Scalewell / Blog

AI washing in B2B marketing: how to talk about AI in ads without overclaiming

What regulators have charged as AI washing, where the US rules are heading, and a five-level ladder for writing AI claims in B2B ads that your team can back.

A magnifying glass held over a cracked AI badge beside a stack of evidence documents

AI washing is claiming more artificial intelligence than a product has, or claiming its AI does more than you can prove. In B2B marketing, you avoid it by describing AI in ads and landing pages only in words the team can back with a test, a document or a live demo, and by cutting every claim it can't.

The pressure usually shows up two weeks before a launch. Product marketing wants "AI-powered" in every headline, the sales deck has started saying "autonomous", and the paid team plans to run both in front of a buying committee on LinkedIn. Then someone from legal asks a fair question in the review thread: what exactly does the AI do, and where's the proof? Nobody has a one-line answer.

If you approve that copy as a CMO, VP of Marketing or demand gen lead, this guide is for you. It sets out what regulators have actually charged, by name and date, where the US rules are heading, and a five-level ladder for AI claims with the evidence each level needs. It's not legal advice, so have counsel read anything you plan to run. The wider risk questions are in our guide to the generative AI risks a CFO or legal team will ask about, and the team rules are in our AI policy template for a marketing team. This one stays on the claims in your ads and pages.

AI washing means claiming AI the product can't show

AI washing covers two different mistakes. The first is claiming AI that isn't there, such as an app sold as automated while people do the work behind it. The second is real AI described as doing more than it does, such as an accuracy figure that only holds on one type of data.

The FTC set out the questions it asks in a February 27, 2023 post by Michael Atleson of its Division of Advertising Practices, titled "Keep your AI claims in check". It asks whether you are exaggerating what your AI product can do, whether you are promising it does something better than a non-AI product, whether you are aware of the risks, and, bluntly, "Does the product actually use AI at all?" The post called AI "a marketing term" and said false or unsubstantiated efficacy claims "are our bread and butter." The page no longer loads on ftc.gov (it returned a 404 error when we checked in November 2026), but the Wayback Machine copy from March 2025 is still the clearest checklist a marketer can get from a regulator.

One line in it matters more for marketing teams than for anyone else: "merely using an AI tool in the development process is not the same as a product having AI in it." If your team wrote the ads with Claude and drew the hero image with a generator, the product still has exactly as much AI as it had before. Plenty of launch copy blurs that line without meaning to.

A marketing team that runs AI in its own weekly work learns where the technology stops, and that makes a product's AI far easier to describe without stretching it. Putting AI into that weekly work, across strategy, planning, implementation and operations, is what Scalewell does with paid and demand gen teams.

US regulators have charged AI washing under laws that already exist

US regulators have brought AI claims cases under the advertising and securities laws they already enforce, and in several of them the buyers said to be misled were businesses. The record below covers the cases a B2B marketer should know, each checked on the regulator's or NAD's own page in November 2026.

DateWho actedCompanyThe claim at issueOutcome
March 18, 2024SECDelphia and Global PredictionsDelphia said its AI would "predict which companies and trends are about to make it big"; Global Predictions called itself the "first regulated AI financial advisor"Settled without admitting or denying the findings; $225,000 and $175,000 in civil penalties (SEC)
September 25, 2024FTC, Operation AI ComplyDoNotPay and four othersDoNotPay sold itself as "the world's first robot lawyer"; the FTC alleged it never tested that claimOrder made final on February 11, 2025: DoNotPay pays $193,000 and may not claim to replace professional services without evidence (FTC)
January 14, 2025SECPresto AutomationClaimed, in the SEC's words, that its own AI product "eliminated the need for human order-taking"Settled cease-and-desist order without admitting or denying the findings; no penalty, citing cooperation (SEC)
April 9, 2025SEC, with a parallel criminal caseNate's former CEO, Albert SanigerTold investors purchases were completed by AI; contract workers entered many orders by handCharged in a civil SEC complaint and a criminal indictment in New York; both are allegations, not findings (SEC, Justice Department)
June 9, 2025NAD, BBB National ProgramsMicrosoft 365 CopilotProductivity claims built on a survey of how productive users feltRecommended discontinuing the claims or modifying them to disclose their basis; Microsoft disagreed in part but said it would follow the recommendations (BBB National Programs)
August 28, 2025FTCWorkadoDetector "over 98 percent" accurate "across all types of text"Final order, 3-0 vote (FTC)
May 21, 2026FTCCox Media Group and two smaller firmsAn "Active Listening" ad service said to target ads from conversations picked up by smart devicesOrders made final on August 27, 2026, with $930,000 in total payments for customer redress (FTC)

