Before an AI draft becomes a care update

What we found in the news

On October 19, 2023, WHO outlined considerations for AI in health (opens in a new tab), including clarity about intended use and opportunities for human intervention. Its earlier May statement on large language models (opens in a new tab) warned that their responses:

“may be completely incorrect or contain serious errors”

— World Health Organization, May 2023

These are broad cautions about health AI, not an evaluation of genus or of daily activity summaries. They prompt a narrower question for us: what should happen between an AI-assisted draft and an update another person is asked to trust?

Why this matters to us

Consider an afternoon at a community program. A participant watches some gardening, helps choose herbs, then leaves before planting begins. A hypothetical draft turns those notes into a sentence saying they planted herbs with the group.

The sentence sounds ordinary. It also changes what happened. A family member reading it might ask about an activity the person never did. Staff may then need to explain an avoidable discrepancy.

We care about the modest details of a day: the choice someone made, the company they enjoyed, the activity they decided to watch. There is plenty worth sharing without making the account fuller or tidier than the evidence allows.

Give the reviewer something to check against

A review needs more than a polished paragraph and an approval button. The person reviewing should be able to compare the proposed update with the relevant observations, understand where they came from, and notice gaps or contradictions.

In our example, the note about choosing herbs belongs in the account. Planting does not, unless another reliable observation supports it. When notes disagree, the team needs a way to resolve the difference or leave an uncertain detail out.

Decide who can do that work and allow time for it. A colleague who was not present may still review an update, but they need a clear route to someone who can answer a factual question. Requiring approval without providing context makes the signature less meaningful.

Leave room for the person’s own account

An observation records a moment from a particular point of view. Someone may have watched the gardening because they were interested, because they were waiting for a friend, or for a reason nobody asked about. A fluent sentence should not quietly select an explanation.

Where appropriate, invite the person to contribute to or correct the account. They might care more about choosing the mint than about how long they stayed outside. They may prefer that part of the day remain private.

Agree what kind of everyday information is suitable for a particular update and who may receive it. Family involvement does not remove the person’s choices or the organization’s responsibility for appropriate access. Review the intended audience alongside the wording.

Learn from the corrections

If reviewers repeatedly remove the same kind of overstatement, treat that as useful feedback about the drafting process. The answer may be a clearer instruction, better source notes, a narrower task, or a decision that this part should remain entirely human-written.

Use fictional or suitably de-identified examples when a wider team discusses those patterns. Personal information does not need to become training-room material simply because a sentence needs improving.

Look at the work the process actually creates. How long does checking take? Which errors are easy to miss? Do staff know when to stop and ask? Time spent correcting a draft belongs in the evaluation, alongside any time saved in writing it.

A team should also be able to pause the assisted step when it is not useful. Continuing to use it should follow from what the team can verify about the task.

What this means for the people you serve

An everyday update should remain recognizable to the person whose day it describes. People receiving support, families, and staff need a way to raise a concern about an account and reach someone responsible for correcting it. Trust has room for an honest correction.

The genus connection

Today, supported genus care-team workflows can prepare daily summaries from recorded information for staff to review, edit, and approve. Which modules and delivery steps are available must be confirmed for the organization.

The intended role is assistance with a bounded task. Staff remain responsible for checking the account and its audience before sharing. A well-written draft is a starting point for that judgment; it cannot establish that an activity happened or that a person wanted it shared.

One question for your next team meeting

Before we approve an assisted update, can we trace each meaningful claim to an observation and give the person a way to correct their story?