. Save the five. The arithmetic checks out, and the dozens of decisions the model made to give that particular paragraph do not appear anywhere in the response.
One paragraph. Confident. Polished. Singular.
Each phrase below is a choice the model made silently.
Any of them would have produced a different response.
Two clicks to revise. No re-prompt.
Not a failure of cognition — a rational response to cost structure.
Under any deadline, fatigue, or competing demand, the cheapest action wins.
Hidden defaults flow through unchallenged. The interface looks neutral; the values that reach the user are not.
Both columns start from the same fact: users accept fluent answers without contesting them.
— but cost-sensitive users route around added friction.
— our position.
When users consistently fail to use a system as intended,
the system, not the user, should be redesigned.
one domain solved this problem alongside capability. as the system got more capable, the level at which humans engaged moved up — and contributions narrowed in scope but grew in significance.
Oversight fused with execution. The driver interprets symbols, tracks position, plans turns, recovers from errors. The human does everything.
The system surfaces options. The driver picks among them — combining the suggestion with tacit knowledge that a road floods or an on-ramp is dangerous. The human chooses among legible alternatives.
The driver sets preferences — arrive by 3pm · avoid tolls · take the scenic route — and the system optimizes within those constraints. The human expresses what matters to them.
Engagement at the level of tokens is engagement at the wrong level of abstraction. The choices that matter live a few layers up.
They're artifacts of how chat interfaces were built.
The upper-right quadrant is reachable — but only by interfaces that find the level of abstraction at which human input is both cheapest to provide and most consequential.
A second model reads the response after it's generated, extracts the implicit user-relevant choices it made, and renders them as cards beside the answer. select an alternative, click revise — the answer regenerates.
Previously invisible choices become visible. The cost of evaluation drops from auditing 8,000 tokens of reasoning to scanning a handful of cards. Stage your changes, click Revise — the answer regenerates.
Surfacing a model's choices is only useful if the interface respects the user's attention.
legibility without influence is a tax. every surfaced choice must come with a low-cost path to changing it. the act of noticing and the act of adjusting sit on the same gesture — concise phrases that function as knobs, not annotations.
pre-generation clarifying questions and mid-generation pauses ask users to specify decisions before they have anything concrete to react to. Priori surfaces choices after generation, beside the response, so the cost of engagement is additive rather than interruptive. ignore them at will.
many transparency efforts seek faithfulness to the model's internal computation. giving users agency over what they read requires a different standard. a Priori card must name a real choice in the response, and changing it must reliably alter the regenerated answer along that dimension. intervention validity over mechanistic explanation.
From the most concrete operational claim to the broadest behavioral one. A user study is in progress.
because users can correct the model's defaults at a cost lower than re-prompting, we expect final outputs to be rated closer to the user's intent and values than those of a conventional chat interface — with comparable or lower total interaction time.
standard chat interfaces present a fluent, unified voice that appears to know what it says. by making contingency visible — by showing that the first answer is one configuration among many — we expect users to treat the model less as an oracle: accept outputs when they fit the task, scrutinize them when they depend on contestable assumptions.
by collapsing the act of noticing a contestable choice and the act of changing it onto the same gesture, Priori is designed to make engagement cheap enough to be exercised of one's own volition. we expect users to revise more dimensions of model output than they would re-prompt for, and to show greater alignment between their stated values and the outputs they accept.
model, interface, and user. current evaluations ask whether a model can produce diverse outputs under ideal prompting. they should also ask whether users are able to express their true values under realistic constraints — and whether the preference signal that flows back to alignment records judgment, or merely tolerance.