How to Stop AI Inventing Details in Your Email Drafts

AI writes a confident reply. You skim it, catch the meeting date it invented — '2pm Tuesday,' which you never mentioned — and fix it before sending. That is the good scenario. The version where you miss it is a credibility problem with a real person, and it happens because the model's default behavior is to fill any factual gap with whatever sounds most plausible.

How to stop AI inventing details in email drafts is a solvable problem, and the solution is not more careful proofreading after the fact. It is three prompt constraints added before the model writes anything, plus one verification step on the output. The constraints change the model's default from 'generate what a fact would sound like' to 'stop and flag the gap.' They work in ChatGPT, Claude, Gemini, Copilot, and any other assistant you already use.

This guide covers the constraints, how to apply them, a verification prompt you can run on any existing draft, a table of how different platforms handle persistence, and what to do when the model ignores the constraints anyway.

Why Does AI Invent Details in Email Drafts?#

AI email drafts fail in two distinct ways. The first is factual invention: dates, prices, order numbers, names, and figures the model generates because you did not supply them. The second is commitment invention: phrases like 'as we agreed,' 'per our last call,' or 'I can confirm' attached to things you never agreed, discussed, or confirmed. Factual invention is embarrassing. Commitment invention creates legal and relationship exposure.

Both failures share the same root cause. A language model generates the statistically likely next sequence of text. When it does not have a concrete fact, it generates what a fact would plausibly look like in context — a price near the range of values it encountered during training, a date that sounds like the kind a business email would name. It does this without uncertainty signals. There is no 'I made this up' marker in the output. The client receives a confident sentence about a discount you never authorized.

The OWASP Top 10 for Large Language Model Applications names this pattern — a model producing output that is factually incorrect but stated with apparent confidence — as a primary risk in drafting and agentic workflows. In email the stakes are concrete: a hallucinated price goes to a client, a hallucinated commitment becomes a dispute. The mechanism is not malicious and it is not a bug; it is how these systems work when the prompt leaves gaps for them to fill.

How to Stop AI Inventing Details: Three Prompt Constraints#

These three constraints work as a system. Apply all three to any prompt where the AI is drafting an email you plan to send.

1. 1

   Add a closed-world instruction

   Add this sentence to your prompt before any other instructions: 'Use only the facts I provide in this prompt. Do not add any detail I have not given you.' This is the closed-world assumption, borrowed from database query logic — if a fact is not in the given information, treat it as absent rather than inferring it. Without this instruction, the model's default is open-world: any plausible-sounding fact is a candidate to fill silence. With it, the model shifts toward the information you actually supplied. Use the exact phrasing 'I have not given you' rather than 'you do not know' — the second phrasing can be misread as an invitation to draw on training-data knowledge rather than restricting to the prompt.

2. 2

   Require bracketed placeholders for unknowns

   Add a second instruction: 'If any fact you would normally include — a date, price, name, figure, or commitment — is missing from what I gave you, write [UNKNOWN: describe what is missing] in that position instead of inventing a value.' This converts the model's instinct to fabricate into an instinct to flag. Instead of a confident '10% discount,' you get '[UNKNOWN: discount percentage, confirm with finance]' — a visible gap you can fill in. The bracket format is deliberate: it is visually distinctive, does not read as part of the email's prose, and survives most copy-paste operations intact. Angle brackets disappear in HTML; curly braces conflict with template syntax in some tools. Square brackets are the safe default across platforms.

3. 3

   Demand an assumptions list at the end

   Add a third instruction: 'After the draft, list every assumption you made — any inference, any guess, any detail you treated as true but that I did not explicitly state.' This is your audit trail. Even a well-constrained model makes logical inferences: it may assume a formal tone is appropriate, guess that 'Tuesday' means next week rather than last, or infer from your role that a figure is in USD. The assumptions list surfaces those inferences, which the closed-world instruction and placeholder rule do not catch because the model filled them logically rather than by inventing a missing fact. Review the list before sending. Any wrong assumption gets corrected in a follow-up prompt; any correct one confirms the model understood your brief.

4. 4

   Run the verification prompt on the output

   After you have a draft, run this in a separate prompt: 'Read the draft below. For every specific factual claim in it — date, price, name, figure, or commitment — quote the source line from my original prompt where I provided that fact. If you cannot find the source line, flag the claim with [UNVERIFIED].' Paste the draft and your original prompt together. The model then maps each claim to its source or flags it as unverified. This catches two things the earlier constraints miss: facts the model inferred correctly but without a stated source in your prompt, and any fact it invented despite the closed-world instruction. Treat every [UNVERIFIED] marker as a required edit before sending.

Does This Work the Same Way in Every AI Tool?#

The three constraints work in any text-based AI assistant, but the mechanics differ by platform. The most important variable is persistence: whether you need to re-add the constraints to every prompt or whether the tool lets you save them once and apply them by default.

What to Do When the Constraints Still Don't Work#

A constrained model will occasionally still invent a detail — in longer drafts, when the factual density of your prompt is high, or when the tool's context window is near its limit. When this happens, the cause is usually one of three things.

First, the constraint wording was too vague. 'Don't make things up' does not anchor the model's behavior as reliably as the specific closed-world instruction above. Second, the model treated an inference as a task completion rather than a factual invention. If you write 'reply to this client complaint,' the model may infer a resolution, a timeline, or a refund you never mentioned — not because it fabricated a fact, but because it completed the task of writing a resolution email. Only the assumptions list catches this category. Third, the constraint was placed too far from the task instruction in a long prompt. Move the closed-world instruction to the line immediately before your task description, not to the opening of the prompt where it can be deprioritized in a long context.

For any draft still containing suspected hallucinations after the constraints, run the verification prompt from the steps above. That is the diagnostic layer. If a claim comes back [UNVERIFIED], delete it or ask the model to replace it with a placeholder. Do not ask the model to 'fix' a hallucinated fact by rewriting the sentence — rewriting can produce a different hallucination. Removal or an explicit placeholder is the safe edit.

A Faster Way: Built-In Constraints Instead of Manual Prompting#

The constraints above work. They do require you to write them into every prompt, re-verify every draft, and manage all of it in a chat tool that sits outside your inbox. That is friction you pay on every email.

AI Emaily is an AI-native email client that removes that friction. Instead of a chatbot that forgets your session's instructions, it drafts replies grounded in the actual thread in front of you — not in inferences about what a reply might plausibly say. When a fact is missing, it flags the gap rather than filling it. Copilot mode holds every draft for your explicit approval before anything leaves your outbox, which is the structural equivalent of the verification step above, built in by default on every draft. We build AI Emaily. Start a 7-day free trial at aiemaily.com.