Model Transparency

When AI features first went into products, the assumption was that only the output quality mattered. Users would judge whether a summary read well, not who wrote it, and saying a machine produced it seemed likely to make them trust it less. So generated sentences went onto the screen looking exactly like sentences a person had typed.

Pagedeck's first incident with its automatic meeting-note summaries arrived that way. A customer took a date from a summary at face value, forwarded it to a partner company, and that date had never come up in the meeting. When the customer asked where the sentence came from, neither Sofia nor Mia could answer. The screen carried no mark that the text was generated and no pointer to the part of the transcript it supposedly came from. Someone ended up replaying forty minutes of recording just to confirm the passage did not exist.

Model transparency is the principle the team put in place to stop that on the product side. The starting move was separating two questions: how to make the output better, and how to make the output possible to doubt. Machines get things wrong. The conclusion the team landed on was that when they do, the screen should already carry enough material for the reader to notice it.

It comes down to surfacing three things — that AI was involved, what the answer rests on, and a record of what ran.

1. The involvement notice. Mark next to the output that a machine, not a person, produced it. Pagedeck labels a summary as an automatically generated draft and adds one line asking people not to forward it without review. It looks like a small thing. In practice that one mark changes how people read what follows it.
2. The grounding. Point each sentence of the summary back to the passage it came from. At Pagedeck, clicking a sentence jumps to that segment of the recording. A sentence with nothing behind it gets no marker, so a fabricated date has nowhere to jump to and gives itself away on the spot.
3. The record. Log which model ran, what went in, and what came out. Reproducing an incident afterward depends on having that log, and when a regulator asks for evidence, this record is usually what they mean.

All three exist because of hallucination — the way a model states things that aren't true in perfectly plausible sentences. Wrong answers arrive in the same confident tone as right ones. Human writing leaves tells: hedging, an assertion with nothing behind it. Model output leaves none, so the reader has no way to know where to be suspicious, and the product has to point for them.

How far to go varies by feature. A drafting tool that a person rewrites anyway needs only the notice; a feature that answers customer tickets on its own needs the grounding too. The deciding question is not how advanced the model is but what the user stands to lose when the output is wrong.

Three places ask for it in practice: the law, enterprise purchasing reviews, and your own team's response when something goes wrong.

The law is the first of them. The AI Act rates each use of AI by risk level and attaches different duties at each level, and this notice is the duty at the lightest tier. For summaries and drafts, saying that AI made it is enough. Move up to screening job applications or deciding loans and the duty grows to include documenting what the system was built from and keeping records of its decisions.

Purchasing reviews catch it too. Enterprise customers send a questionnaire on security and privacy before signing, and questions about AI features now sit in it: which model runs, whether generated output carries a notice, whether customer documents feed model training. That last one belongs to training data rights, which covers the copyright and permitted use of material a model learned from. A blank answer on that sheet stalls the contract right where it stands.

The third place is the incident itself. With grounding in place, the date incident would have ended in five minutes — the customer clicks the sentence, nothing opens, everyone knows. This is also why the principle pairs with human in the loop, the design of leaving a person to check and approve before consequential decisions go out. If the person doing the checking can't see what the output rests on, that step hardens into clicking an approve button.

"Does this summary carry the generated-content notice? We can't ship without it."

The minimum disclosure required by regulation and internal policy is missing.

"The customer questionnaire has new questions about our model. We need to update our answer doc."

It feeds contract review, so it should match what the product actually does.

"Source links aren't happening this quarter. Let's at least put the notice on."

A proposal to take the lower tier now and add the rest later.

At Pagedeck, Grace asks for a one-page checklist whenever a new AI feature starts. Four lines: which model, what the user is shown, what the user loses when the output is wrong, how long the logs are kept. It gets filled in at design time rather than after the feature is built. Retrofitting means opening a screen that already runs, and at that point adding one notice turns into a quarterly item.

Many people read this as a demand to publish the model's internals. The word transparency suggests being able to see straight through something, so the connection is an easy one to make. But the principle asks for the information a user needs to judge an output, not for the algorithm or the trained weights. What actually gets asked is usually three things: was AI involved, what was it based on, how far can this be trusted. Set the scope too wide and the company concludes that trade secrets make it impossible, and the disclosures that were perfectly shippable never ship either.

The other common reading is that the notice is the whole job — small grey text under the box saying this was AI-generated. That line shifts responsibility to the user without telling them what to doubt. Had it been there during the date incident, the customer would have forwarded the wrong date anyway. The notice is the opening move; what changes a reader's judgment is showing where each sentence came from.