Your AI draft sounds right. Can you prove it?

The easiest sentence to miss in an AI-generated email is often the one that sounds most helpful: “We will have this resolved by Friday.” It has a clear owner, a deadline and a reassuring tone. It may also be a promise nobody made.

That is a useful place to begin learning AI fluency. The question is whether you can tell which parts of the output are supported, which are assumptions and which require somebody’s approval. A fluent user can make those distinctions before the draft becomes somebody else’s problem.

Separate the writing task from the decision

Consider an illustrative example. A small hotel wants to turn a maintenance update into a guest email. The source says a technician will inspect a faulty air-conditioning unit on Thursday. It says nothing about a repair date, compensation or moving the guest to another room.

AI can help make that update clear and considerate. The hotel still has to decide what it can promise. Ask for a draft based only on the supplied facts, with any missing decision left as a question for the manager. That instruction helps define the task. It is not proof that the resulting draft obeyed it.

Read it in three passes

First, inspect the claims. Underline dates, amounts, names, assurances and statements about what will happen. Compare each one with the original material. The model’s own list of supporting sentences is a navigation aid; read the source yourself. A convincing explanation can repeat the same error as the draft.

Second, inspect the omissions. A draft can contain no false statement and still mislead. Did it omit that the visit is an inspection? Did it remove an important uncertainty? Is the next update time clear? Ask what a reasonable reader would believe after reading the message, then compare that belief with what you actually know.

Third, inspect the authority. Highlight anything that commits time, money, access or somebody else’s work. Route those decisions to the person who owns them. Editing a sentence does not create permission to make its promise.

Practice the judgement, not just the prompt

The AI Fluency Framework developed by Rick Dakan and Joseph Feller with Anthropic groups the relevant skills into delegation, description, discernment and diligence. In everyday terms, you need to choose the work, explain it, judge the result and remain responsible for its use. Better instructions cover only part of that job.

For a short practice session, use a public notice or an invented example. Ask a colleague to add one unsupported commitment and remove one useful qualification from a draft. Try to find both. Then compare your reasoning. This is a learning exercise, not a validated measure of ability.

Leave a trace somebody else can use

Keep the approved source beside the final draft. Note the reviewer and the date. If the task recurs, keep a small checklist of the errors you actually encounter. “Check everything” is too vague to teach. “Confirm that inspection dates have not become completion dates” tells the next reviewer where to look.

For private or regulated information, begin with your organisation’s approved tools and data rules. Do not paste guest records, employee details or account information into a service simply because the prompt works on a fictional example.

Before you approve the wording, approve the claims.

The next time an AI draft looks finished, pick its strongest promise. Find the source for it and the person authorised to make it. If either is missing, the draft still needs work.