# AI Washing in Marketing Software: How to Test a Tool Before You Buy

> AI washing is easy to miss in a polished demo. Here's what it looks like in marketing software, a 15-minute vendor test, and why a real trial matters more.

- Source: https://calven.ai/resources/ai-for-pmm/ai-washing-marketing-software
- Author: David Kolinek
- Published: 2026-09-19
- Tags: AI washing, AI for PMM, vendor evaluation, buyer's guide, product marketing software

---

AI washing is claiming a product uses artificial intelligence, or more of it than it really does, to win customers or investors. It's the software cousin of greenwashing. In marketing tools it usually takes one of three shapes: rules relabeled as AI, a chat box on someone else's model, or people doing the work behind a screen the vendor calls AI.

Every vendor demo this year says "agentic," and the word tells you nothing about what the product does on a Tuesday with nobody watching. Finding out takes fifteen minutes on the call and two weeks with the product on your own.

## Is AI washing illegal?

No statute uses the term, but existing law covers the conduct. The SEC charges false AI statements to investors as securities fraud, and the FTC treats false AI claims to customers as deception. Launching [Operation AI Comply](https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdown-deceptive-ai-claims-schemes) in September 2024, FTC Chair Lina Khan said "there is no AI exemption from the laws on the books."

- March 2024. The SEC [settled with Delphia and Global Predictions](https://www.sec.gov/newsroom/press-releases/2024-36), investment advisers that claimed AI they weren't using. $400,000 in penalties.
- September 2024. Operation AI Comply opened with five FTC actions, including DoNotPay, "the world's first robot lawyer," which the FTC says never tested its output against a human lawyer.
- January 2025. The SEC [settled with Presto Automation](https://www.sec.gov/enforcement-litigation/administrative-proceedings/33-11352-s), which hid that a third party ran its speech recognition and, the SEC found, claimed its AI "eliminated the need for human order-taking" when most orders needed a person.
- April 2025. The SEC [sued the founder of Nate](https://www.sec.gov/enforcement-litigation/litigation-releases/lr-26282), alleging contract workers entered by hand the purchases its AI supposedly completed "without human involvement."
- August 2025. The FTC [finalized its order against Workado](https://www.ftc.gov/news-events/news/press-releases/2025/08/ftc-approves-final-order-against-workado-llc-which-misrepresented-accuracy-its-artificial), whose AI detector claimed **98 percent accuracy** and, per the FTC, scored **53 percent** on general content.
- August 2026. The FTC [finalized orders against Cox Media Group and two other firms](https://www.ftc.gov/news-events/news/press-releases/2026/08/ftc-finalizes-orders-cox-media-group-two-other-firms-settling-charges-they-deceived-customers-about) over "Active Listening," a marketing service sold as AI that, per the FTC, "wasn't based on voice data." The firms pay $930,000 between them.

Read as a buyer, the cases sort into three failures: hidden humans (Presto, Nate), an undisclosed third-party model (Presto) and untested accuracy (DoNotPay, Workado). The orders bind the seller, so your protection is a test that catches all three before the contract. (None of this is legal advice.)

## How do you test an AI tool?

Bring one real signal from your own work, a competitor's launch post or the transcript of a lost deal, and ask the vendor to turn it into an updated artifact while you watch. If deal material can't leave before a security review, use a public signal such as a competitor's pricing page. A vendor who won't run it live on a public signal has already answered.

| Check | What to say on the call | What passes |
|---|---|---|
| **1. Signal to artifact** | "Here's a competitor launch from this week. Show me the battle card that changes because of it." | The card updates and shows what changed. <br />**Red flag:** it ends at an alert, a summary or a mock-up. |
| **2. Three citations** | "Open the source for this line. And this one. And this one." | Each opens a real page or transcript. <br />**Red flag:** a vendor cache, or nothing. |
| **3. A run nobody prompted** | "Show me something it did last week that no one asked for: the trigger, the log, who approved it." | A trigger, a run record and an approval step you can switch off. <br />**Red flag:** a scheduled prompt, or a person on their side. |
| **4. Model and data** | "Which model runs this, whose is it, and does my data train anything?" | A direct answer. <br />**Red flag:** "proprietary AI" with nothing named. |

The call shows you the product at its best, run by someone who knows every shortcut. It can't show you what it's like without them.

## Why should you insist on a real free trial?

Because a trial is the only test the vendor can't steer. A demo runs on prepared data. A two-day proof of concept where you upload one file shows the happy path. Neither tells you how much of the work the AI does and how much stays with you.

A real trial looks like this:

- **One to two weeks.** Long enough for real signals to arrive on their own, not just the ones you feed it.
- **Your own data.** Your messaging, your competitors, your call transcripts.
- **No vendor help.** No sales engineer configuring it, no onboarding call steering what you try. Your team uses it the way it would after the contract.

While it runs, count the manual actions: every signal you had to paste in, every run you had to trigger, every output you had to rewrite or carry into another tool yourself. Count edits too. Translation researchers have scored machine output by counting a person's edits [since 2006](https://aclanthology.org/2006.amta-papers.25), and the same logic holds: if fixing the battle card takes as long as writing it, the tool saved you nothing. At the end, paste the same signals and your messaging doc into ChatGPT or Claude. The tool should win on sources you can open, on what it updates, or on hours.

A vendor who won't give you that trial is the biggest red flag in this whole guide. A "guided pilot" with their team at the keyboard tells you what their team can do, not the product.

This test checks whether the AI is real. Whether the product is the right one is a separate question, covered by the [six criteria in the AI product marketing platform buyer's guide](/resources/ai-for-pmm/ai-product-marketing-platform-buyers-guide) and the [CI tools checklist](/resources/competitive-intelligence/how-to-evaluate-competitive-intelligence-tools).

## Would ChatGPT and your messaging doc do the same job?

For plenty of tools sold as AI, yes, which is why the baseline belongs in every trial. An AI wrapper is a prompt on someone else's model with no data, memory or write access of its own. Selling it as more than that is where AI washing starts. [AI tools for product marketing](/resources/comparisons/ai-tools-for-product-marketing) covers where assistants stop and what separates an agent from a scheduled prompt.

The subtler failure is real AI bolted onto a point tool that stops at an alert. The detection works, but you still carry the signal into the battle card, the messaging and the enablement deck, which leaves the PMM as the [integration layer of the stack](/resources/product-marketing/product-marketing-tech-stack).

## Which vendor phrases should make you suspicious?

None proves AI washing on its own; each tells you what to check.

- "Actionable AI insights" or "agentic," when the demo ends at a notification. Run check 1.
- "Proprietary AI," with no model named. Check 4.
- "Human in the loop," with no answer to whose human. Check 3.
- "Fully autonomous," with no setting to keep approvals on.
- "We'll set it up for you," when you ask for a trial. Insist on running it alone.
- A recorded demo, or a planned feature shown as if it runs.

Calven, an [AI product marketing platform](/resources/ai-for-pmm/ai-product-marketing-platform) whose agents cover research and competitive intelligence, audience, and positioning and messaging from one shared knowledge base, should pass the same test on the same terms, including a 14-day trial on your own data. A product that does what it claims can show you. One that can't will show you slides.