“The AI Made It Up” Is Not a Defense: FTC Substantiation for AI-Written Copy
An AI-drafted nurture email arrives for review. Second paragraph: “Teams using our platform cut onboarding time by 43 percent, according to a 2025 industry study.” It reads fine. It is the kind of sentence marketing emails have. Nobody on the team has seen the study, because there is no study. The model produced the sentence because sentences like it appear in emails like this.
If that email goes out, the company has made an unsubstantiated performance claim. The person who approved it did not invent the number, and it does not matter. The Federal Trade Commission’s standard has been the same for decades: an advertiser needs a reasonable basis for an objective claim at the time the claim is made. Nothing in that sentence mentions who wrote the copy.
This post is not legal advice, and a company with real exposure should have a lawyer read its claims. What it offers is a working method for a marketing team of one to five that has no in-house counsel and is now drafting most of its copy with a model: the claims map, five classes of claim that every sentence in a draft falls into, the source each class needs, and an annotated email showing where the lines are.
The standard did not change when the tool did
Three things are settled and worth stating plainly.
Substantiation attaches to the advertiser. The FTC’s long-standing policy is that the advertiser must possess and rely on a reasonable basis for objective claims before disseminating them. “Before” is the operative word; you cannot substantiate after the complaint. And “advertiser” is you, whether the sentence came from an agency, an intern, or a language model. The agency has said in public, repeatedly, that there is no AI exemption from the laws it enforces, and it ran an enforcement sweep in 2024 aimed specifically at companies whose AI-related claims were unsupported.
Fake testimonials are a rule, not a guideline. The FTC’s rule on consumer reviews and testimonials took effect in October 2024. It prohibits creating, purchasing or disseminating testimonials from people who do not exist, who did not have the experience described, or that misrepresent it. A model asked for “a customer quote about time savings” will produce one, attributed to a plausible name at a plausible company. That output is the prohibited conduct, drafted for you.
Health and financial claims have their own bar. The FTC’s health products guidance, updated in December 2022, and its financial-services enforcement expect competent and reliable evidence, often scientific, for claims in those categories. A model drafting copy for a supplement, a clinic, a lender or an insurance product will produce claims that need a level of support the model cannot supply and the marketer usually does not have on file.
Civil penalties per violation are adjusted annually and are now well above fifty thousand dollars, and the more common outcome, a consent order that dictates how the company markets for years, is arguably worse for a small team than the fine.
The claims map
The reason AI copy is dangerous here is not that models lie more than people. It is that models produce claim-shaped sentences fluently and without a source, in every paragraph, and a reviewer reading for tone will not see them. The fake stat post covers how to spot fabricated evidence. This is about what to do with every sentence that makes a claim, fabricated or not.
Sort each sentence into one of five classes. Each class has one substantiation source. A sentence in a class does not ship without its source.
| Claim class | What it sounds like | The source it needs |
|---|---|---|
| Fact about us | ”We support 40 integrations.” “Founded in 2019.” “Our plans start at $249.” | The spec sheet, the pricing page, the facts file. Current, not the model’s memory. |
| Performance or result | ”Cuts onboarding time by 43%.” “Customers see ROI in 90 days.” | Your own data, documented and representative, or a real study you can produce. If the result is not typical, say what is. |
| Testimonial | Any quote attributed to a customer, user or partner | The real person, the real words, a signed release, a record of when and where they said it. |
| Comparative | ”Faster than HubSpot.” “Half the price of Intercom.” | A documented, current, apples-to-apples comparison you could show a regulator. |
| Health or financial | Anything about health outcomes, safety, returns, savings, approval odds | Competent and reliable evidence to the standard of that category. Usually: not in a nurture email at all. |
Two rules make the map fast. The point survives the claim. “Teams get onboarded faster” needs no study; “43 percent faster” does. If the source is not on file, drop to the version that does not need one. Your own facts get a file. A one-page claims file (the numbers, names, prices and results you are allowed to state, with the source for each) is the highest-leverage compliance document a small team can own, because it turns “is this a stretch?” from an instinct into a lookup. The knowledge base post covers building it, and the founder bottleneck post puts it where it belongs: approved upstream, once, so nobody re-litigates it per email.
