AI Email Copy That Invents Facts About Your Product: 7 Fixes

TL;DR

AI-drafted emails state prices, deadlines, integrations and testimonials nobody gave the model. Seven checks before send, tied to CAN-SPAM, FTC pricing and review rules, and UK CMA guidance.

AI email copy invents product facts for the same reason all generated text does: it writes what a plausible email would say, and a plausible email names a discount, a deadline, an integration and a happy customer. If those specifics were not in the input, the model supplies them. Seven fixes stop that before send: ground the draft in a fact sheet, use placeholders for anything that changes, allow testimonials only from a quote bank, check the subject line on its own, bind every offer term to a record, check each claim with a remove-and-mark pass, and keep the evidence with the send.

Scope: marketing and sales email drafted with a general-purpose language model, against US federal rules and UK guidance as published on 2026-09-17. Not legal advice.

Key Takeaways

Email is the one channel with no correction. A wrong web page can be edited in place; a sent email stays in every inbox as sent (ANALYSIS).

US penalties are counted per message. "Each separate email in violation of the CAN-SPAM Act is subject to penalties of up to $53,088" (FTC CAN-SPAM compliance guide, retrieved 2026-09-17).

The subject line has its own test. It "must accurately reflect the content of the message" (same FTC guide, retrieved 2026-09-17).

The UK wants evidence for factual claims, and bans false countdowns. "Any factual claims made by a trader about a product should be supported by evidence" (CMA, Unfair commercial practices guidance, updated 18 November 2025, retrieved 2026-09-17).

A model provider documents the fix, and its limit. Anthropic's guidance has the model match each claim to a supporting quote and remove what it cannot support, and says these techniques "don't eliminate" hallucinations (Anthropic, Reduce hallucinations, retrieved 2026-09-17).

On this page: Where the invented facts show up · Why email is worse · The rules · Fix 1 to Fix 7 · How This Guide Was Sourced · FAQ

Where the invented facts show up

Six places, and they recur in almost every campaign:

What the model writes	Why it writes it	Where it should come from
A feature or integration ("works with your CRM")	Emails in the category usually mention one	The current product page or changelog
A price, discount or "was" price	Promotional emails usually carry a number	The pricing record for this campaign
A deadline ("ends Friday", "48 hours left")	Urgency is a common email pattern	The offer's actual end date
A customer count, logo or "trusted by" line	Social proof is a common email pattern	An approved, dated figure, or nothing
A testimonial or quote	Same	A signed-off quote bank
A compliance claim ("SOC 2 compliant", "HIPAA compliant")	It sounds reassuring in B2B email	The audit report, or the legal team

The middle column is ANALYSIS, not a measurement; we publish no count of how often each appears. The underlying mechanism, that models reproduce the shape of a typical claim more reliably than a specific fact, is sourced in why AI makes up statistics.

Why email is worse than a web page

Three properties make an invented fact costlier in email (ANALYSIS).

It cannot be corrected in place. A fixed page is what every later reader sees. A sent email stays as sent, and the only correction is a second email that reaches fewer people.

It ships at volume in one action. One wrong sentence in a template is multiplied by the list.

Personalised variants multiply the surface. Each per-segment or per-recipient version is a separate draft that can invent something different. Checking one sample does not check the others.

The product-page version of these errors, with the refund and marketplace rules they trigger, is in 9 AI product description mistakes and what each one costs a seller.

The rules an invented fact breaks

CAN-SPAM covers business email and points past itself. The FTC says "the law makes no exception for business-to-business email", and that "email that makes misleading claims about products or services also may be subject to laws outlawing deceptive advertising, like Section 5 of the FTC Act" (FTC CAN-SPAM compliance guide, retrieved 2026-09-17, SOURCED).

Delegating does not move the responsibility. Same guide: "even if you hire another company to handle your email marketing, you can't contract away your legal responsibility to comply with the law" (SOURCED). It is written about agencies, not software; we read it as applying at least as strongly to a model you chose to use (ANALYSIS).

