The Anti-Hallucination Habit: Keep a Source Box Next to Every Draft
Use the anti-hallucination habit to build a source box, expose evidence gaps and verify factual claims before publishing AI-assisted drafts.
A source box gives an AI-assisted draft a clear factual boundary before the writing begins.
The anti-hallucination habit is simple: place an approved set of facts, figures, names, dates, quotations and links beside every draft, then tell the AI to use only that material for factual claims. The box does not make the model infallible. It makes unsupported additions easier to prevent, flag and find during review. That turns factual control into a visible part of the writing process.
The Short Version
- Build a small, approved source box before asking AI to draft.
- Require every factual claim to stay within the supplied evidence.
- Tell the AI to flag gaps instead of inventing names, figures, dates or quotations.
- Check each material claim against its supporting passage before publication.
Why Fluent Drafts Still Need Boundaries
Language models produce text by predicting likely words from learned statistical patterns. That process can create coherent, confident prose, but fluency does not guarantee factual accuracy. A model may produce material that fails to reflect the supplied context or conflicts with established knowledge. Hallucinations can therefore look plausible enough to pass an ordinary read-through.
The academic paper Beyond Misinformation: A Conceptual Framework for Studying AI Hallucinations in (Science) Communication distinguishes errors of faithfulness from errors of factualness. In the first case, the answer does not accurately reflect the source or input. In the second, it does not align with established real-world knowledge. That distinction matters because a draft can fail even when every sentence sounds polished.
The risk is especially easy to miss when an invented detail has the right shape. A false statistic can resemble a real one, while a fabricated reference can follow the usual pattern for titles, authors and dates. General proofreading often catches clumsy wording but may miss a small, fluent factual error. The answer is to give the drafting process a firm evidence boundary and then review the result at claim level.
This habit fits naturally after a clear writing brief. A useful brief defines the audience, purpose and format, while a source box defines what the draft is allowed to say as fact. These controls solve different problems and work well together: the brief directs the writing, while the source box limits its factual material.
The Anti-Hallucination Habit: Build the Source Box
A source box is a compact collection of approved evidence for one draft. It can sit in the prompt, in an attached document or in a clearly labelled section of a working note. The format matters less than the boundary: the model should know which material is approved and which gaps remain open. Keep the box small enough for a human editor to understand and inspect.
Start with the facts that the finished piece genuinely needs. Record exact names, job titles, dates, quantities, units, quotations and source links. Beside each item, keep the supporting passage or a precise note showing what the source establishes. Do not reduce a qualified statement to a stronger shorthand that changes its meaning.
For example, “may qualify” should not become “qualifies”, and an observed association should not become proof of causation. A source box preserves those distinctions when it stores the wording and context, not merely a loose conclusion. It also helps separate confirmed information from an editor’s assumption. That separation is central to the anti-hallucination habit.
A practical source-box entry might contain five fields: the proposed claim, the exact evidence, the source link, any necessary qualification and its approval status. The evidence should support the whole claim, including its number, time period, population and degree of certainty. If it supports only part of the sentence, split the sentence or remove the unsupported part. Avoid treating the mere presence of a citation as proof.
The source box should also state what the AI must do when evidence is missing. A useful instruction is: “Use only the approved material below for factual claims. Attach the relevant source label to each factual assertion. If the evidence does not support a name, date, number, quotation, cause or obligation, write [EVIDENCE NEEDED] instead of completing the gap.” In a publication workflow, replace that drafting marker before release rather than allowing it into reader-facing copy.
This five-field format is practical editorial guidance, not a guarantee that every error will be prevented. Source material may be unreliable or conflicting, and a model may still produce a conclusion that goes beyond accurate passages. The editor therefore remains responsible for checking whether the final wording matches the evidence.
What Belongs in the Box
Include only material that bears directly on the planned article. A short product update may need an approved product name, release date, price, availability, quoted comment and links to the relevant documents. A results summary may need the reporting period, currency, revenue, profit measure, comparison period and definitions for any adjusted figures. Each item should be understandable without guessing what its label means.
Numbers deserve extra care because a digit can look authoritative even when its basis is unclear. Store the unit, period and comparison alongside the value. If a figure is an estimate, range or adjusted measure, retain that description. The discussion of adjusted profit and add-backs illustrates why definitions and boundaries can matter as much as the headline number.
Names and quotations should be copied accurately from their supporting material. Keep the speaker’s role and the context required to interpret the words. Do not ask the model to recall a surname, complete a quotation or infer a job title from general knowledge. A blank field is safer and more useful than a plausible invention.
Dates also need labels. “Announced on”, “effective from”, “reported for the year ended” and “accessed on” describe different things. A source box that records only a bare date invites accidental substitution. Clear labels allow the drafter and reviewer to see whether the finished sentence makes the same temporal claim as the evidence.
