The Timeline Check: How to Stop AI-Assisted Writing From Reversing Cause and Effect

A clear timeline is necessary for a causal claim, but it is not enough. The timeline check orders cited events, preserves uncertainty, and stops an AI-assisted draft from turning “after” into “because.”

A Smooth Story Can Run Backward

AI-assisted drafts are very good at turning scattered notes into a continuous story. That fluency becomes risky when the notes contain several kinds of dates: when an event happened, when someone noticed it, when a record was created, when a source was published, and when a later author interpreted the result.

A model may compress those dates into one neat sequence. It can say that a policy produced a result even though the first signs of the result appeared before the policy. It can place a public announcement at the start of a process that had been running privately for months. It can treat the publication date of a retrospective report as though it were the date of the events inside it.

The sentences may remain factually recognizable while the causal direction changes.

The timeline check is an editing method for catching that error. It asks the writer to reconstruct the order and status of events before deciding what relationship the prose can support. Its central rule is simple: a proposed cause must come before its proposed effect, but coming first does not prove that it caused what followed.

Chronology is a gate, not a verdict. If the order is impossible, the causal claim fails. If the order is possible, the claim still has to survive questions about mechanism, delay, selection, measurement, and alternative explanations.

Build An Event Ledger Before Writing The Narrative

Do not begin with a paragraph. Begin with an event ledger. Give every consequential event one row and keep source language separate from editorial interpretation.

For each row, record:

• Event: What was observed, changed, announced, measured, or alleged?
• Event time: When did it actually happen?
• Precision: Is that time exact, bounded by a range, or only estimated?
• Time type: Is it an event, onset, collection, upload, publication, update, or retrieval time?
• Source: Which log, document, interview, dataset, or firsthand record supports it?
• Evidence status: Is it observed directly, reported by someone else, or inferred by the draft?
• Evidence timing: Was the record created contemporaneously or reconstructed later?
• Notes: Which time zone, unit, definition, exclusions, or corrections affect interpretation?

Keep unknown times unknown. “Between Tuesday afternoon and Thursday morning” is more accurate than inventing Wednesday at noon. If a source says “about six weeks later,” record the estimate and its anchor rather than converting it into a false calendar date.

Then sort by event time, not by the order in which the sources were found. A search result, press release, or later case study may be the first document the editor sees while describing something that occurred much earlier.

Separate Event Date From Publication Date

A page can contain several clocks. An article published on June 12 may describe a rollout on May 3, quote an interview conducted on May 20, analyze data collected from April through May, and include a correction added on June 14. “According to a June 12 article” identifies the source. It does not date every event inside it.

Record publication and update dates because they help establish what information was publicly available at a given point. Record event and collection dates separately because they establish the sequence being discussed. If a source does not distinguish them, the draft should not quietly choose one.

This distinction matters most with retrospective sources. A year-end report may provide the best summary of a January decision, but its December publication date cannot move the decision to December. Conversely, a later recollection may add context that was not known in January. The ledger should preserve both the event date and the date from which the interpretation comes.

Normalize Time Zones, Calendars, And Units

Two timestamps cannot be ordered safely until they use a comparable frame. “9:10” in Tel Aviv and “8:45” in London may refer to the opposite order from what the clock faces suggest. A midnight UTC log may appear under the previous or next local calendar date. Daylight-saving changes can produce repeated or missing local times.

Keep the original timestamp, time-zone identifier or offset, and normalized comparison value. The W3C guidance on working with time and time zones distinguishes instants on a timeline from local or “floating” times that do not identify one unique instant. That is a useful editorial distinction: a calendar invitation with a zone can be compared as an instant; “Tuesday morning” cannot be made equally precise without more evidence.

Normalize duration units too. Do not compare 48 hours with “two business days” as though they are interchangeable. Separate calendar months from fixed day counts, averages from medians, and a reporting week from a rolling seven-day window. Preserve the original unit beside any conversion and state the assumption that made the conversion possible.

