We tell you which model wrote your answer
The industry default is to hide the model. The answer arrives from nowhere in particular — no name on the door, no indication of what produced it or what that thing is bad at. The interface is polished specifically so you stop asking.
That is a design choice, not a neutral state, and it has a cost the person paying it cannot see.
Attribution is ordinary, and we removed it
When a colleague answers a question, you factor in what you know about them without noticing you are doing it. Their background, their tendencies, the areas where they are excellent and the areas where they are confidently approximate. You know whose number to take at face value and whose to check. This is not cynicism. It is the ordinary calibration that makes working with other people possible at all.
The same calibration applies to models, and for the same reason: they differ. Different training emphases, different knowledge boundaries, different behaviour when the evidence is thin. One may be strong at synthesis and weaker at precise recall. One tuned toward caution will hedge where directness would have served you. These are real, they are stable enough to learn, and they are relevant to whether you should act on what you just read.
Hide the model and you remove the reader's ability to do any of that. Every answer has to be treated as coming from the same source with the same reliability — which means treating all of them with the caution the worst case deserves, or, far more commonly, treating all of them with the trust the best case deserves. Neither is right, and the interface has made the correct behaviour unavailable.
What we show
When you receive an answer in askFinz you can see which model produced it, the sources it drew on, and where a claim came from retrieved material rather than from the model's own reasoning. On models you can read about what is available and what each is known to be weak at.
The third of those is the one that took the longest to get right and it is the one that matters most. "Here are some sources" and "this specific sentence rests on this specific source" are very different products. The first is a bibliography. The second is checkable.
Two alternatives we tried and rejected
Show the model only on request. The compromise version: keep the interface clean, put attribution behind a disclosure. We built toward this and abandoned it, because a default is not a neutral container for a preference — it is the answer for almost everyone. Something behind a click is something that exists for the small number of users who were already suspicious, which is precisely the population that needed it least. If attribution only reaches the people already checking, it has changed nothing.
Explain the choice instead of naming it. The more sophisticated version: rather than a model name, which means nothing to most readers, show a plain-language rationale — this was routed here because the question was long-form synthesis. We liked this and dropped it, for a reason worth stating plainly. An explanation of a choice is not evidence about an output. It reassures without informing, and reassurance that is not backed by anything the reader can check is the exact failure mode this whole note is about. It also has a nasty property: the better written the rationale, the more it substitutes for the reader's own judgement.
What survived both attempts is the plain, slightly ugly version: name the thing, show the sources, mark which claims are retrieved. It is less elegant than either alternative. It is the only one that gives the reader something to act on.
Citations are not a footnote
The same logic runs through sources.
A result that cites is not merely more rigorous. It is differently useful, because it converts a closed conclusion into an open thread. You can follow it. You can check whether the source says what the summary claims it says — which, often enough to matter, it does not quite. You can revise your view when a source turns out to be less authoritative than its confident summary implied.
A result without sources hands you a conclusion and no means of testing it. In professional work that is not a smaller version of the same product. It is a different product, and a worse one, because the thing you needed was never the conclusion.
This is what we mean by trust: not a privacy checkbox or a compliance badge, but the specific act of giving you what you need to make up your own mind.
What visibility does not fix
Two limits, stated rather than discovered later.
A model name is not a quality signal. Knowing which model answered tells you what to be alert to. It does not tell you whether this particular answer is right, and it is possible to read attribution as reassurance — a good model wrote this, therefore it is good — which is worse than no attribution at all. We show the name because it enables calibration, not because it certifies anything, and the distinction is easy to lose.
Attribution has a floor of usefulness that depends on the reader. For someone who has never compared model behaviour, a name is noise. The value accrues over time, as you build a sense of where different capabilities serve you well. That is a real onboarding cost and we impose it on everyone. We think it is the right cost — the alternative is a tool whose behaviour you can never learn — but it is a cost, and it lands on new users hardest.
What we would do differently
We would have shown the absence of a source with the same emphasis as the presence of one.
Everything we built pointed at the citations: making them accurate, making them land on the right sentence, making them fast to open. What we underweighted is the far more important negative case — the claim that came from the model's own reasoning with nothing behind it. That case is the one where a reader most needs a signal, and for a long time it was simply the quiet default: a sentence with no marker, which reads identically to a sentence nobody has got round to citing yet.
Marking what is grounded is only half the job. Marking what is not is the half that changes behaviour, and we did it second.
There is a version of AI that is confident without being accountable. It sounds authoritative and leaves no trail, and it is very good at making you feel the question has been answered. We are trying to build the other one — the version that tells you where it looked, what produced the output, and what it is unsure of. It is less magical. It is a great deal more useful.