In the UK the test is whether the average consumer would decide differently. The CMA says traders "should not give consumers information which is objectively false", and that not every error counts: "a typo in an email or a trader's address might be unlikely to cause the average consumer to take a different decision" (CMA guidance, retrieved 2026-09-17, SOURCED).

So the risk is not typos. It is the invented facts a buyer acts on: price, contents, availability, deadline.

Fix 1 — Ground the draft in a fact sheet, and nothing else

Before the model writes a word, give it a short, dated fact sheet for this campaign: product name, what is included, current price, offer terms, end date, integrations that exist today, approved proof points. Then tell it to use only that.

Anthropic's guidance lists this as "External knowledge restriction": instruct the model "to only use information from provided documents and not its general knowledge" (Anthropic, Reduce hallucinations, retrieved 2026-09-17, SOURCED). Research points the same way: retrieval-augmented dialogue models "substantially reduce the well-known problem of knowledge hallucination", verified by human evaluation (Shuster et al., 2021, retrieved 2026-09-17, SOURCED).

Reduce, not remove. The same Anthropic page: "while these techniques significantly reduce hallucinations, they don't eliminate them entirely" (SOURCED). The remaining six fixes exist for what gets through.

Fix 2 — Put placeholders where the volatile facts go

Anything that changes between campaigns should never be generated: price, discount, end date, stock level, customer count, plan name. Instruct the model to write a token — {{PRICE}}, {{END_DATE}}, {{DISCOUNT}} — and fill it from your pricing and offer records at merge time.

The model is good at the sentence around a number and cannot know the number, so split the two jobs (ANALYSIS). It also makes Fix 5 mechanical: every token has one source. A draft that returns a literal price where a token belongs has failed the brief. The same idea applied to questions is in 50 prompts that make AI ask before it assumes.

Fix 3 — Testimonials and customer names come from a quote bank only

Never let a model write a testimonial, a customer quote, a named customer or a logo line. Keep a bank of approved, attributable, dated quotes, and let the model select from it and nothing else.

Under the FTC's rule on reviews and testimonials, it is "an unfair or deceptive act or practice" for a business "to write, create, or sell a consumer review, consumer testimonial, or celebrity testimonial that materially misrepresents" that the testimonialist exists, used the product, or had the experience described (16 CFR 465.2, retrieved 2026-09-17, SOURCED). A quote a model wrote and attached to a customer misrepresents, at minimum, the experience it describes (ANALYSIS). The FTC's endorsement guidance adds that an endorsement "must reflect the honest opinion of the endorser" (FTC, Endorsement Guides: What People Are Asking, retrieved 2026-09-17, SOURCED).

Fix 4 — Check the subject line and preview text on their own

When subject lines are generated as a separate request — "write five subject lines that get opens" — the request rewards the claim the body does not make: a bigger discount, an earlier deadline, a feature still on the roadmap (ANALYSIS).

CAN-SPAM sets a standalone test: "The subject line must accurately reflect the content of the message" (FTC CAN-SPAM compliance guide, retrieved 2026-09-17, SOURCED). Check it against the finished body, not the brief: every number, date and promise in the subject line should appear, identically, in the body. The guide does not name preview text; we hold it to the same standard (ANALYSIS).

Fix 5 — Bind every offer term to a record

Discounts, "was" prices and deadlines are where invented detail becomes a pricing claim. Each should trace to a record that existed before the email was written.

"Was" prices. The FTC's pricing guides treat a former price as legitimate when it is "the actual, bona fide price at which the article was offered to the public on a regular basis for a reasonably substantial period of time"; where it is "not bona fide but fictitious", the bargain "is a false one" (16 CFR 233.1, Guides Against Deceptive Pricing, retrieved 2026-09-17, SOURCED). A model that writes "was $99, now $49" has invented a pricing history.

Deadlines. In the UK, "falsely stating that a product will only be available for a limited time, or that it will only be available on particular terms for a limited time, in order to elicit an immediate decision" is banned practice 7; the CMA's example is a countdown after which "the offer does not in fact end". The guidance says breaches of the banned practices can bring monetary penalties of up to the higher of £300,000 or 10% of worldwide turnover (CMA guidance, retrieved 2026-09-17, SOURCED).