An AI-drafted email, annotated
Here is a plausible draft, the kind a model produces from a two-line brief, with each sentence classed.
Subject: See why teams switch to Northwind
Hi Dana,
(1) Northwind helps operations teams at growing companies get new hires productive faster. (2) Teams using our platform cut onboarding time by 43 percent, according to a 2025 industry study. (3) “We went from three weeks to five days,” says Priya M., Head of Ops at a 200-person logistics company. (4) Unlike legacy tools, Northwind connects to your existing HR system in minutes. (5) Plans start at $249 a month. (6) Book a 20-minute walkthrough this week and we’ll map your current process for free.
- (1) Fact about us, soft. No number, no result. Ships as is.
- (2) Performance claim. Needs the study or your own documented data. There is no 2025 industry study. Either replace with a real internal figure that is typical and documented (“customers who completed setup in Q2 reduced onboarding by a median of X days”), or drop to (1)‘s level. As written, this is the sentence the FTC would ask about first.
- (3) Testimonial. Priya M. does not exist. This is the fake-testimonial rule, verbatim. Replace with a real quote from a real customer with a release on file, or cut. Never keep an invented quote “as a placeholder”; placeholders ship.
- (4) Comparative. “Unlike legacy tools” is vague enough to be puffery; “in minutes” is a performance claim about setup time. Keep the comparison generic or document the minutes.
- (5) Fact about us. Check the facts file. If the starting price changed last month, the model does not know.
- (6) Offer. Not a claim, but “for free” is a representation. If the walkthrough has conditions, state them.
Two of six sentences fail outright and two need a check. A reviewer reading for tone would have approved all six.
CAN-SPAM and TCPA do not care who wrote it either. Consent, sender identification, opt-out handling and send-time rules attach to the message and the sender. An AI that drafts a text does not change whether you had consent to send it. Compliance lives in the send system’s rules; the drafting tool has to operate inside them. The SMS compliance post covers the texting side.
Where the check lives
A claims map in a document gets skipped on the day it matters. It has to run at the gate.
The shape that works for a small team: drafts are written with the claims file attached as context, so class 1 is right at the source and the model is told which results and quotes it may use. A review pass, before the human reads, flags every sentence in classes 2 through 5 that lacks a source: every number, every quote, every comparison, every health or money word. The human review starts with that list. Nothing moves to ready until the flags are cleared, either by attaching the source or by softening the sentence to the version that needs none.
That is how Marqeable’s content review works: your brand facts and approved claims live with the account, drafts are written inside them, the review pass flags unsourced performance claims, quotes and comparisons with the quoted line, and a person clears the list before the piece can be scheduled. The model still drafts in seconds. It cannot ship a 43 percent that nobody can produce.
Frequently asked questions
Does the FTC treat AI-generated claims differently?
No. The reasonable-basis standard attaches to the advertiser at the time of the claim, regardless of who or what drafted the sentence. There is no AI exemption, and the agency has brought cases over unsupported AI-related claims.
Can I use an AI-generated customer testimonial?
Not if the customer does not exist or did not say it. The FTC’s testimonial rule, in effect since October 2024, prohibits exactly that. Real quotes with permission and a record are fine; the model formats them, it never writes them.
What is a claims map?
Sorting every sentence in a draft into five classes (facts about us, performance, testimonial, comparative, health or financial), each with a named source. A sentence does not ship without its class’s source.
Do CAN-SPAM and TCPA apply to AI-written messages?
Identically. Consent, identification, opt-out and timing rules attach to the message and the sender, not the drafting tool.
The bottom line
The FTC’s standard is older than the models and did not move when they arrived: a reasonable basis for every objective claim, held by you, before you publish. Models produce claim-shaped sentences in every paragraph, and a reviewer reading for tone will approve them. Sort every sentence into its claim class, attach the source or soften the sentence, keep a claims file approved upstream, and run the check at the gate before the human reads. “The AI made it up” is a true statement. It is not a defense.
Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access