Label Observation, Mechanism, And Inference

A chronological paragraph often hides three different kinds of statement:

• Observation: A documented event or measurement, such as “error reports declined from Monday through Thursday.”
• Mechanism: The process proposed to connect events, such as “the routing change sent fewer requests to the failing service.”
• Inference: The conclusion drawn from the pattern, such as “the routing change reduced errors.”

Put those labels in the working draft even if they disappear from the final prose. The observation may be well supported while the mechanism remains plausible but untested. The inference may be reasonable enough to investigate without being strong enough to publish as fact.

A model tends to bridge gaps with connective language: “therefore,” “leading to,” “as a result,” or “which caused.” Search for every such bridge. Ask what evidence supports the connection itself, not only the statements on either side.

Mark Contemporaneous And Retrospective Evidence

A server log written during an incident, a memo approved before a launch, and a survey response recorded that week are contemporaneous evidence. A quarterly review, later interview, and postmortem are retrospective. Neither category is automatically correct or incorrect.

Contemporaneous records can be incomplete, automated, or based on assumptions that later prove wrong. Retrospective accounts can benefit from broader evidence while also being shaped by memory, hindsight, or the outcome that became known. Tagging the timing helps the editor explain why two credible sources disagree.

When a later source changes the interpretation of an earlier event, write that change explicitly: “The team initially attributed the outage to traffic; a postmortem published two weeks later identified a configuration error.” Do not rewrite the earlier record as if everyone knew the final explanation at the time.

Test The Main Alternatives To A Causal Story

Once the ledger is ordered, challenge the most appealing explanation. At minimum, test these four alternatives:

• Reverse causation: Did the supposed outcome actually trigger the supposed cause? A support backlog may prompt a staffing change, rather than the staffing change creating the backlog.
• Lag: Could the proposed effect occur only after a delay? An immediate change may be too early to result from a process that requires days or months.
• Selection: Did the people, records, or periods included after the event differ from those included before it?
• Common cause: Could a third event have influenced both? A seasonal demand spike might trigger a product change and raise support volume at the same time.

Also look for measurement changes. A metric can jump because logging improved, definitions changed, a dashboard refreshed, or reporting became mandatory. The event is real, but the apparent trend may belong to the measurement system rather than the world it describes.

Causal inference asks more than whether one event preceded another. The open textbook Causal Inference: What If frames causal questions around contrasts between outcomes under different possible actions. An ordinary content review rarely has that complete evidence. It should therefore keep observational sequence, proposed mechanism, and causal conclusion at different levels of certainty.

Use A Before, During, And After Table

A three-window table makes missing baselines and overlapping events visible:

Window	Capture	Questions	Safe Output
Before	Baseline trend, prior decisions, selection rules, expected variation	Was the outcome already changing? Was the intervention responding to it?	Describe the baseline without implying a cause.
During	Exact rollout stages, exposure, concurrent events, measurement changes	Who or what was actually affected, and when?	Name overlap, partial exposure, and unresolved timing.
After	Outcome timing, plausible lag, follow-up period, reversals	Did the change persist, and is the window long enough?	Report the observed association within its window.

“Before” should be long enough to reveal an existing trend. “During” should reflect staged rollouts rather than a single ceremonial launch date. “After” should match the speed at which the proposed mechanism could reasonably operate.

Example: A Product Rollout That Appears To Cause Churn

An AI-assisted launch review says, “The new checkout caused subscription cancellations to rise.” The evidence list appears orderly: checkout version 2 launched Monday, cancellation requests increased Tuesday, and the weekly report was published Friday.

The ledger changes the story. A billing email with incorrect renewal dates went to a subset of customers on Sunday night. The checkout flag opened to 10 percent of traffic Monday morning and reached 100 percent only Thursday. Most Tuesday cancellation requests came from customers who received the email but never used the new checkout. A support dashboard also changed from counting conversations to counting individual messages on Monday.