Three of those cases read like B2B marketing problems. Presto sold drive-thru ordering technology and told investors, in SEC filings and public statements, that humans weren't needed. The SEC found the vast majority of orders on its own AI version needed human help, and that for a time a third party owned and ran the speech recognition in every deployed unit. The Cox Media Group service was marketed to small businesses that wanted to reach people in their area, and the FTC alleged it resold email lists bought from other data brokers instead of listening to anything. Workado's figure fell apart outside its training data: per the FTC's complaint notice, the model was "trained or fine-tuned to accurately classify only academic content," and testing showed it classified non-academic AI text with "around 53 percent accuracy."

The Microsoft decision is the one to show your product marketers, because no regulator was involved and the claim had a real study behind it. NAD, the advertising industry's self-regulation body, found the study "demonstrates a perception of productivity" but "does not provide a good fit for the objective claim." A customer survey can support a claim about how customers feel. It can't support a claim about what the product does.

The US direction on AI ad rules is open, but deception cases continue

The rules for AI in US advertising are moving in two directions at once. Federal policy has turned against new AI-specific restrictions, while ordinary deception cases about AI keep coming from the same agency.

On the loosening side, the FTC reopened and set aside its 2024 order against the AI writing tool Rytr on December 22, 2025, by a 2-0 vote. It said the complaint failed to show a violation and that the order "unduly burdens innovation in the nascent AI industry." Then, on July 1, 2026, the FTC proposed a policy statement on AI accuracy under Executive Order 14365, signed December 11, 2025. The proposal says companies that "distort their systems' outputs to achieve undisclosed ideological objectives" could be deceiving consumers, and says a state law that pushes companies to alter accurate outputs, naming Colorado's AI Act as one that appears to, is "impliedly preempted" to the extent it conflicts with federal law. Comments closed on July 31, 2026, and as of November 21, 2026 the FTC's policy statements page still lists it as proposed.

On the enforcement side, nothing has softened for false capability claims. The Workado order, the Air AI case and the Cox Media Group settlements all came in 2025 and 2026, from the same agency that set aside the Rytr order. Air AI, sued on August 25, 2025, marketed "conversational AI" it said could replace human customer service representatives, and the FTC says many of the buyers were small business owners; a stipulated order was filed on March 24, 2026, and the FTC's case page still lists the case as pending. The SEC set up a Cyber and Emerging Technologies Unit on February 20, 2025 that lists "fraud committed using emerging technologies, such as artificial intelligence and machine learning" among its priorities.

The states add rules of their own. New York General Business Law section 396-b, in effect since June 9, 2026, requires an advertiser who knows an ad contains a synthetic performer to disclose it "conspicuously", with penalties of $1,000 for a first violation and $5,000 after. Audio-only ads and AI used only to translate a real performer are exempt. Our guide to LinkedIn video ads for B2B with AI covers what that means for a shoot.

So the practical reading for a marketing leader is narrow and fairly stable. The theory that a tool could be misused is out of favor in Washington. A claim that your product does something it doesn't is as exposed as it has ever been, and the ad platforms say the same in their own rules. LinkedIn's advertising policies state that "any claims in your ad must have factual support," and Google's misrepresentation policy bans "inaccurate claims or claims that entice the user with an improbable result."

