How Certain Is This AI Answer? Use Self-Calibration Before You Act

An AI assistant can write a confident client email, summarise a policy document, suggest a spreadsheet formula, or answer a question about a competitor in seconds. Problems arise when a polished answer is subtly wrong. Most professionals do not need an AI that is never uncertain. They need one that signals when an answer needs a second look.

Self-calibration draws out that signal. Rather than treating the first response as final, ask the model whether a specific answer is likely true. Better still, ask it to generate several independent candidate answers before it judges the proposed one. That review step can separate routine, high-confidence work from tasks that need a source check, a colleague’s input, or more context.

What It Is

Self-calibration asks an AI to assess the reliability of its own proposed answer in a clear True/False or multiple-choice format, rather than answering, “Are you sure?” For an open-ended question, the model produces an answer, reviews it as a distinct claim, and estimates whether it is correct. Generating several alternative answers first can strengthen that review. The research found that explicit judgment formats improved calibration and helped models distinguish correct from incorrect generated answers.

Before/After

Consider a common task: preparing a short briefing note before a sales call. You want the AI to identify a prospect’s recent strategic priorities from a set of notes without introducing unsupported claims.

Before prompt

Read the notes below and write three bullet points on the company’s strategic priorities for the next 12 months.

[Paste meeting notes, annual-report extracts, and news clips]

This often produces a useful-looking answer. It does not show which bullet the material directly supports and which reflects an inference from vague language. When source material is incomplete, the model may fill gaps without flagging them.

After prompt

Read the source notes below. First, draft three possible sets of strategic priorities independently. Keep each set concise and base it only on the supplied material.

Then produce one recommended set of three bullets. For each recommended bullet:

1. Assess the proposed answer as True or False.
2. Give a confidence estimate from 0–100%.
3. If confidence is below 80%, label the bullet Needs verification and state what information is missing.

Do not use outside knowledge. Do not include a claim merely because it sounds commercially plausible.

[Paste meeting notes, annual-report extracts, and news clips]

The confidence percentage alone is not the key change. Many systems will supply a number when asked, whether or not it carries much meaning. A more useful workflow generates alternatives, selects a proposed answer, then requires the model to judge it in an explicit format. In the research, comparison samples improved self-evaluation, particularly for short answers.

For recurring workflows, provide examples of well-calibrated output. One might show a claim with direct documentary support receiving high confidence. Another can show an attractive but weakly supported inference marked for verification. The research found that examples improved self-evaluation; without a prior format, models often gave unreliable, middle-of-the-road confidence estimates.

Formatting also matters. In the research, clear True/False decisions and lettered answer choices produced better calibration than loosely framed questions. When a task has a finite set of options—choosing a policy category, assigning a support ticket, or identifying an applicable contract clause—ask the model to select from labelled choices. Avoid “none of the above” where possible; it reduced both accuracy and the quality of confidence estimates in the tests.

When to Use It

• Before publishing a proposal, press release, account plan, or executive briefing, ask the model to check whether the supplied documents support each factual assertion. Send low-confidence claims to a human review queue.

• Use self-calibration to extract decisions, deadlines, owners, or risks from meeting transcripts and policy documents. Plausible errors in these details create follow-up work.

• When analysing interview notes, survey responses, competitor material, or market reports, require the model to distinguish direct evidence from inference. That keeps a tidy narrative from posing as a documented finding.

• For HR, finance, IT, or legal-adjacent requests, let the model answer only when an approved knowledge base supports the answer. Otherwise, have it identify the relevant policy and route the question to an owner.

• Treat low confidence as a cue to retrieve more information: search approved sources, attach a relevant document, or ask a clarifying question. In the underlying study, supplying relevant reference material substantially increased the model’s estimated likelihood of answering correctly in one example.

Teams should not treat self-calibration as an audit trail or a guarantee. The research found that models can be poorly calibrated on unfamiliar tasks and may find their own generated answers more plausible than they deserve. Extreme confidence also warrants caution: a highly confident answer can still be wrong. Use the technique to enforce a decision rule—for example, publish only claims with cited support, send medium-confidence items for review, and require external verification for low-confidence ones.

For teams building more advanced AI systems, the research also describes a trained “probability I know” predictor. Developers train it by repeatedly testing whether the model answers a question correctly. That approach can be useful, but it requires model fine-tuning and substantial evaluation data. Most professional users do not need to build it. A structured self-review prompt, comparison answers, and a clear escalation path provide a practical starting point.

Start with one repeatable task. Define what counts as evidence, set confidence thresholds, and decide how the team will handle uncertainty before the answer arrives. Track the results over time: which claims needed review, which confidence estimates proved misleading, and where better source material changed the outcome.

Source: Kadavath et al. (2022). Language Models (Mostly) Know What They Know. arXiv:2207.05221