The original sentence puts the visible launch before the reported increase, but the exposure and measurement details do not support the claimed mechanism. The billing email is a common-cause candidate, selection differs across customer groups, and the dashboard creates an artificial jump in support volume.

A defensible revision is narrower: “Cancellation requests rose during the checkout rollout week, especially among customers who received an incorrect renewal email. The available records do not isolate the checkout’s effect.” The sentence keeps the operational signal worth investigating without assigning a cause the timeline cannot establish.

Example: A Workplace Health Claim

Suppose an internal article says, “The new meeting policy reduced employee headaches within two weeks.” The policy began July 1. A July 15 pulse survey contains fewer headache reports than a survey sent in June.

The event ledger should ask when symptoms were experienced, not only when responses were submitted; whether the same employees answered both surveys; whether the questions, reminder schedule, and response options stayed the same; whether workload, leave, weather, or office attendance changed; and whether two weeks is an appropriate observation window for the proposed explanation.

The CDC Field Epidemiology Manual notes that exposure must precede a health outcome for a causal interpretation and that report date may be only a surrogate when onset time is unavailable. Its guidance also illustrates why ordering events generates hypotheses rather than proving them. This editorial example is not medical advice or a clinical conclusion.

A responsible internal summary might say, “Fewer respondents reported headaches in the first pulse survey after the meeting policy changed. Differences in respondents and timing mean the survey cannot establish why reports fell.” That wording preserves the observation and the uncertainty.

Ask AI To Order Facts, Not Supply Missing Time

AI can help organize the ledger when its role is deliberately narrow. Give it the cited excerpts and use a prompt such as:

“Extract each event stated in these sources. For every event, return the original date or time wording, normalized time only when the source provides enough information, time-zone or unit, precision, source URL, and whether the record is contemporaneous or retrospective. Separate observed facts from mechanisms and inferences. Sort only comparable timestamps. Mark missing, conflicting, or ambiguous times as unresolved. Do not invent timestamps, infer an unstated time zone, or add causal language.”

Then run a second pass:

“List every phrase in the draft that implies cause, response, prevention, increase, reduction, or sequence. For each phrase, identify the proposed cause, proposed effect, evidence for their order, plausible lag, and any reverse-causation, selection, measurement, or common-cause alternative. Do not decide that an association is causal.”

Return to the original sources for every correction. The model can expose inconsistencies; it cannot convert an absent timestamp into evidence.

Write Uncertainty Into The Sentence

Uncertainty should sit beside the claim it limits. Do not hide it in a final disclaimer after a page of confident causal prose.

Useful patterns include:

• “The increase began before the public announcement, so the announcement cannot explain the start of the trend.”
• “The records place both events in the same week but do not establish which happened first.”
• “The decline followed the rollout, although a concurrent pricing change offers another explanation.”
• “Logs show an immediate association; the proposed mechanism would require a longer follow-up period.”
• “Retrospective interviews support this interpretation, but contemporaneous records do not document it.”

The CDC’s guidance on analyzing and interpreting data advises considering chance, selection, information bias, confounding, and error before treating an association as causal. Editors can apply the same discipline without turning an article into a formal study.

The Final Timeline Check

Before publishing, confirm that:

• every consequential event has its own ledger row and cited source;
• event, observation, collection, publication, update, and retrieval dates remain distinct;
• exact, estimated, ranged, and unknown times are labeled honestly;
• time zones, calendar boundaries, durations, and units are normalized without losing originals;
• events are sorted by event time rather than source discovery order;
• observation, mechanism, and inference are separated;
• contemporaneous and retrospective evidence are identified;
• the proposed cause precedes the proposed effect by a plausible interval;
• reverse causation, lag, selection, measurement change, and common causes were tested;
• the before, during, and after windows use comparable populations and definitions;
• uncertainty appears beside the claim it limits; and
• AI organized only cited facts and did not invent timestamps or causal bridges.

A timeline can disprove an impossible story. It cannot prove a causal one by itself.

Put the events in order, show what each clock means, and let the prose stop exactly where the evidence stops.