Five levels of AI claim, and the evidence each one needs

The risk in an AI claim rises with what it promises, so sort every claim by level before anyone argues about wording. We'd use five levels, from what you used to make the ad up to what the product replaces.

LevelWhat the copy saysExample wordingEvidence to have on file
1. ProcessAI helped make the marketing"Our team drafted this guide with AI and checked every figure"Nothing about the product; keep it out of product claims
2. PresenceThe product contains an AI feature"Includes an AI assistant that summarizes call notes"The feature ships in the current version, and a named owner confirms it
3. CapabilityThe feature does a specific task"Drafts a reply to an inbound lead in under a minute"A test on realistic data, dated, with the conditions written down
4. Comparison or outcomeAI beats the old way or moves a number"Cuts weekly reporting time in half"A controlled comparison against the non-AI method, on customers like the ones the ad targets
5. Autonomy or replacementThe AI works without people, or replaces them"Fully autonomous", "replaces your SDR team"Proof that no human steps in, which most products can't supply
A five-step staircase rising left to right, with a taller stack of documents on each higher step
The five levels of AI claim, from process (lowest step) through presence, capability and comparison to autonomy (top step). The evidence a claim needs grows with each step.

The cases above map onto these levels. DoNotPay's robot lawyer, and the claims about removing people described in the Presto order and the Nate and Air AI complaints, are level 5. The Copilot productivity claims were level 4, resting on a survey of how users felt. Workado's accuracy figure was a level 3 claim tested on one kind of text. That pattern gives you the rule: when you're unsure, drop the claim a level and say exactly what the product does.

Here is a hypothetical to make it concrete. Company A and Company B both sell revenue intelligence software with the same transcription model inside. The LinkedIn ad from Company A says "AI that runs your pipeline reviews for you", a level 5 claim nobody can prove. Company B runs "Transcribes every sales call and flags deals with no next step booked", which is a level 3 claim it can demo in the first meeting. B's ad is also the one a skeptical buying committee is more likely to believe, because it names a job the reader recognizes.

Rewrite overclaims into claims you can back

The safest rewrite of an AI claim names the job the AI does, the conditions it was tested under, and where a person still steps in. The table below covers the phrases that turn up most often in B2B launch copy.

Risky phraseSafer rewriteWhat you need on file
"AI-powered platform""Uses AI to score inbound leads against your ICP"The feature spec and the version it shipped in
"Fully autonomous""Runs the weekly pull on its own; your team approves budget changes"A description of every human step
"98% accurate""98% accurate on [named test set], tested [month year]"The test, the data and who ran it
"Replaces your analysts""Takes the weekly data pull off your analysts' list"Time studies, or nothing about headcount
"Customers are 30% more productive""In a survey of [number] customers, [share] said they felt more productive"The survey, its sample and its question wording
"The first AI-native [category]"Drop it, or cite who says soA source that is not your own press release

Two habits make the rewrites stick. Put a date on every capability figure, because a model swap or a retrain can change it without anyone in marketing hearing about it. And keep roadmap features out of live ads entirely; "coming soon" in a headline reads as "available now" to a buyer scrolling past it.

Your vendors' claims need the same treatment. If your product runs on another company's model, you can't borrow that company's benchmark as your own, and the 2023 FTC post was blunt that "you can't just blame a third-party developer of the technology."

The AI claims check belongs in the creative review you already run

An AI claims check works best as a few extra fields in the creative review the team already does, with one named person who confirms each claim before an ad goes live. In our view, a separate committee slows launches until people start routing around it.

Start a claims register if you don't have one: every number, comparison and capability your ads and landing pages may use, with its source and a review date. For AI claims, add four fields: the claim level, the shipped version the feature is in, every step where a person is involved, and the date of the last test. Our creative approval workflow shows how an AI pre-check can compare each new ad against the register before brand and legal see it.

Give each field an owner. Product confirms the feature exists in the shipped version; product marketing owns the wording; legal sees any level 4 or 5 claim before it runs. If your company is public, have someone compare AI claims in the ads with the AI language in investor materials, since the Presto order rested on statements made in SEC filings and public statements together.

