The answer ends with a small sentence: “I may be wrong.” Above it sits a confident recommendation. You are about to send a message, change a setting or repeat a claim. Does that final sentence change anything you do?
Uncertainty can be a useful signal. It can also become wallpaper—always present, rarely noticed. The question is not whether an AI sounds modest. It is what remains unresolved, and what you should check before taking the next step.
A warning is not a measurement
“Probably,” “almost certainly” and “I am 90% confident” do not tell you how that confidence was established. Was the system tested on comparable questions? Does the score come from an evaluated prediction method, or is it simply part of the generated reply? Without that context, a percentage can look more informative than it is.
The NIST Generative AI Profile, published in July 2024, describes confabulation: generated content can be false while being presented confidently. It also cautions against generalizing system capabilities from narrow or anecdotal assessments. This is established guidance, not a new announcement about every model available today.
Neither confidence nor hesitation settles the claim.
Ask what kind of uncertainty you face
Imagine asking an assistant whether a community venue will be open on Saturday. It produces an opening time, then adds a disclaimer. Three different problems may be hiding in that answer.
The source may be missing: no official schedule was consulted. The information may be old: last month’s schedule is real but no longer decisive. Or the question may be incomplete: which venue, which Saturday, and does a special event change public access?
Each gap calls for a different response. Find the official page. Check its date and any closure notice. Specify the location and day. “Be more certain” is a weaker request than “show the current official schedule for this venue and date.”
A useful answer does not need to resolve everything. It needs to keep the unresolved part from disappearing inside a polished sentence.
Make the next step smaller
Suppose you are drafting an invitation based on that opening time. You can prepare the wording without sending it. You can leave the hour visibly unconfirmed. You can contact the venue before asking other people to travel.
These are not equivalent actions. Drafting preserves options; announcing a time creates expectations. The same uncertain fact can be tolerable for an internal sketch and inadequate for a public commitment.
This is a practical recommendation, not a guarantee: while evidence is incomplete, prefer a bounded, reversible step over an action whose consequences depend on the disputed detail. Sometimes the appropriate step is to wait. Sometimes it is to continue the part that does not depend on the answer.
Human review needs something to work with
“A person will check it” sounds reassuring. But what will that person inspect, and what knowledge can they add?
In their 2020 study of AI-assisted decision making, Yunfeng Zhang, Q. Vera Liao and Rachel K. E. Bellamy found that displaying confidence could help calibrate trust, but better trust calibration alone did not ensure better joint decisions. Their experiments concerned a specific prediction task, not today’s conversational assistants. They are a reason to ask what a reviewer contributes—not proof that a warning label solves the problem.
For the venue example, useful review means comparing the invitation with the current schedule, resolving the special-event question and correcting the draft. Merely rereading the assistant’s explanation is not an independent check.
Leave room for an answer to change
Keep a short record of the consequential detail: what is known, what is missing, which source could resolve it and which action is waiting. If you need to inspect a cited document, our guide to verifying AI-generated references separates finding a source from checking what it supports. The distinction between facts and interpretations helps with the next layer: what you conclude from it.
Do not turn every ordinary task into an endless investigation. Match the check to the consequence, and stop when the relevant question has been answered well enough for that action. A corrected official notice can outweigh a confident old summary; repeating the summary cannot make it current.
The useful “I don’t know” is not an ornament. It is a junction: it shows what cannot yet carry your decision, and leaves you free to choose a better-supported path.
Before you act, name the missing piece.
M.R. Orion
