AI Washing: Who Audits the Words AI-Powered in 2026?
On 24 March 2026, the Federal Trade Commission attached a price to a marketing adjective. Air AI and its owners agreed to a settlement that would bar them from marketing business opportunities. The proposed order includes an $18 million judgment. Most of that amount is suspended because the company cannot pay. Another $50,000 will go to the Commission for consumer relief.
The details of the charge sheet make an important distinction. The alleged offence was never the use of AI itself. The FTC focused on false earnings claims and a refund guarantee that the company failed to honour. It also challenged statements about what the service could actually deliver. AI appeared on the label. The wider promises attached to that label created the legal problem.
Regulators call the practice AI washing. In 2026, three separate institutions began examining different parts of it at once. The FTC assigns a financial cost to exaggerated claims. Reddit searches for planted grassroots posts meant to influence AI-generated answers. Anthropic, which makes Claude, now places invisible watermarks inside AI-generated text. Sellers once received the benefit of the doubt. This year, that doubt acquired auditors.
For buyers of marketing services, the change affects every pitch containing the words AI-powered. Agencies can no longer expect the term itself to prove technical depth. This guide defines AI washing, explains the role of each enforcement layer and sets out the five questions that distinguish a working AI process from an ordinary software subscription sold at a premium.
What is AI washing?
AI washing means overstating the role, capability or results of artificial intelligence in a product or service to secure customers, investors or attention that the underlying technology has not earned. The term follows the model of greenwashing. Regulators approach both practices as questions of truthful claims rather than debates about the technology itself.
The FTC stated that position clearly when it began Operation AI Comply in September 2024. The sweep targets businesses that overstate what their AI products can accomplish. The Air AI settlement shows the operation moving from warning to enforcement. Strip away the technical language and the alleged conduct resembles a conventional fraud case. Customers were told they could recover tens of thousands of dollars within 30 days. Some were told they could make millions.
Buyers should focus on the distance between the claimed technology and the promised result. Calling something AI is a marketing assertion. That assertion carries the same legal obligations as clinically proven or money back guaranteed. In 2026, an AI claim also draws more scrutiny than either phrase. The technology may be new, but the duty to support an advertising claim is familiar.
Where did AI washing start?
The conduct appeared before the present enforcement push, with financial regulators spotting it first. In March 2024, the Securities and Exchange Commission reached settlements with two investment advisers, Delphia and Global Predictions. Both faced allegations that they had exaggerated how their platforms used AI. Their combined penalties were $400,000. That was a modest amount by SEC standards, selected to send a warning rather than generate substantial Treasury revenue. The agency's message was direct: a company saying AI drives its product must be able to show that AI actually does so.
AI washing takes its name from greenwashing, and the resemblance extends beyond the wording. In each case, the market places a premium on a desirable quality. Some sellers claim that quality instead of doing the expensive work required to build it. Writing the claim takes little effort. Creating the underlying capability can take years. The practice grows while outsiders struggle to verify what sits behind the promise. It eventually attracts auditors and disclosure rules. Buyers then learn to demand evidence before accepting the label.
Marketing services reached that point later than finance. An agency presentation does not face the type of oversight that the SEC applies to a fund prospectus. The FTC's consumer-protection authority began closing that gap. Finance received its first signal in 2024. Marketing received its own in 2026. The move from the first SEC penalty to the first FTC agency-adjacent scalp took two years.
The history of greenwashing suggests what comes next. Claims become less obvious as sellers learn which language attracts attention. Misrepresentation becomes harder to spot. Verification then develops into a service category in its own right. Buyers can expect fewer loud promises and more carefully worded ones. Their job remains the same: examine whether the product does what the seller says it does.
Who audits AI claims now?
Three institutions now bring three different enforcement tools to the issue. None occupied this particular role two years ago.
| Auditor | What it polices | The enforcement | Since |
|---|---|---|---|
| The FTC | Exaggerated AI marketing claims | Bans and monetary judgments (Air AI: $18 million, largely suspended) | Operation AI Comply, September 2024 |
| Stealth brand posts planted to influence AI answers | Its own AI catches roughly 25,000 spam posts and comments a day | Confirmed 6 July 2026 | |
| Anthropic | Unlabelled AI-generated text itself | Invisible watermarks embedded in Claude's output | Announced 11 August 2026 |
Each one covers an escape route left open by the others. The FTC pursues businesses accused of making false outcome claims. Reddit looks for companies manipulating the public sources that AI engines use to form recommendations. Anthropic's watermark can expose machine-generated writing presented as human analysis.
