How to Evaluate AI Answers When Evidence Is Missing

An AI response says, “I'm 90% sure.” Another says, “The source doesn't establish that.” Which deserves the better evaluation?

Check the evidence before choosing. A confident guess may exceed what the source supports, while a blanket refusal may leave answerable parts untouched. A useful response can state the known facts and identify the specific information it cannot determine.

This guide shows how to assess those boundaries without automatically rewarding certainty or caution.

What recent research contributes

An October 22, 2026 study, Cost-Aware Abstention for LLM-Based IT Ticket Classification, evaluates allowing a classifier to defer cases under specified costs. Its findings depend on one ticket dataset and illustrative assumptions, so they do not supply a universal threshold for declining to answer. The useful evaluation question is whether abstaining fits the task and its consequences. Read the original study.

Earlier research offers a close example for source-based questions. SQuAD 2.0 includes questions that a supplied passage cannot answer, requiring systems to recognise that gap alongside answering supported questions. It is an extractive reading-comprehension benchmark, rather than a general rule for every conversation. Read the original paper.

The method below is an original practice aid, not a scoring system validated by either study.

A five-step evidence-boundary check

1. Identify the allowed information

Read the evaluation instructions. Does the task restrict the answer to a supplied passage, allow outside knowledge, or require current verification?

A fact might be true somewhere else but unsupported in a source-only task. Conversely, an open research task may allow the assistant to verify a missing fact. Judge the response under the actual rules, including any tools it was permitted to use.

2. Break the question into answerable parts

List each requested fact separately. For each one, mark whether the available material establishes an answer, contradicts a proposed answer, or leaves it unresolved.

“Not established” and “false” are different findings. A biography that omits someone's birthplace does not prove they were born elsewhere. It simply cannot support a birthplace answer on its own.

This distinction also helps identify a useful partial answer: uncertainty about one detail need not erase everything the source does establish.

3. Inspect the substance behind confidence language

Words such as “definitely,” “probably,” and “I think” describe how an answer is presented. They do not supply the missing evidence.

A hedge attached to an unsupported birthplace still leaves an unsupported birthplace. A precise percentage also needs a basis; do not treat it as proof of correctness. This guide assesses individual responses, not statistical confidence calibration across a labelled test set.

Check the concrete conclusion or recommendation, including any summary at the end. A cautious introduction does not resolve an unsupported final claim.

4. Assess the scope of the uncertainty

Look for a clear statement of what is missing and what can still be answered. “The profile names her exhibition city but does not give her birthplace” is more informative than “There isn't enough information.”

Do not require an unnecessary follow-up question when the task is simply to report what a passage supports. Where further investigation is allowed, a useful response may identify the source or detail needed next without claiming it has already checked it.

5. Apply the rubric to usefulness as well as restraint

An answer that avoids guessing can still fail by withholding an explicitly supported fact. Compare accuracy, evidence use, completeness, and instruction following as the assigned rubric requires.

Do not invent an automatic bonus for refusing or a fixed penalty for uncertainty. Explain the particular consequence: the response guessed a missing fact, or it failed to supply an answer the material already contained.

Worked example: an incomplete artist profile

Fictional practice case. The profile and all responses below were created for this guide.

Profile: “Leena Rao is an illustrator currently based in Port Mere. Her first solo exhibition opened in Brighton in 2012. She works mainly in ink.”

Task: “Using only this profile, where was Leena born, and where and when did her first solo exhibition take place?”

Response A: “Leena was born in Brighton—I'm 90% sure. Her first solo exhibition was also in Brighton in 2012.”

Response B: “The profile does not state Leena's birthplace, so I cannot determine it from this source. Her first solo exhibition opened in Brighton in 2012.”

Response C: “I can't answer this with the information provided.”

Check each requested fact:

Evaluation: B best meets the task. It answers the supported parts and identifies the precise gap. A gets the exhibition details right but guesses the birthplace. Its confidence percentage provides no evidence for that guess. C avoids the guess but fails to provide the supported city and year.

A suitable justification is: “B keeps the birthplace unresolved while supplying Brighton and 2012. A adds an unsupported birthplace, and C withholds answerable exhibition details.”

This evaluates the responses against the profile. It does not establish Leena's birthplace beyond this fictional source.

Short exercise: is refusal justified here?

Fictional source: “Tickets for the town illustration fair cost £12 for adults and £8 for students. The notice does not list prices for other visitor categories.”

Question: “Using this notice, how much does a student ticket cost?”

AI response: “I can't determine that because the notice does not contain a complete price list.”

Is the response appropriately cautious?

Explained answer: No. The requested student price is explicitly £8. Missing prices for other categories do not prevent answering this question. A suitable answer is: “A student ticket costs £8, according to the notice.” If the question instead asked for an unlisted category, the answer would need to identify that gap.

Common mistakes to avoid

• Accepting a guess because it sounds cautious. Check the claim's support.
• Rewarding every refusal. Look for answerable parts the response omitted.
• Calling an unsupported statement false. Use the evidence to distinguish absence from contradiction.
• Requiring certainty about everything. Judge each requested detail separately.

Practise supporting your judgment

Try AtomicEval's first challenge to practise comparing responses and explaining your choice with evidence. It has a five-minute target and can be started without signing up. A free profile is required to view and save your result.

When you're ready to continue preparing, create a free AtomicEval profile.

Sources

• Gupta, Cost-Aware Abstention for LLM-Based IT Ticket Classification, submitted October 22, 2026.
• Rajpurkar, Jia, and Liang, Know What You Don't Know: Unanswerable Questions for SQuAD, ACL 2018; arXiv submission June 11, 2018.