Itai Liptz.
← Articles

Work & learning

Practice a responsible AI drafting routine

Work through a fictional drafting task with explicit data boundaries, source checks, human review, and a record of the final decision.

Updated
A person reading notes on a laptop beside an open notebook

Before using an AI tool for a piece of work, decide what you will ask it to do and how you will judge the result. A polished draft still needs a person who can explain its sources, limitations, and intended use.

This exercise follows a fictional colleague drafting a workshop invitation. It provides a personal routine that can sit within an organization’s approved AI policy. Your employer’s requirements and permitted tools determine whether the exercise is appropriate for a real task.

Write the task in a sentence

The colleague wants a clearer invitation that tells staff who the workshop is for, when it takes place, and how to register. They have approved event details and a previous public invitation.

They decide to ask for two possible structures, then write and check the final message themselves. This makes the requested assistance specific. It also gives them criteria for reviewing the output.

Before opening a tool, identify who will read the message and what an error could cause. An incorrect date or inaccessible registration link would matter even in this modest example.

Establish the information boundary

List what the tool needs and what it does not need. The public workshop title, published date, and general audience description may be enough. A staff mailing list, private accessibility request, or internal discussion is unnecessary for this draft.

Use only information you are authorized to enter into the approved tool. When that authorization is unclear, resolve it with the responsible person before proceeding. Replacing names with placeholders can help a practice exercise, but it does not by itself establish permission for a real use case.

Record the boundary beside your task description so a later reviewer can understand the choice.

Check the output against original material

Suppose a suggested invitation adds a promise that attendees will receive a certification. The approved event brief contains no such promise. Remove it and check the rest of the draft for additional details that the source does not support.

Compare dates, names, eligibility, and links directly with the original event information. Read source material rather than relying on a tool’s assurance that it has checked its own work.

NIST offers broader voluntary guidance for considering risks throughout AI use. Organizations can use that guidance to support a more comprehensive assessment of the task, information, responsibilities, and possible consequences. NIST AI Risk Management Framework.

Keep responsibility attached to the release

Choose who can approve the invitation. If another colleague owns event details, ask them to verify those details. If you own the final communication, confirm that the finished message says what you intend before sending it through the usual process.

A useful review note states what assistance was used, which sources were checked, and who approved the message. Follow your organization’s disclosure and recordkeeping requirements.

Learn from the correction

After the task, note the invented certification promise and how you detected it. Decide whether the next exercise needs a narrower instruction or an earlier check.

Also review the effort involved. If checking and correcting the draft took longer than writing it directly, that result belongs in the evaluation. The AI tool assessment exercise develops a comparison that includes review time and error correction.