Can Generative AI Identify Missing or Impossible Values in Research Data?

Can AI Tell Which Research Data Values Are Missing or Impossible?

You have collected questionnaire responses from 600 university students. Most entries appear reasonable, but several deserve attention. One respondent reports being 250 years old. Another selects 8 on a five-point Likert scale. Some cells are blank, while others contain values such as 0, 99, and "N/A."

Could generative AI examine the dataset and identify these problems? And how would it distinguish a genuinely missing response from a valid zero, or an impossible value from an unusual but legitimate observation?

The challenge is not simply finding suspicious cells. It is determining what those cells mean according to the research instrument, coding scheme, and circumstances of data collection.

Yes, Provided the AI Uses Explicit Validation Rules

How AI Can Identify Missing and Impossible Research Data

What Counts as a Missing Value?

A missing value occurs when information expected for a particular observation is unavailable, unrecorded, or not provided. However, the appearance of missingness depends on how the dataset was designed and coded.

In a spreadsheet, missing information might appear as an empty cell, a software-specific missing marker, or a special code such as 99 or -9. Some datasets also use text labels such as "Not answered," "Refused," or "Not applicable."

These representations are not interchangeable. A blank answer caused by a participant refusing to respond differs from a question that was intentionally skipped because of questionnaire branching.

A chatbot cannot safely assume that 99 always means missing or that zero is an invalid response. These interpretations depend on the variable.

For example, zero is a perfectly valid number of research publications, but it would be invalid for an item explicitly coded from 1 to 5.

What Makes a Value Impossible?

An impossible value violates a known constraint governing the variable or observation. The constraint may come from the questionnaire, measurement system, study eligibility criteria, or a logical relationship between variables.

Consider three hypothetical observations:

• A five-point Likert item contains a response of 7.
• A variable representing the number of children contains -3.
• A recorded interview completion date precedes the participant's documented recruitment date, although the study procedure requires recruitment before interviewing.

Each violates a specified rule, assuming the variable definitions and records are correct.

However, impossibility should not be confused with improbability. A participant reporting 18 hours of daily internet use may warrant investigation, but that response is not inherently impossible. The observation could reflect overlapping device use, a misunderstanding of the question, or a genuine reporting pattern.

Researchers should therefore avoid defining impossible values solely through what appears typical. The separate problem of distinguishing data errors from legitimate unusual observations becomes important when an anomaly cannot be resolved through explicit rules.

How Can Generative AI Detect These Problems?

Generative AI can support validation in several ways. It may inspect a supplied codebook, suggest logical checks, generate Python or R scripts, or explain why a flagged observation violates a rule.

Some AI-enabled tools can execute code directly. Others only produce instructions that must be run in statistical software. Researchers should verify which operations were actually performed.

A useful approach is to ask AI to create a validation table containing the variable name, allowed values, missing-value conventions, and conditions that should trigger review.

For instance, a five-point academic self-efficacy scale may require every item response to be an integer between 1 and 5, except for documented missing values. A validation script can identify observations that violate this condition.

For deterministic checks, ordinary programming functions are often more reliable and easier to reproduce than repeatedly asking a language model to inspect individual cells. AI's principal contribution may be helping researchers translate the codebook into those explicit checks.

Can AI Identify Logical Inconsistencies Across Variables?

Yes. Some problems only become visible when two or more variables are considered together.

Suppose a questionnaire asks whether a participant has ever used an AI writing tool. Respondents who answer "No" are instructed to skip a subsequent question about frequency of use.

If a respondent selects "No" but reports using the tool daily, the record contains a logical inconsistency worth investigating.

Possible explanations include incorrect data entry, confusing questionnaire wording, faulty branching logic, or a participant changing their interpretation between questions. The inconsistency does not, by itself, reveal which response is wrong.

AI can help formulate cross-variable rules, but the researcher must establish which relationships are genuinely required by the study design.

Can AI Determine Why Data Are Missing?

AI can summarize observed patterns of missingness and help generate statistical diagnostics. However, the observed dataset does not necessarily reveal why values are missing.

Statistical literature commonly distinguishes three missing-data mechanisms:

• Missing completely at random (MCAR): The probability of missingness does not depend on observed or unobserved data values.
• Missing at random (MAR): Given the observed information used in the missingness model, missingness does not additionally depend on the missing values themselves.
• Missing not at random (MNAR): Missingness still depends on unobserved information, potentially including the missing values, after accounting for the observed information.

These are assumptions about the missing-data process, not labels that a chatbot can reliably assign simply by examining blank cells. In particular, MAR and MNAR generally cannot be conclusively distinguished from observed data alone.

Rubin's foundational work on inference with missing data provides the theoretical basis for these distinctions. Little's work on testing MCAR offers a related statistical diagnostic, although failing to reject MCAR does not prove that the mechanism is MCAR.

Identifying missing values is therefore separate from deciding how they should be analyzed or whether they should be estimated.

Should AI Automatically Replace Missing or Impossible Values?

Detection does not establish the correct replacement.

If a participant's age is recorded as 250, the researcher cannot simply change it to 25 because 25 looks plausible. The original value might have been 20, 50, or something else entirely.

Similarly, replacing every missing Likert response with the midpoint may distort distributions, relationships among variables, and estimates of uncertainty.

Where original records establish a transcription error, a documented correction may be appropriate. Where the intended value cannot be recovered, researchers should preserve the uncertainty and use a defensible missing-data procedure.

The choice of whether to estimate missing research values involves additional statistical assumptions. It should not be bundled into an unrestricted data-cleaning command.

How Should Flagged Values Be Documented?

A validation report should make it possible to locate each flagged observation and understand why it was flagged.

Useful fields include the record identifier, variable name, original value, violated rule, investigation result, approved action, and status of the correction.

For example, a flag might state: "Record P017, variable SE04, value 7, permitted range 1–5, unresolved pending source verification."

This is more informative than a chatbot's general statement that the dataset contains invalid responses. It also supports an audit trail of AI-assisted cleaning if changes are subsequently made.

Checking Missing and Impossible Values in an Academic Self-Efficacy Survey

The researcher confirms that P014's original questionnaire contains a response of 3, which was incorrectly entered as 8. The cleaned dataset can therefore record 3, with the correction documented.

For P028, the original age cannot be verified. The researcher does not replace 250 with a guessed age. Instead, the value is flagged as unresolved and handled according to the study's documented missing-data rules.

P093 requires no correction. A low self-efficacy response is entirely legitimate within the instrument's permitted range.

Interpretation: The validation procedure successfully distinguishes a confirmed transcription error, an unresolved anomaly, an ordinary missing response, and a valid observation. That distinction matters more than simply reducing the number of flagged cells.

Misconceptions About Missing and Impossible Values

How Should You Use AI to Validate Missing and Impossible Values?

Begin with the codebook rather than the chatbot. Establish what each variable measures, which values are permitted, how missing responses are represented, and which logical relationships should hold across variables.

Then use AI to help translate those definitions into explicit validation procedures.

For large datasets, ask the AI to generate a validation script that produces a separate report of suspected problems without changing the data. Review that report before authorizing any transformations.

When corrections are eventually made, verify the resulting changes against the original dataset and the approved rules.

Before Accepting AI-Generated Data Validation Results

Frequently Asked Questions About AI Data Validation

Detect Problems With AI, but Establish Corrections From Evidence

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.