Together, those layers cover the claim and the reputation built around it. They also reach the deliverable itself. A seller may face questions about the results it promises, the public discussion it tries to shape and the origin of the work it sends to a client. AI washing is therefore becoming easier to investigate from several directions.
What did the Air AI settlement actually change?
The settlement turned AI exaggeration from an abstract reputation concern into an offence with an assigned price and a public order. Any prospective buyer can read that order. Its effect reaches beyond the amount named in the judgment because it gives future complainants a documented model for describing similar conduct.
Before March, a founder harmed by an inflated AI sales pitch could make a complaint. After March, that founder could point to a template. The FTC's allegations can be read as a due-diligence list. They cover earnings statements that lack evidence and guarantees that fail when customers seek payment. They also cover descriptions of a service's performance and central characteristics that differ from the actual product.
The case remains subject to court entry. That is why the proposed order says the owners will be banned, rather than stating that the ban already applies. The legal process has yet to reach that final step. The direction of the enforcement action, however, is clear from the terms the FTC is seeking.
Coverage in the trade press reached a common conclusion. PYMNTS and CFO Dive both presented the matter as the first high-profile scalp of the AI-washing era. More cases are likely because the sweep was created to identify and pursue this type of conduct. The settlement also gives buyers language they can use before signing a contract, while the claims are still being tested.
Why is Reddit policing AI claims?
Reddit has become a place where AI engines look for evidence of what people believe. Brands recognized that value before the platforms fully responded to the resulting manipulation. Reddit confirmed on 6 July 2026 that it uses its own AI systems to identify brands placing undisclosed posts intended to appear in answers from ChatGPT and Gemini.
During the first quarter of 2026, Reddit's systems detected roughly 25,000 spammy posts and comments a day. They also prevented about 23 million spam views a day, according to figures reported by Bloomberg and MediaPost. Neither publication says how much of that activity came from brands planting content. Without that breakdown, nobody can honestly assign a figure to the brand-planted share.
The method matters more than the total volume. AI answer engines give substantial weight to community sources when deciding which brands to mention or recommend. A favourable post placed in the right subreddit can cost far less than earning a genuine recommendation. The calculation changes once the post is detected. The company then leaves a public trace of an attempt to manufacture its own reputation inside a venue whose value depends on opinions being authentic.
The practical question for buyers is straightforward. When an agency promises citations from ChatGPT, ask how it plans to earn them. A valid answer involves publishing content engines can verify. The improper route resembles the activity behind 25,000 removals a day. The desired outcome may sound identical in a pitch. The process behind it makes the difference.
What does Anthropic's watermark change?
From August 2026, writing produced by new Claude models contains an invisible watermark. That means an agency's statement that its writers use AI can cease to be private once someone with detection access examines the text. Anthropic announced the system on 11 August. The company introduced it to meet the transparency code under the EU AI Act, which took effect on 2 August and requires AI-generated material to be identifiable by machines.
According to Anthropic's own documentation, models released on or after 2 August 2026 place a statistical pattern within their word choices. Readers cannot see the pattern, but someone with the correct key can detect it. Copying and pasting the text does not remove it.
The signal has limits. Extensive editing can weaken it, as can translation or combining the output with human writing. Anthropic says older models will receive the capability over the coming months. It has also promised detection tools for third parties in documentation that is still forthcoming.
The wider consequence reaches the standard agency sales pitch. For years, clients had little practical way to verify a statement that senior strategists personally crafted their content. A growing category of output can now be checked. An agency that sends unedited model output while charging for senior judgment could soon face an uncomfortable client conversation triggered by a detection tool.
A more candid description has always been easier to defend: AI prepares drafts while named people guide the work, edit it and accept responsibility for the final material. The watermark makes that account more durable because it describes a process that can survive inspection. Agencies that rely on ambiguity will have less room to do so.
