AI for Journalists, Historians, and Investigators

Trace the claim to its origin and preserve the difference between repetition and corroboration.

A claim appears in a blog, a newsletter, and several social posts. An AI assistant summarizes them as “multiple reports confirm.” On inspection, every version comes from the same unsupported post.

The number of links increased. The amount of original evidence did not.

AI can help build timelines, compare accounts, and organize documents. Reporting and historical investigation still require attention to provenance: where information came from, what the source could know, and how later writers changed it.

Define the claim precisely

Replace a broad request such as “Investigate the project” with an assertion that can be checked. For example: “Did the council award a $2 million construction contract on August 3, or did it authorize a grant application?”

Those actions have different meanings. A vote to apply for funding does not by itself establish an award, contract, construction start, or completed project.

Identify why the question matters and what evidence could answer it. A contract claim points toward procurement records, an executed agreement, meeting records, and appropriate responses from the responsible body. A photograph of a building cannot establish the legal or financial action described.

For this lesson, all events concern an invented town called Pine Crossing. The dates and documents are synthetic training material, not real reporting about any community.

Build a source packet with provenance

Here is the fictional packet:

The source chain for the disputed award claim is R5 to R4 to R3. Neither R4 nor R5 adds independent evidence. R1, R2, and R6 are original institutional records about the action, though their common institutional origin also deserves recognition.

Do not count institutional documents as independent witnesses merely because they have different filenames. Their usefulness comes from what each records and how it can be checked.

Ask AI to preserve the chain

Trace each substantive claim in this packet to its earliest supplied evidence. Distinguish original records, firsthand statements, interpretation, and repetition. Build a dated timeline and flag contradictions, missing records, and unanswered questions. Do not infer a contract award from a budget estimate or treat linked repetition as independent confirmation.

A useful output identifies which document supports which action. It should not merge “application authorized” and “contract awarded” into a vague compromise such as “the project moved forward” when the disputed distinction is the central issue.

Keep document metadata separate from event dates. An agenda published August 1 concerns a meeting on another date. An approved record may be published after the meeting it describes. A historical document can also be copied or cataloged years after creation. Record both dates when they affect the chronology.

Write the finding within the evidence

An original example account based on the packet is:

Pine Crossing’s August 3 vote authorized submission of a grant application for a proposed learning center. The supplied meeting record lists an estimated project budget of $2 million and a 5–2 vote. A later social post described the action as a construction-contract award, and subsequent items repeated that claim without adding an original award record. The clerk’s supplied written response says the vote did not award a construction contract. The packet does not establish later funding, procurement, or construction decisions.

This account identifies what the records establish, explains how the inaccurate claim spread, and limits the conclusion to the supplied period. It does not claim that no contract could ever have been awarded later.

A useful headline would preserve the distinction: “Records show application vote, not contract award, in Pine Crossing example.” In an actual article, remove the training framing only after doing real reporting about real records. Never publish this fictional packet as a local news event.

Verify quotations, images, and names separately

A timeline can be accurate while a quotation is invented. Compare quoted words with the recording or original text and check the surrounding passage. If you summarize, label it as a summary rather than adding quotation marks.

For images, inspect source, date, location, caption, and whether the image depicts what the story claims. A genuine image of another project can still mislead. An AI-generated illustration should not be presented as documentary evidence.

For names, confirm identities carefully. People can share names, organizations can change names, and historical records may use variant spellings. Do not merge two individuals because a model considers their biographies similar. Keep uncertain matches marked for review.

OCR requires its own checks. A damaged scan can change a number or turn a place name into another word. Preserve the original image and identify uncertain readings. A neat transcription is not proof that the source was legible.

Handle absence and contradiction carefully

An incomplete packet with no contract does not prove there was no contract. It proves the packet lacks one. Here, R6 adds a direct statement about the August 3 action, which strengthens the narrow finding, but later developments still remain outside scope.

When sources disagree, ask whether they refer to the same event, period, definition, or stage. A budget authorization, grant award, contract execution, and payment can all involve the same project while occurring separately.

Do not manufacture balance by treating a well-supported record and an unsupported rumor as equally established. Describe the evidence each has. At the same time, do not conceal a material uncertainty because it makes the article less tidy.

Give people a fair opportunity to respond

The Society of Professional Journalists’ existing code emphasizes verification, original sources, context, correction, and seeking responses from subjects of criticism. These are professional ethical principles, not a substitute for applicable law or an organization’s editorial process. SPJ: Code of Ethics.

For a real story, prepare precise questions tied to the evidence and use authorized reporting methods. Preserve what was asked, when, and what response was received. “No response received by the stated deadline” is different from “refused to comment.” Silence does not establish guilt, motive, or agreement.

SPJ has also published proposed 2026 revisions, explicitly describing them as pending approval. Do not cite a proposed revision as an adopted rule. That status distinction is itself a useful source-checking lesson. SPJ: Proposed 2026 revisions.

Keep historical interpretation connected to the text

Historians face similar problems across longer periods. A later memoir may describe an event differently from a contemporary record. A term may have changed meaning. A modern summary can erase uncertainty about dates or authorship.

Use AI to compare passages and propose questions, then check the original editions, archival descriptions, and relevant specialist sources. Separate what a document states from what you infer about its author’s motives or the wider society.

A useful note might say, “The two sources use different dates; one appears to refer to authorization and the other to completion.” That is a hypothesis to test, not a reason to silently harmonize the dates.

Maintain a visible correction process

Keep a claim ledger with the source, draft wording, verification status, and final wording. If an error is found after publication, record what changed and why according to the publication’s correction practice. Repair related headlines, captions, summaries, and distribution copy where they repeat the error.

AI can locate repeated wording and draft a correction. A responsible editor should verify that the correction actually addresses the mistake and does not overstate the new evidence.

For students: investigate the source chain

Students can use R1–R6 to build a timeline and compare evidence strength under course rules. A history class can repeat the method with authorized archival excerpts; a journalism class can draft response questions and a correction note.

Do not invent interviews to make an assignment appear complete. Label unanswered questions and distinguish fictional practice records from original research. Showing why three links represent one unsupported source is a substantive research finding.

Practice: write an evidence-based account

Trace R3 through R5, compare the claim with R1, R2, and R6, and write a short account limited to the packet. Add a claim ledger, five follow-up questions, and a correction note for the fictional newsletter.

Completion check: The account distinguishes records, allegations, repetition, and interpretation. Quotes and dates are traceable, missing evidence is identified, and no claim of corroboration rests only on circular sourcing.

For a stretch exercise, add a later fictional contract record and revise the timeline without changing what the earlier vote did or did not establish.