Give AI the reader level as a number: 90 of 100 understand── Replace "easy to understand" with a reader group and the share who must understand, and the pass mark is set before the AI starts revising

01The operator replaced "make it easy to understand" with a reader and a share

On 9 September 2026, the operator gave an AI a new job. It was to summarise each day's developments in AI and turn them into a 15-minute explainer video. The request also set a level for the content: material that 90% of a group with an IQ of 100 could understand.

Before making any video, the AI returned a concept for the video page. One of its principles was to publish only material that met the target for plain language. The number the operator supplied had become, on the AI's side, a condition for publication.

Two days later, on 11 September, the operator set the same level for a design page and asked for a rewrite. Neither request said "make it clearer" or "make it simpler". Both said who the reader was and what share of those readers had to understand.

The point worth testing is why this kind of instruction works on an AI, and whether it carries over to documents in promotional review and medical affairs.

02Stated as a number, the goal splits into a reader group and a share

"Easy to understand" is a single adjective. Restated as a number, it becomes two parts. One is the group of readers. The other is the share of that group who must be able to understand.

IQ is the score on an intelligence test, scaled so that the population average is exactly 100. A group with an IQ of 100 therefore means ordinary readers, neither unusually expert nor unusually weak. Ninety percent means 90 of every 100 such readers must follow the text. Only 10 can be left behind.

03An adjective gives neither the AI nor the person a place to stop

Ask for "easy to understand" and the AI changes something and reports that the text is now easier. The requester reads it and still finds it hard. They ask again. This loop has no finish line.

Without a finish line, two costs follow. The first is the number of rounds. Each time, the person has to find words for what is still missing. The second is a split judgement. The AI believes it has fixed the text, and the person believes it has not. Nothing between them decides who is right.

A numerical target puts that standard in place first. If the line is 90 out of 100, the AI can review its own revision against it. The person can point at a remaining term or sentence and say it will not reach 90 readers. The reason for a redo moves from a feeling to a specific place in the text.

04A numerical target needs four parts: reader, share, test and scope

Writing a percentage alone is not enough. If the reader group is not fixed, nobody knows 90% of whom. The target becomes usable only when four parts are in place.

The target also has limits. First, the number is a goal, not a measurement. When an AI reports that a text now meets the 90% level, that is the AI's estimate. Whether 90 readers really understand becomes known only when people close to that group read it.

Second, when a document is written for specialists, the reader group becomes those specialists. Applying the ordinary-reader level to material for physicians would strip out technical terms the readers need. Keep the numerical form and change who is in the group.

05Pharmaceutical labelling already sets a 90% pass mark for understanding

Judging a document by the share of readers who understand it is not new in pharmaceuticals. The European Commission's guideline on the readability of the labelling and package leaflet of medicines, revised in 2009, is one example. It expects patient leaflets to be tested with people who resemble the intended users.

The test method the guideline describes uses at least 20 participants. A satisfactory result is when 90% of participants can find the requested information, and 90% of those can show they understand it. With 20 people, that is 16. The criterion must be met for each question, and results cannot be averaged across questions.

Web content has a comparable rule. WCAG 2.1, the W3C's guidelines on web accessibility, says that text needing more reading ability than lower secondary education, about 7 to 9 years of schooling, should come with a simpler version or supplemental content.

Teams in promotional review and medical affairs can set a reader group and a share for each kind of document.

Passing this line to the AI along with the drafting request cuts the number of "hard to follow" returns before the draft ever reaches review. Which line to adopt depends on company rules and on the readers. It is not the AI's decision.

06An AI that is not told the reader writes at the level of its working context

AI text drifts toward difficulty because the AI does not know the reader. If the request names no reader, the AI infers one from what it has. What it has is a long conversation, specialist documents and working labels. Writing to fit that context produces text the requester can follow and a first-time reader cannot.

Anthropic's published guidance on prompting Claude recommends being specific about the desired output and its constraints, and explaining the context or purpose behind an instruction. A reader group and a share deliver both the constraint and the purpose in one line.

A numerical target also gives the AI something to check its own draft against. Told "90% of ordinary readers", it can list words that group would not know, split overlong sentences and fill in skipped assumptions. An adjective gives it no such yardstick.

That is as far as the AI can go. It cannot see into a reader's head. The final check, whether the text reaches 90 readers, is done by people giving it to readers close to the target group.

07From tomorrow: name the reader, give the share as a number, and test with real readers

There are four steps.

1. Name the reader group in one sentence. For example, "This will be read by sales representatives in their first year." Use a role or a level of experience.
2. State the share as a number. Add "90 of every 100 of them should understand it without help." Do not write "make it easy to understand".
3. Have the AI list the stopping points. After the draft, ask it to list the terms that group could not follow and the sentences they would need to read twice, then fix them.
4. Give it to people close to the reader group. Two or three are enough. Ask them to restate the key points, and send any gaps back to the AI.

The first two steps add two lines to a request. The third can be left to the AI. Only the fourth needs a person. Skip it, and the AI's estimate gets treated as a result.