Buyers also need to understand the limits listed in Anthropic's documentation. Detecting a watermark proves that text passed through Claude. It does not establish that Claude created the underlying ideas. People also use AI to translate existing material, summarise it or improve writing that began with human thought.
The absence of a watermark proves nothing. Heavy editing can reduce the signal. Short passages may contain too little material for a reliable reading. Older models did not mark their output in the first place. A watermark functions like a smoke detector. It signals that further examination may be warranted, but it cannot issue a final judgment on authorship or quality.
The useful response is narrow. Add a question to every agency-vetting call: how has your process changed now that AI-generated text can carry a watermark? An agency with a considered disclosure policy should answer plainly. A seller built around promises of undetectable content may treat the question as hostile. That reaction can reveal as much about the deliverables as the formal response.
Why does AI washing matter more in B2B than anywhere else?
B2B buyers complete much of their evaluation before speaking with a vendor, and AI exaggeration contaminates the sources used during that evaluation. Forrester's 2024 Buyers' Journey Survey found that 92% of B2B buyers start with at least one vendor already under consideration. It also found that 41% begin with a single preferred vendor. Forrester describes the buying process as confirmation rather than selection.
Gartner's March 2026 survey of 646 B2B buyers shows where much of that confirmation now occurs. The research found that 67% prefer an experience without a sales representative. It also found that 45% used AI during a recent purchase.
Read together, those findings show the risk. A future client may assemble a shortlist before the first sales call. Part of that decision may come from an AI engine whose answer relies on public sources. Some vendors are trying to manipulate those sources. The distortion can therefore reach the buyer long before the vendor faces direct questions.
A buyer who fails to distinguish genuine AI capability from AI washing risks more than wasted spending. The buyer may also assume legal or reputational exposure created by a vendor's conduct. Agency assessment has moved beyond routine procurement hygiene. It now belongs within risk management because the supplier's claims and methods can affect the client.
How do you vet an agency that says it uses AI?
Use five questions during the first call. A failure to answer any one of them should be treated as useful evidence. The questions come from the FTC's allegations, reversed into a due-diligence process.
| The question | A real answer sounds like | The red flag |
|---|---|---|
| What does the AI do and what does a human do, step by step? | A named workflow with stages | "Our proprietary AI handles everything" |
| Can I watch the workflow run live? | A screen share on your account | A slide of tool logos |
| Which senior person touches my work? | A name, a role, an entry point | "Our team," no names |
| One case study from the last 12 months? | A dated result with the metric named | Case studies predating 2024, or none |
| What happens when the output is wrong? | A correction process with an owner | A pause, then a pivot to pricing |
First, ask for a step-by-step account of what the AI performs and what people perform. Then request a live demonstration using your account instead of accepting a slide filled with tool logos. Ask which senior person will handle the work, including that person's name and role. Request one case study from the last 12 months that includes both a date and a number. Finally, ask what happens when an output is wrong and who takes responsibility for correcting it.
Agency buyer guides released this year describe additional warning signs. They cite claims about fully autonomous systems and refusals to demonstrate workflows. They also warn about promises of undetectable AI content. Agencies produced those guides about their competitors, so they should be read as evidence of buyer concern rather than independent research.
Even so, the shift around detection is worth noting. Undetectable AI content could be promoted as a benefit in 2025. It became a potential liability as soon as the watermark arrived. A promise tied to secrecy now raises questions about disclosure, authorship and whether the client receives the judgment described in the proposal.
One further test is free. Ask the same operational question twice, with a week between the conversations. Changing explanations about who performs the work point to the same inconsistency that buyers punish suppliers for everywhere. McKinsey's 2026 data identifies inconsistency as the leading reason B2B buyers leave a supplier. Repetition can expose a weak process faster than another presentation can.
What separates a real AI operation from a subscription markup?
The distinction rests on how the final result is created. In a genuine operation, AI may prepare material that senior judgment shapes into a usable outcome. A subscription markup offers access to the drafting tool while charging as though the tool itself supplied the expertise. Research now shows how wide the performance gap can be.
