A marketer asks an AI system to draft a product page. Within seconds, the draft contains a specific performance figure, a comparison to a competitor, a causal claim about what the product prevents, and a sentence implying the approach is validated by independent research. None of this required unusual effort. It required a prompt.

The natural next question is whether any of it is true. But that is not quite the right first question, because “true” and “supportable” are not identical, and an organization publishing the sentence does not need it to be merely true in some abstract sense — it needs to actually possess, before publication, whatever basis would justify saying it. The draft can be accurate by coincidence and still be something the organization is not yet in a position to publish responsibly, because no one has confirmed the basis for the claim. This is not a new problem introduced by AI. It is an old regulatory principle meeting a production environment that generates finished-sounding sentences much faster than any organization can generate the evidence behind them.

Fluency Creates a False Sense of Readiness

A polished draft looks finished. That is precisely what makes it useful and precisely what makes it risky.

Linguistic completion and evidentiary completion are different states, and a fluent AI-generated paragraph can reach the first without coming anywhere near the second. The sentence is grammatically sound, well-paced, appropriately confident in tone — and none of that tells you whether the underlying figure is current, whether the comparison is fair, or whether anyone at the company has actually checked. A rough draft announces its own incompleteness; a fluent one doesn’t. This is a genuine operational hazard, separate from any question about whether AI systems are prone to factual error. Even a completely accurate draft can create a false sense that the organizational work behind it — sourcing, scoping, authorization — is also done, when it may not have started.

Editing and Verification Are Different Functions

Most organizations already have someone review AI-generated marketing copy before it goes out. The relevant question is not whether review happens, but what the reviewer is actually checking.

Editing asks whether something is clear, whether it fits the brand’s voice, whether the structure is strong, whether the language does its persuasive job well. These are real and valuable questions. They are also, on their own, silent about a separate set of questions: Is this specific figure accurate and current? Does the underlying evidence actually establish what the sentence claims, or something narrower? Does this apply to the product being described, or was it true of an earlier version, a different market, a different customer segment? Is the sentence implying a causal relationship the evidence doesn’t support? Is the comparison to a competitor fair, or does it depend on a technicality? Is the certainty being expressed — “the fastest,” “proven to,” “guaranteed” — actually warranted, or is it just how the sentence happened to come out?

An editor asking the first set of questions can substantially improve a sentence without ever touching the second set. The sentence gets sharper, more persuasive, more on-brand — and precisely as unverified as it was before. This is the central distinction the rest of this piece depends on: a reviewed draft is not automatically a verified one, and the review that improves how something reads is not the same activity as the review that establishes whether it’s true.

The Claim Changes What’s Required

Not every sentence in a piece of marketing content carries the same burden. A statement of brand personality, an expression of enthusiasm, a subjective description of experience — these operate differently from an objective, checkable assertion about what a product does, how it compares, or what results a customer can expect. The more a sentence functions as a factual, causal, comparative, or quantified claim, the more the organization needs to know, in advance, what actually supports it. Treating a stylistic flourish and a performance statistic as though they carry identical verification requirements is its own kind of error — one that either under-scrutinizes real claims or buries genuine communication in unnecessary process. The judgment about which category a given sentence falls into is itself part of what verification, as a function, has to do.

What the Governing Standard Actually Says

This distinction is not new, and it did not originate with AI. It is the same distinction advertising law has built its substantiation framework around for four decades.

The FTC’s Policy Statement Regarding Advertising Substantiation, issued in 1984 and still the Commission’s operative guidance, states the core requirement plainly: advertisers and ad agencies must have “a reasonable basis for advertising claims before they are disseminated,” and this obligation applies to “express and implied claims, however conveyed, that make objective assertions about the item or service advertised.” The word “before” is doing real work in that sentence. The requirement is not that a claim eventually turn out to be true. It is that the basis for believing it exists prior to publication — which means the verification question has to be answered before the sentence goes out, not discovered to be a problem afterward.

The Commission’s framework also makes clear that the required level of substantiation depends on what the ad actually communicates, not just what it explicitly states. An advertisement can imply more support than it literally claims, and the advertiser is responsible for that implied level too — “when a seller’s representation conveys more than one meaning to reasonable consumers, one of which is false, the seller is liable for the misleading interpretation.” This matters directly for AI-assisted drafting, because generated language frequently produces confident, fluent phrasing that implies a stronger evidentiary basis than anyone actually possesses — not through any intent to deceive, but simply because confident phrasing is what the model was optimized to produce.

