Can You Say AI-Powered in Ads? FTC AI-Washing Orders, SEC Fines, and Claim Checklist

On April 28, 2025, the Federal Trade Commission announced an enforcement action against Workado, LLC (formerly Content at Scale AI), a vendor marketing its "AI Content Detector" to businesses and publishers as "98% accurate" across ChatGPT and GPT-4 text samples (FTC, Apr 28, 2025). Regulators checked the underlying data. The model had been trained mainly on academic writing; tested on general-purpose content, its accuracy rate dropped to 53% of non-academic test cases, barely above a coin toss. Finalized in August 2025, the FTC order barred unsubstantiated accuracy claims, required corrective buyer emails, and imposed four years of compliance reports (FTC, Aug 2025).

Guessing at ad metrics carries real corporate and personal penalties. DoNotPay paid $193,000 and sent corrective subscriber notices over its "world's first robot lawyer" ads (FTC, Feb 2025). Air AI drew an $18 million FTC judgment ($50,000 paid on inability-to-pay review plus a permanent business-opportunity ban) after buyers lost roughly $19 million on "human-replacing" sales agents (CFO Dive, 2026). For fundraisers, the SEC fined Delphia and Global Predictions $400,000 combined (SEC, Mar 18, 2024) and charged Presto Automation over undisclosed call-center labor behind its drive-thru voice AI (SEC Order, Jan 14, 2025). Paired with our customer AI disclosure decision tree, this guide offers general operational information, not legal advice.

How does the FTC's prior-substantiation rule judge "AI-powered" ad copy?

Under Section 5 of the FTC Act (15 U.S.C. s. 45), calling an offer "AI-powered" is lawful only when a machine-learning or generative model actually drives the advertised feature and the seller holds competent and reliable evidence on file before publishing any objective performance claim. Running benchmarks after a buyer complains or an agency writes is already unlawful.

Think of a UL fire-rating plate riveted onto a commercial kitchen hood. A manufacturer cannot bolt a "3-hour fire-rated" stamp onto sheet metal on a hunch; an independent test log must exist before the hood ships. Advertising law applies that exact sequence, called prior substantiation, to software and service claims.

Where the fire-rating analogy breaks is who writes the test standard. No federal office issues an official "Certified AI-Powered" seal. Instead, your own copy sets the legal benchmark you must prove. Call a tool "a drafting assistant" and you need only show an LLM helps draft text. Claim a "98% accuracy rate," "replaces two bookkeepers," or "compliant with all hiring laws," and you must hold pre-launch empirical tests across representative sample inputs proving those exact outcomes.

Federal shifts in 2025 and 2026 sharpened that line. On December 22, 2025, the FTC set aside its 2024 Rytr order, dropping a theory that faulted a general AI writing tool because buyers could misuse it for fake reviews (Mintz, Feb 13, 2026). Yet in the same cycle, the FTC finalized direct deception orders against Workado, Growth Cave (DLA Piper, Feb 2026), and Air AI. Neutral tools are lawful; lying about what your AI achieves is not.

What did FTC Operation AI Comply and state AG orders actually penalize?

Launched in September 2024 and expanded through 2026, Operation AI Comply (FTC, Sep 25, 2024) and state Attorney General settlements drew clear red lines across software, agencies, and coaching offers.

Notice Workado's mistake: it borrowed an academic essay score (98% benchmark rate) and pasted it onto a general commercial headline. You can run an honest task test in one afternoon using our 50-example spreadsheet method to test an AI tool, and schools evaluating detector pitches should check our teacher checklist before running an essay through an AI detector.

When do pitch decks and investor updates cross into SEC AI-washing?

Pitch decks, investor updates, Form ADV filings, and public releases cross into SEC AI-washing whenever a firm claims proprietary machine learning or autonomous AI drives its product while relying on basic rules, unused data, or hidden human workers. Even private seed startups raising capital from accredited investors remain subject to Exchange Act Rule 10b-5 and Securities Act Section 17(a).

Three settled SEC cases map the boundary:

1. Delphia (USA) Inc. ($225,000 penalty, Mar 18, 2024): Claimed for four years that it trained predictive machine learning on collective client data when it never built that pipeline (SEC Release 2024-36).
2. Global Predictions Inc. ($175,000 penalty, Mar 18, 2024): Claimed to be the "first regulated AI financial advisor" offering "expert AI-driven forecasts" without proof (SEC Release 2024-36).
3. Presto Automation Inc. (SEC Order, Jan 14, 2025): Promoted drive-thru Presto Voice as proprietary speech AI eliminating human order-taking, while failing to disclose in SEC filings that third-party models powered key lanes and off-site human workers handled over 70% of voice orders across deployed restaurant sites (SEC File No. 3-22413).