Content Marketing Institute's 16th annual B2B survey, conducted with MarketingProfs among more than 1,000 B2B marketers, found that 9 in 10 already use AI to create content. Fewer than 4 in 10 report that it improved performance.
That gap explains the commercial issue. Access to the tools is widespread, while successful results remain much less common. A client should therefore avoid paying an agency merely for software access. The same subscription may be available directly for $20 a month.
The agency fee should purchase judgment. That includes deciding what deserves publication and rejecting material that cannot survive verification. It also means selecting the claim that can be supported and determining which number belongs at the front of the argument. If a process demonstration reveals software the client could operate alone, followed only by an invoice, the agency is selling a markup instead of a meaningful partnership.
A useful test can be expressed in one sentence: ask what the agency deletes. A real operation discards weak drafts. It removes statistics that cannot be verified. It also declines assignments beyond its area of competence. A markup tends to send whatever the model produces because its commercial value depends on volume rather than judgment.
Clients can apply the same logic to their invoices. An agency may spend between $20 and a few hundred dollars a month on a collection of writing tools. Suppose the monthly retainer is 50 times the software cost. If the agency cannot identify the senior person or demonstrate the workflow, the premium requires scrutiny. The same applies if it cannot produce a dated result. Under those conditions, the difference between the software bill and the retainer is effectively the price attached to the word AI in the presentation.
A genuine judgment layer leaves evidence in the finished work. Statistics include their sources. Claims remain defensible after a week of examination. The output also shows that somebody declined to publish material that could have been sent. Tool prices appear on rate cards. The value of judgment has to be established through evidence.
How should an honest agency respond to all this?
Disclose the machine and put a human name behind the judgment. Agencies that endure this period of audits will treat transparency as part of their positioning. They will make their process visible before a regulator or client forces the issue.
The disclosure playbook contains four actions, and none requires a lawyer. State which models and tools are used in the workflow. Put that information in the proposal instead of hiding it under an NDA. Identify the people who guide and edit the machine, along with their roles. Write a correction policy before the first mistake occurs. Keep an updated inventory of claims so that every number on the website has a date and source.
Each action can take an afternoon. Taken together, they can turn the 2026 enforcement wave into a competitive barrier. A vendor that checks its own claims has less reason to fear outside review. Competitors built on AI washing cannot copy that degree of transparency without exposing how their own service is produced.
We have a direct interest in this standard. Moxie Digital is an AI-powered content studio, which places it inside the category now examined by the FTC and Reddit, as well as by watermark detection. This article states the standard we invite buyers to apply to us.
Our responses to the five questions are public. The workflow is documented. AI supports drafting and research. Two named operators, rather than junior staff, guide and edit every item that reaches a client. Apply the five questions to us. Apply them to every other vendor under consideration.
The uncomfortable conclusion is simple. When a vendor's sales pitch fails the vetting checklist it recommends, the failure belongs to the pitch. Blaming the checklist only confirms that the promised transparency cannot withstand examination.
Frequently Asked Questions (FAQ)
What is AI washing in simple terms? AI washing occurs when a business exaggerates the work artificial intelligence performs in its product or service. The purpose is to secure business that the technology itself has not earned. Regulators classify the conduct as a deceptive-claims issue in the same category as greenwashing.
Is AI washing illegal? Overstated AI claims can breach consumer-protection law. The FTC has run Operation AI Comply since September 2024. Its settlements include the March 2026 proposed Air AI order, which carries an $18 million judgment that is largely suspended and would prohibit the marketing of business opportunities.
How do I check if an agency's AI claims are real? Ask five questions. Request a step-by-step division of work between people and AI, followed by a live demonstration. Ask for the name of the senior person assigned to the account. Seek one case study from the last 12 months that includes a date and number. Then examine the process for correcting an output when it is wrong.
Does AI-generated text carry a watermark? Some AI-generated writing now contains one. Anthropic announced on 11 August 2026 that new Claude models place invisible watermarks in generated text to meet the EU AI Act's transparency requirements. Extensive editing or combining the material with human writing can weaken the signal. Older models are scheduled to receive the capability over time.
If you want these five questions applied to the claims made by your current agency, that audit is the work we do. You retain the findings regardless of the result.