The web-side version of stale prices is in AI is quoting your old pricing.

Fix 6 — Check claim by claim, and remove what cannot be supported

List every factual claim in the draft — each feature, number, date, name, integration and certification — and match each to a line in the fact sheet. No match, no claim.

Anthropic documents this loop with a press release example: "review each claim", find a supporting quote in the documents, and "if you can't find a supporting quote for a claim, remove that claim" and mark where it was removed with empty brackets (Anthropic, Reduce hallucinations, retrieved 2026-09-17, SOURCED). The marker shows a reviewer exactly where the model wanted to say something it could not back up.

A model can run the first pass. A person runs the last one, for every variant that will send. The general method is in how to trace an AI claim back to its original source.

Fix 7 — Keep the evidence with the send

Store the fact sheet version, the offer record and the approved quotes with the campaign that used them. When someone asks why the email said what it said, the answer is a file, not a memory.

Both regulators treat evidence as something you hold when you make the claim. Under the FTC Act, "advertisers must have evidence to back up their claims" (FTC, Advertising FAQ's: A Guide for Small Business, retrieved 2026-09-17, SOURCED). The CMA says that if a trader "is not able to produce such evidence, the CMA or the courts may find the claim to be false on this basis" (CMA guidance, retrieved 2026-09-17, SOURCED). An evidence file does not make a claim lawful. It lets you show a true claim was true.

The seven fixes, in the order they run

Fact sheet — written before the prompt.
Placeholders — in the prompt instructions.
Quote bank — the only source of social proof.
Subject line and preview check — against the finished body.
Offer terms bound to records — prices, "was" prices, deadlines.
Claim-by-claim check with remove-and-mark — every variant that sends.
Evidence stored with the send.

None of the seven makes a campaign compliant or replaces legal review. They move an invented fact from "found by a customer" to "found before send". The wider taxonomy is the AI content errors hub; the full pre-publish list is 42 checks before you publish.

How This Guide Was Sourced

Written and maintained by the LogicBalls editorial team (logicballs.com). Disclosure: LogicBalls builds AI writing tools, including a sales email generator. Every failure described here can occur in output from our tools, and every fix applies to them.

How it was made. AI assisted with drafting. Every source was fetched on 2026-09-17 and each quoted sentence matched against the fetched text before use.

Sources. US: the FTC's CAN-SPAM compliance guide, Advertising FAQ's and Endorsement Guides FAQ; 16 CFR 233.1 and 16 CFR 465.2 via eCFR. UK: the CMA's Unfair commercial practices guidance (CMA207), updated 18 November 2025. Grounding: Anthropic's documentation on reducing hallucinations and Shuster et al. (2021). All linked inline.

What was not checked. EU and other national law. The UK's PECR consent rules, which govern who you may email rather than what it says. The DMCC Act itself — only the CMA's guidance on it.

What is not claimed. No rate at which AI-drafted emails contain invented facts, no frequency ranking of the six error types, and no enforcement case involving AI-written email specifically. The fixes are practices, not guarantees, and nothing here is legal advice.

No LogicBalls telemetry is used in this guide.

Frequently Asked Questions

Is it illegal to send an AI-written marketing email?

No rule we read prohibits AI-drafted email. The rules apply to what the email says, whoever or whatever wrote it — and the FTC says you "can't contract away your legal responsibility" when someone else handles your email marketing.

Can the model write a sample testimonial we replace later?

Use a placeholder instead. A realistic sample quote in a template is one missed edit away from a fabricated testimonial, which 16 CFR 465.2 treats as an unfair or deceptive practice.

Will a better model stop inventing product facts?

It can reduce it, and grounding reduces it further, but the provider documentation we read says these techniques do not eliminate it. No model knows your current price or offer end date unless you give it one.

Conclusion

An AI email draft invents product facts because it was asked for a convincing email and not given the facts. The seven fixes give it the facts, keep the volatile ones out of its hands, and check every claim before send — because email gives you one send and no edit.