Run the check wherever the claim lives: LinkedIn Sponsored Content and Document Ads, Google responsive search ad headlines, Microsoft Advertising, Meta, the landing page and the sales deck. The deck is where claims drift furthest, because it gets edited by whoever has the next meeting. Generated people in a video or image need the New York disclosure wherever that law applies, and our guide to AI generated ads for B2B covers when generated creative helps and when it looks cheap.

Where AI washing slips into B2B copy without anyone deciding to overclaim

Most AI washing in B2B marketing comes from small, reasonable-looking edits made by people under deadline. Nobody sits down to mislead a buying committee; the claim gets one notch bigger at each hand-off.

Watch four places where it usually happens. A product brief says "uses machine learning to suggest", the ad says "AI decides", and the landing page says "autonomous". An agency or freelancer writes variants, and the A/B test picks whichever one gets clicked, true or not. A customer quote with a big number goes into an ad without anyone asking how the customer measured it. And a year-old capability figure stays live after the product has changed.

The first three are fixed by the register and a named approver. The last needs a calendar: review every AI claim when the product ships a new version, and at least once a quarter. And be wary of the test result that says the bolder claim won. A higher click-through rate on "replaces your SDRs" tells you the claim is attractive. It tells you nothing about whether it's true, and the sales team will be the ones explaining the gap on the first call.

How Scalewell gets your team fluent enough in AI to describe it accurately

A team that uses AI every week is better placed to describe it accurately, because it has seen what the tools get wrong. Scalewell's work ends with that team trained: your people learn to run, check and change the AI jobs built with them, so the know-how stays in-house. The first conversation is a free call of 30 minutes, followed by an audit of how your team spends its week, with each recurring job scored for AI. A ranked roadmap puts those jobs in order. Then the first ones get built in a sprint, usually lasting 8 to 12 weeks and occasionally more, in your existing ad platforms, CRM and reports. From then on, your team runs them without Scalewell.

Where to start

Export every live ad and landing page that mentions AI this week, from LinkedIn Campaign Manager, Google Ads, Microsoft Advertising and your site, and give each claim a level from one to five. Anything at level 4 or 5 without evidence on file gets rewritten or paused before the next review. If you'd like a partner to help your team build AI into its own work and leave the people able to run it, Scalewell can help.

Frequently asked questions

What is the difference between AI washing and greenwashing?

Both describe overstating something buyers reward. Greenwashing overstates environmental benefits, such as a product sold as carbon neutral on weak offsets. AI washing overstates artificial intelligence, either claiming AI a product doesn't have or claiming its AI does more than the evidence shows. The legal test in the US is the same for both: is the claim truthful and backed by evidence?

Does AI washing only matter for public companies?

It matters for private companies as well. The SEC's AI washing cases mostly concern statements to investors, and they have reached private companies: the Nate case involved money raised in seed and Series A rounds. The FTC's deception cases apply to advertising by any company, public or private, and several recent ones involved products sold to small businesses and entrepreneurs.

Can a competitor challenge our AI claims?

A competitor can. BBB National Programs' National Advertising Division reviews ad claims brought by competitors as well as cases it opens through its own monitoring, as it did with Microsoft 365 Copilot in 2025. A competitor challenge usually asks you to show the evidence behind the claim, so a claims register with tests on file is the best preparation.

How should a marketing team check an AI vendor's claims before buying?

Ask the vendor for the evidence behind each capability claim on its own site: the test data, the date and who ran it. Ask which parts involve a person and which model the product runs on. If the answers are vague, treat the claim as unproven. Our guide to AI vendor due diligence for marketing has the questionnaire to send.

Should B2B ads stop using the word "AI" altogether?

There's no need to drop it, and buyers expect to hear about it. The word itself carries little risk; the trouble starts with what follows it. "AI" next to a named job, such as scoring leads or summarizing calls, is a claim you can demo. "AI" as a promise about results or replacing people is the claim regulators have charged.