The rule for founders is crisp: human-in-the-loop review is an asset when disclosed and fraud when concealed. Also check your D&O policy: as shown in our guide to business insurance AI exclusions (CG 40 47 and Berkley D&O gaps), absolute AI exclusions on 2026 D&O renewals frequently strip defense cover for AI-washing inquiries.

How does Massachusetts Chapter 93A raise the stakes beyond the FTC?

Massachusetts Attorney General Andrea Joy Campbell's April 16, 2024 AI Advisory confirmed that M.G.L. c. 93A, s. 2 applies directly to false AI capability claims, unverified "bias-free" promises, and defective AI tools (Orrick, Apr 2024). Unlike the FTC Act, which has no private right of action, Massachusetts Chapter 93A lets both consumers (s. 9) and businesses (s. 11) sue directly for double or treble damages plus mandatory attorney's fees.

AG Campbell reinforced that advisory in July 2025 with a $2.5 million settlement against lender Earnest Operations over unlawful and disparate AI underwriting practices (Mass.gov, Jul 10, 2025).

Worked math: Hyped agency landing page vs substantiated copy

Assume an 8-person Massachusetts agency sells a $1,500/month "99% Accurate, Zero-Staff AI Receptionist" to 12 local dental offices ($108,000 billed over six months). Missed urgent calls cause four clinics ($50,000 combined fees and lost bookings) to serve a Chapter 93A, s. 11 demand.

• Unverified "99% accurate / zero-staff" copy: Knowing misrepresentation without prior test logs risks $150,000 in trebled damages (3 x $50,000) plus $45,000+ in plaintiff legal fees, wiping out nearly two years of revenue.
• Substantiated copy ("AI-assisted call triage via OpenAI API; routed 46 of 50 test calls [92% sample rate] in our Oct 2026 dental benchmark, with live SMS front-desk fallback"): Running a 50-call benchmark takes 3 hours (~$180 staff cost) and removes the deception claim.

Audit your copy before launch: The ed.cafe Marketing & Sales AI Claim-Substantiation Table

Before publishing a landing page, ad, sales script, or investor deck, cut three phrases unless counsel signs off on empirical logs: "autonomous," "replaces [licensed role]," and "bias-free." State plainly whether you wrap a third-party API vs an in-house model, and archive dated benchmark files in a /Substantiation/ folder before launch.

Copy the ed.cafe Marketing & Sales AI Claim-Substantiation Table below into your pre-launch review doc. The ed.cafe Marketing & Sales AI Claim-Substantiation Table maps eight common AI marketing claims to exact pre-publication proof files.

THE ED.CAFE MARKETING & SALES AI CLAIM-SUBSTANTIATION TABLE
Publisher: ed.cafe by No Worker Left Behind (Boston 501(c)(3), EIN 88-3268312)
Apply to ads, landing pages, sales scripts, proposals & investor pitch decks.
Operational compliance tool; not legal advice.

PART 1: 8-CLAIM EVIDENCE MATRIX (SAVE IN /SUBSTANTIATION/ *BEFORE* PUBLISHING)
1. "AI-Powered" / "Machine Learning"
   - Proof Needed: Architecture doc or API log showing an ML/LLM model materially
     performs the advertised task (not plain if/then rules or manual labor).
2. "[X]% Accurate" / "Cuts Errors by [X]%"
   - Proof Needed: Dated test log (n >= 50 representative real-world inputs),
     confusion matrix, model version, and explicit domain scope note.
3. "Replaces [Role]" / "Matches a Human [Lawyer/CPA/Rep]"
   - Proof Needed: Controlled blind comparison against qualified practitioners
     across edge cases + professional-licensing legal sign-off.
4. "100% Automated" / "Zero Human Touch"
   - Proof Needed: Production telemetry confirming 0% human contractor fallback;
     otherwise disclose exact human-in-the-loop review or escalation share.
5. "Maximize Revenue" / "Pays for Itself in 30 Days"
   - Proof Needed: Audited buyer cohort distribution (median result), written
     typicality disclosure, and FTC Business Opportunity Rule review.
6. "Proprietary AI Model" / "Trained on Collective Data"
   - Proof Needed: Training logs, weights repo, and data-consent audit. If calling
     OpenAI/Anthropic/Google APIs, write "built on enterprise LLM APIs."
7. "Bias-Free AI" / "100% Compliant with All AI Laws"
   - Proof Needed: Avoid blanket "bias-free" or "compliant" badges; keep dated
     disparate-impact audits and state only the specific test completed.
8. "Saves 10 Hours a Week"
   - Proof Needed: Pre/post time study naming sample size (n), role, verification
     minutes subtracted, and date measured.

