What Severity Level Should We Assign to Wrong Hours Versus a False Safety Allegation?

Why use severity levels at all?

Because the queue of wrong AI answers never empties, and the loudest complaint is not always the most harmful. Incident management teams use severity levels for the same reason. Atlassian, for example, defines its top level as a critical incident with very high impact, and its lowest as a minor incident with low impact. It notes that a team without severity levels spends the first minutes of a major incident working out how important it is. A shared scale lets marketing, support, legal and communications agree in advance who does what for each kind of error.

What scale should you use?

Four levels are enough for most companies:

The response column is a starting point. Adjust it to your team's size, but keep the order.

How do you adjust for exposure?

Harm sets the starting level. Exposure then moves it, within limits:

• Raise one level if the error appears in more than half of runs across two or more assistants, for questions customers commonly ask.
• Raise one level if the error appears in answers to questions asked at the point of decision, such as "Is [business] open now?" or "Is [product] safe for children?"
• Lower one level if the error appears in fewer than 1 in 10 runs, in one assistant only, and only for unusual phrasing.
• Never lower an allegation of wrongdoing or a safety claim below S1, however rare it is. Rarity changes how often you retest, not the level. One screenshot shared widely can do the damage.

To measure exposure, run each affected question several times in each assistant and record the share of runs with the error. For example, wrong Sunday hours appearing in 8 of 12 runs across three assistants gives 67%. A false recall claim appearing in 2 of 12 runs gives 17%. The hours error is far more frequent, yet the recall claim still outranks it.

How do wrong hours and a false safety allegation compare?

Wrong hours become S2 when the business is one people rely on urgently, such as a pharmacy, clinic or emergency repair service. They also become S2 when the error keeps appearing in "open now" answers. How to fix hours at the source is covered in why AI shows incorrect business hours and how a local business can fix them.

What does an S1 response look like?

1. Record the answer fully: question, assistant, date, full text, cited sources. The steps are in how to document a false statement generated by an AI assistant.
2. Assign an owner and bring in legal and communications on the same day.
3. Establish the facts: is there any basis for the claim, such as a real recall of a different product, an old incident, or a similarly named company?
4. Find the source: cited pages, or the likely origin if none is cited.
5. Report to each platform, using the routes that exist for serious or legal issues as well as ordinary feedback. For ChatGPT, see how to report a false allegation through OpenAI's content reporting.
6. Correct the source where you can: ask publishers to correct or update pages, and publish an accurate statement where appropriate. How to do that without spreading the claim further is covered in how to correct a false claim without amplifying it in search.
7. Retest daily until the claim stops appearing, then weekly for a month.
8. Brief frontline staff so they can answer customers who mention it.

What does an S3 response look like?

Fix the sources you control: website, Business Profile, directories and feeds. Check the cited sources for stale data, and retest a week later. No legal review or public statement is needed. Many S3 issues close once your own sources are corrected; if one is still present at the next retest, check the cited source.

Who should own each level?

• S1: a senior owner, such as the head of communications or general counsel, with marketing supporting.
• S2: the marketing or digital lead, with legal or compliance consulted in regulated sectors.
• S3 and S4: the person who manages your website and profiles.

Write the owners into the scale, so nobody has to decide under pressure.

How should you track incidents?

Keep a log with one row per error: the date first seen, the question, assistants affected, severity, exposure rate, source, owner, actions and date closed. Review S1 and S2 incidents monthly, looking for patterns such as the same source causing several errors. The log also shows leadership that the scale is used consistently.

What might an illustrative case look like?

Consider an illustrative children's furniture brand that found two errors in the same week. Assistants gave wrong Saturday hours for its showroom in 7 of 12 runs. In 2 of 12 runs, answers to "Is [brand] cot safe?" said the cot had been recalled, which it had not. A different brand with a similar name had issued a recall. The hours error was logged as S3 and fixed through the website and Business Profile within four days. The recall claim was logged as S1: communications and legal were involved the same day, and each platform received a report. A short statement confirmed that no recall had been issued for the brand's products, linking to the official recall database. The recall claim stopped appearing within three weeks, and the hours error within two.

What do people ask about severity levels?

Should a false claim that appears only once still be S1?

Yes, if it alleges wrongdoing or a safety risk. Rarity affects how you monitor it, not how seriously you treat it.

Can one error change level over time?

Yes. Review severity when exposure changes or new facts emerge, and record the reason.

Do we need a separate scale for each assistant?

No. Use one scale, and record which assistants are affected in each incident.