Why we never generate a crowd
No fake reality.
Democrez never generates crowd data. No model writes responses, fills a thin card out to a believable size, smooths a jagged distribution, or estimates what a missing population would probably have said. Where verified humans have not answered, no number appears at all — and that is the whole rule, with no exceptions written into it.
Why this rule has to be absolute.
Most rules on a platform can hold partially and still be worth something. This one cannot. The moment a single figure is generated rather than counted, no figure anywhere on the platform can be trusted — because from the outside there is no way to tell which was which. A reader cannot inspect a percentage and see how it was produced. They have to take the whole set on faith or reject the whole set, and there is no middle position available to them.
That is why it is written as an absolute rather than as a preference. A preference survives until the first genuinely good reason to break it, and a genuinely good reason always arrives: a launch that needs a demonstration, a journalist on a deadline, a card that is nearly at threshold and would be so much more useful today.
What specifically does not happen
No model is asked what a typical person would choose. No historical card is used to predict a new card's distribution. No sparse result is topped up, rounded to a friendlier figure, or reweighted toward what a representative sample would supposedly look like. No placeholder number is written into a screen with the intention of replacing it later, because the intention is not visible to anyone reading the screen.
Where AI is used, and where it stops
AI drafts situations. It proposes wordings, suggests options, and helps catch a card that leans one way. That work happens entirely before a card reaches anybody, and every draft is reviewed and approved by a person before release.
It stops at the boundary of the response data. Nothing on the answering side is machine-produced: not an answer, not a prediction, not an aggregate, not an interpolation between two aggregates. The line is easy to state because it is the line between what a card asks and what people did.
What it costs
The public readings page has been empty since launch, and it will stay empty until real cards cross their thresholds and pass review. A demonstration would have been trivial to produce and would have made the platform look considerably more finished. Every argument for filling it was a good argument, which is exactly why the rule is not written as a judgement call.
Every figure has people behind it, or it does not exist. There is no third category of number on this platform, and that is the only reason the second screen is worth reading.
The boundary
AI writes the question.
People write the answer.
The distinction is not about how much AI is used. It is about which side of the card it is allowed to touch.
| AI is used | AI is never used | |
|---|---|---|
| The situation | Drafting wordings and options, flagging a card that leans | — |
| Release | — | Approving a card. A person does that, always |
| Responses | — | Writing, adding, weighting or simulating any answer |
| Aggregates | — | Estimating, smoothing, rounding or projecting a figure |
| Thin cards | — | Topping up, filling in, or predicting the missing share |
| Your Gap | — | Interpreting it, explaining it, or advising on it |
Who approves a card, and what they check. The Human Seal
The arguments we said no to
Four good reasons
to break this rule.
None of these is a straw man. Each has been a live argument, and each is refused for the same structural reason rather than on principle alone.
“Just for the demo”
A demonstration figure sits on the same screen, in the same typeface, as a real one. Nobody scrolling past can tell them apart, and the label that distinguishes them is the first thing to be dropped in a screenshot.
“The card is nearly at threshold”
Nearly is a judgement, and a judgement can be made again next week at a slightly lower bar. A threshold that bends once is a suggestion from then on.
“The model would be close enough”
It might well be. But a figure that is close enough is indistinguishable from one that is not, and the reader has no way to check which they were handed.
“It is only one number”
One generated number makes every other number unverifiable, because the reader cannot inspect provenance. There is no way to contain the damage to the one figure.
How you can check
Partly, and that is worth being honest about. You can see that the readings page is empty rather than populated with plausible content. You can see that a card below its threshold shows a count and no aggregate, rather than a figure with a caveat. You can see that every published reading carries its response count, its version and its seal date, so a figure with no people behind it would have nowhere to hide.
What you cannot do from outside is inspect the pipeline. That is a real limit, and the response to it is not a reassurance — it is that the visible behaviour of the platform is expensive and consistent with the claim, and inconsistent with the alternative.
In one line
The definition.
- No fake reality
- The rule that no response, aggregate or distribution on Democrez may be generated, estimated, smoothed or filled in. AI assists in drafting situations before release; nothing on the response side is machine-produced. Where verified humans have not answered, no figure is shown.
Two consequences you can see
- The readings page is empty. Not because nothing is being collected, but because nothing has yet crossed a threshold and passed review. See it
- Illustrative figures carry a label. Where this site shows an example of what a Gap looks like, it is marked as illustrative and stays marked. That label is not decoration; it is the only thing separating an example from a claim.
Questions
Straight answers.
Is any AI used on Democrez at all?
Yes, on one side of the card. AI drafts situations, proposes options and helps catch a card that leans one way, and every draft is reviewed and approved by a person before release. Nothing on the response side is machine-produced — not an answer, not a prediction, not an aggregate.
What if a card has almost enough answers?
It stays closed. Almost is a judgement, and a judgement that bends once can bend again at a slightly lower bar next time. The threshold is applied by the system precisely so that nobody has to decide whether today is close enough.
Are any figures on this website made up?
Where a figure appears as an example — such as the worked Gaps on the Gap page — it carries a visible illustrative label and keeps it. Everywhere else, a figure would only appear if it came from verified responses, which is why the public readings page currently shows none.
Could you generate data and simply not tell anyone?
Technically yes, as could any platform, and no assurance on a website settles that. What can be pointed at is behaviour that would be irrational under the alternative: an empty flagship page, cards that show a count and withhold the figure, and published readings that carry their own response count, version and seal date.
Why not use AI to fill gaps in under-represented groups?
Because the result would describe the model rather than the group, while looking exactly like a description of the group. A missing population is a real finding about who is using the platform, and covering it with an estimate erases the finding and replaces it with a guess.
Does AI interpret my Gap or tell me what it means?
No. Nothing follows the number at all — no interpretation, no advice, no summary of what your answers suggest about you. That restraint has its own page.
The numbers wait for people.
Every card that opens does so because enough real people answered it first.