PART 2: 6-STEP PRE-PUBLISH SIGN-OFF CHECK
[ ] 1. Every number or "%" links to a dated file saved in /Substantiation/.
[ ] 2. Human review or escalation steps are stated plainly as a feature.
[ ] 3. Third-party LLM APIs vs custom models are described accurately.
[ ] 4. Zero "replaces licensed staff," "bias-free," or guaranteed ROI promises.
[ ] 5. Investor deck matches live architecture today (roadmap items labelled).
[ ] 6. Product demos show typical outputs, not unlabelled cherry-picks.

Freelancers and career-changers can market verified AI skills truthfully by building a transparent portfolio piece; see creating a one-page work sample after the Google AI Certificate. Massachusetts residents can take that certificate free via the Mass AI Hub, and No Worker Left Behind is forming Circle #1 at ed.cafe/start/, a free 30-chair cohort closing December 1, 2026 (pre-launch as of October 2026; zero cohorts completed yet).

Where this stops being true

• Subjective puffery: Non-measurable opinions reasonable buyers do not take literally (like "smartest workspace for busy teams") remain lawful puffery if unpaired with fake metrics.
• Regulated clinical, financial, and legal sectors: FDA medical-device rules, FINRA Rule 2210, SEC Marketing Rule (206(4)-1), and state bar advertising rules impose stricter pre-approval rules than general FTC standards.
• Business-Opportunity Rule (16 CFR Part 437): Selling turnkey AI setups or coaching with recoupment promises triggers mandatory 7-day advance disclosures regardless of how cautiously you describe the AI.
• Not legal advice: Consult qualified advertising, securities, or Massachusetts Chapter 93A counsel before publishing high-stakes claims or answering an agency inquiry.

Questions people ask

Key terms

• AI washing: Overstating the presence, autonomy, proprietary nature, or measured accuracy of AI in ads, sales pitches, or investor filings.
• Competent and reliable evidence: Objective tests or analyses using sound methods that substantiate a performance claim before publication.
• Massachusetts Chapter 93A (M.G.L. c. 93A): State consumer and B2B protection law barring deceptive acts and authorizing private suits with mandatory fees and up to treble damages.
• Means and instrumentalities: Liability theory for supplying tools designed to deceive, narrowed for general AI tools when the FTC vacated Rytr in December 2025.
• Operation AI Comply: FTC enforcement sweep launched in September 2024 targeting deceptive AI capability, legal-replacement, and earnings schemes.
• Prior substantiation: FTC doctrine requiring advertisers to hold objective proof supporting performance claims before an ad runs.

Sources

1. FTC, "Operation AI Comply Crackdown" (Sep 25, 2024): FTC
2. FTC, "Final Order with DoNotPay ($193,000)" (Feb 2025): FTC
3. FTC, "Workado AI Detection Orders (98% vs 53% Test Cases)" (Apr & Aug 2025): FTC Apr and FTC Aug
4. CFO Dive, "Air AI Settles FTC Deception Charges ($18M Judgment)" (2026): CFO Dive
5. DLA Piper, "FTC Resolves Growth Cave AI-Washing Case" (Feb 2026): DLA Piper
6. Mintz, "FTC Sets Aside Rytr Order" (Feb 13, 2026): Mintz
7. SEC, "Delphia & Global Predictions AI-Washing Charges ($400k)" (Mar 18, 2024): SEC
8. SEC, In the Matter of Presto Automation Inc., File No. 3-22413 (Jan 14, 2025): SEC PDF
9. Orrick, "Massachusetts AG AI Guidance (M.G.L. c. 93A)" (Apr 2024): Orrick
10. Orrick, "Texas AG Settlement with Pieces Technologies" (Sep 2024): Orrick
11. Mass.gov, "AG Campbell $2.5M Earnest AI Settlement" (Jul 10, 2025): Mass.gov