AI-assisted practical guide. Examples are hypothetical; these are proposed editorial methods, not reported research results.
Ambiguous dates should remain visible during a CSV review. A value such as 05/06/2023 can support more than one interpretation when the source convention is unknown. AI can propose a review list, but the person responsible for the data must decide which evidence is sufficient for conversion. Keep the original values so a proposed change can be checked and reversed.
Specify the allowed interpretations
State what is known about the source. For a limited exercise, you might allow only month/day/year and day/month/year and require a four-digit year. That restriction is an explicit assumption, not a universal rule for date data. Use a prompt such as: “Review these rows under the two supplied conventions. Preserve every original value. Propose a normalized date only when exactly one allowed interpretation is valid. Otherwise mark the row unresolved and explain what source information is missing.”
Request a separate proposed-value field rather than an overwritten original. Do not let a pattern elsewhere in the column settle an ambiguous row when mixed sources are possible. A format used by most records is not proof of the format used by this record.
Hypothetical example
Imagine the allowed conventions are exactly those above. Row A contains 12/01/2021, row B contains 05/22/2021, and row C contains 03/04/2021. Rows A and C remain unresolved because their day and month positions can both be interpreted validly. Under these stated assumptions, row B has the proposal 2021-05-22 because 22 cannot occupy the month position.
The reviewer checks that row A has not silently become December 1. They also verify that no record was dropped or reordered. If a source document later confirms a convention for a particular row, record that evidence beside the approved change rather than treating the model's confidence as confirmation.
Check values and structure separately
Compare every proposed conversion with the original and the permitted formats. Check calendar validity as well as the range of month numbers; a plausible-looking string can still represent an invalid date. Preserve identifiers, row counts and unrelated fields while reviewing the CSV structure.
Leave unresolved rows in the handoff with a specific question for the source owner. Missing values, invalid values and ambiguous values need separate explanations. The finished result is a reviewable set of proposals and unresolved questions, not an apparently complete column produced by guessing.