None of this is specific to AI-generated content. The FTC’s framework applies to how a claim is made, not what tool made it. This matters because it removes a tempting but incorrect escape hatch: the standard does not soften because a machine, rather than a person, generated the first draft of the sentence. Responsibility for substantiation sits with the organization disseminating the claim, regardless of how the words were produced. It is also worth being precise about jurisdiction and scope here: this is U.S. federal advertising law, administered by the FTC under Section 5 of the FTC Act; it does not describe every jurisdiction’s rules, and this discussion of it is not legal advice — an organization with a specific claim in question should seek that from qualified counsel, not from a marketing essay.

A recent enforcement action illustrates exactly how this principle applies when AI is the tool involved, rather than the excuse. In September 2024, the FTC charged DoNotPay, Inc. — which marketed its AI-powered service as “the world’s first robot lawyer,” claiming it could generate documents and give advice at the level of a human attorney across more than 200 areas of law — with deceptive practices under Section 5. According to the FTC’s own account of the case, the company had not tested whether its AI system actually performed at the level it claimed, and had not retained attorneys to check the accuracy of the legal content it generated. The Commission finalized a consent order in January 2025 by unanimous vote, requiring $193,000 in monetary relief, notice to past subscribers about the service’s limitations, and a prohibition on advertising that the service performs like a real lawyer “unless it has sufficient evidence to back it up.” What matters here is narrower: according to the FTC, the company had not done the testing needed to establish that the claims were supportable before publication.

Why Review Has to Do More Than Polish

Ask a straightforward operational question inside any organization already using AI to draft marketing copy: what is the human reviewer in that workflow actually checking for?

If the answer is tone, grammar, persuasive strength, and whether the piece sounds like the brand, then a genuinely important function is being performed — and it is not the function that satisfies a “reasonable basis” standard. A sentence can pass every one of those checks and still be a claim no one has verified. This is not a criticism of editors; editing is a real skill addressing a real need. It is an observation that editing and verification answer different questions, and an organization can staff, train, and reward the first thoroughly while assuming, often without ever deciding to, that it has also covered the second.

The distinction is between a human editor and a human verifier. One person can perform both roles in sequence — but only if the organization has actually designed the workflow to require both, rather than treating a single pass of human eyes as sufficient regardless of what that pass was checking for. A polished, brand-appropriate, factually unverified sentence is a more dangerous artifact than an obviously rough one, precisely because it no longer looks like it needs anything further done to it.

Where This Becomes an Organizational Design Problem

The deeper issue is not whether any individual writer or reviewer is careful. It is whether the organization has actually defined, as a matter of process rather than assumption, who is responsible for establishing that a claim is supportable before it goes out.

That requires answers to questions most content workflows have never explicitly worked through. Who verifies a factual or comparative claim, as distinct from who edits the sentence containing it? What evidence is required before a given type of claim — performance figure, causal statement, comparison, guarantee — is authorized for publication, and does that requirement scale with how consequential the claim is? When does subject-matter expertise need to be consulted, and when does legal review actually need to be involved, as opposed to a general content reviewer assuming the claim is fine because it reads fine? What happens when two sources inside the organization disagree about a number, or when the evidence supporting a claim has gone stale since it was first approved? And, underneath all of that: who actually has the authority to say a piece of content is ready to publish, and what, specifically, does that authorization mean they have confirmed?

These questions do not have a single universal answer, and this essay is not proposing a compliance framework. The point is narrower: verification, when it exists at all, is usually an emergent accident of who happened to be in the review chain, rather than a function anyone deliberately designed. AI-assisted drafting does not create this gap. It removes the friction that used to disguise it, because when drafting was slow, there was more incidental time for someone to notice an unsupported claim before it reached a customer. When drafting is fast, that incidental protection disappears, and whatever verification process existed — or didn’t — becomes visible in a way it wasn’t before.

What Leadership Should Actually Ask

The obvious response to everything above — have a human review the AI output — was never wrong. It was just never specific enough to be useful. The more useful version of that instruction requires answering what, exactly, the human is checking, and whether the organization has ever explicitly decided that.

A more productive set of questions for leadership follows from the distinctions this essay has tried to draw. What evidence currently supports the claims in our highest-visibility content, and who last confirmed that evidence is still current? When a reviewer approves a piece of AI-assisted copy, are they confirming that it reads well, that it’s factually supportable, or both — and does everyone in that review chain agree on the answer? Is there a category of claim in our marketing — a performance number, a comparison, a guarantee — that should require a different, more rigorous check than the rest of the content we publish, and does our current process actually treat it differently? And perhaps most directly: has the speed at which we can now produce language outpaced the speed at which we can establish what that language is entitled to say?

None of these questions has an easy universal answer, and none of them is answered by simply adding another reviewer to the chain. They are questions about whether an organization has designed verification as its own function — separate from editing, separate from brand approval — or whether it has been quietly assuming that a well-written sentence has also, somehow, earned the right to be believed.