Links should point to the material actually used, and the saved passage should support the attached claim. Webiano’s workflow discussion of AI hallucinations warns about unsupported synthesis and citation mismatch. Accurate passages do not justify a stronger claim merely because they were retrieved together. The sentence still has to remain within what those passages establish.
A Practical Worked Example
Imagine a small company preparing a 500-word announcement about a new staff training programme. The draft needs to state the programme name, starting date, number of available places, intended participants and a quotation from the operations director. The editor has an approved internal notice and a signed quotation, but the notice does not state the course length. The source box records that absence instead of treating it as an invitation to guess.
The box lists: programme name, “Practical Data Skills”; start date, 12 October; places, 24; participants, UK customer-support staff; and the exact approved quotation. Each item has its supporting passage and document link. A separate line says, “Course length: no approved evidence.” The drafting instruction prohibits adding a duration or inferring one from similar programmes.
The first AI draft says: “The six-week Practical Data Skills programme begins on 12 October and will train 24 UK customer-support staff.” Most of that sentence is supported, but “six-week” is not. Because the source box exposes the gap, the editor can remove those words immediately. The corrected sentence is: “The Practical Data Skills programme begins on 12 October and has 24 places for UK customer-support staff.”
Next, the AI introduces the quotation with “Operations Director Maya Shah said”. The box confirms both the name and role, so those details can remain. The editor compares the quotation character by character with the signed version and corrects one changed word. This is claim-level checking in practice: each factual unit is tested against its evidence rather than approved because the paragraph sounds convincing.
The final review also checks whether “will train 24 staff” overstates the source. The notice confirms 24 places, not 24 completed participants, so the wording stays as “has 24 places”. That small change preserves the difference between capacity and outcome. A general style review might easily overlook it.
This hypothetical example shows how a visible evidence boundary can expose an unsupported detail before publication. It does not suggest that the source box removes the need to verify the result. Its purpose is to make gaps and altered claims easier to identify while there is still time to correct them.
Check the Draft Claim by Claim
Once the draft exists, do not rely solely on a smooth read from beginning to end. Break each material sentence into atomic factual claims. A sentence containing a person, role, date and figure may contain four separate claims, even if it reads as one neat line. Check each element against the relevant source-box entry.
Use claim-level verification alongside general proofreading because hallucinations can be small, fluent and buried. The editor should ask four questions: what exactly is being asserted, which passage supports it, does that passage support the whole assertion, and has the draft strengthened or altered the source? If any answer is unclear, revise, qualify or remove the claim. Style polishing comes after factual alignment.
Pay close attention to modal verbs and causal language. “May”, “can”, “is” and “must” carry different levels of certainty or obligation. “Associated with”, “contributed to” and “caused” are not interchangeable. A model can create an unsupported promise simply by replacing a cautious term with a stronger one.
Citations need their own check. Locate the cited passage that supports the sentence. Confirm that it covers the relevant subject, period, jurisdiction, number and qualification. A nearby passage or a generally relevant document is not enough when the stated claim goes further.
The checklist described at The Anti-Hallucination Checklist presents checking as a repeatable daily workflow and also describes it as imperfect. That is the right expectation for a source box. It can support more disciplined drafting, while human verification remains necessary. No checklist should be treated as a promise that invented details cannot survive.
In Plain English
Think of the AI as a fast cook and the source box as the tray of approved ingredients. The cook may combine those ingredients into a clear meal, but it must not fetch an unlabelled jar or invent a missing ingredient. If the tray lacks a date or number, the correct response is to leave a visible gap. The editor then decides whether to find evidence or change the recipe.
What This Means For You
You do not need a complex system to begin. Choose one factual draft and create a source box before opening the drafting prompt. Give every important fact an evidence passage and link, then mark anything unconfirmed. This first run can show which information the brief leaves implicit.
Assign ownership as well as evidence. One person should approve the source box, while the writer or editor checks the generated claims against it. For material your organisation judges to need added scrutiny, a second reviewer can check figures, quotations and obligations. The source box creates a shared record, but responsibility must still be clear.
Keep drafting and approval states separate. A model may suggest a useful structure or identify a missing field, but that does not approve new facts. Add fresh evidence to the box only after a human has assessed it. Then regenerate or revise the affected passage and repeat the claim check.
Teams can record evidence gaps found during review, claims that needed qualification and citations that did not support a sentence. Those observations can help them refine the workflow without turning local experience into a universal promise about error reduction.
The final publication decision remains human. The source box narrows the model’s permitted factual material, exposes gaps and makes verification more organised. It cannot make weak sources reliable or resolve every conflict on its own. Used with claim-level review, it turns “be accurate” from a vague request into a repeatable editorial practice.