Most "AI-powered" products are API wrappers. Some don't even use AI. The SEC has charged companies for fabricating AI capabilities. The FTC has filed over a dozen enforcement actions. This is the enterprise version of putting "blockchain" on everything in 2017. This guide covers how to spot AI washing, what the regulators are doing about it, and how to evaluate whether a vendor's AI claims are real.
What AI Washing Is
AI washing is the practice of marketing a product or service as AI-powered when the AI is nonexistent, trivial, or misrepresented. It takes three forms.
No AI at all. The product uses rule-based automation, manual processes, or human labor behind an interface that claims to be AI. Nate Inc. told investors its shopping app processed transactions with AI. The actual automation rate was zero. Human workers in the Philippines completed every purchase manually. The founder raised $42 million before the SEC and DOJ filed parallel fraud charges in April 2025.
Third-party AI relabeled as proprietary. The product calls an external API (OpenAI, Anthropic, Google) and presents the output as its own technology. Presto Automation described its drive-thru ordering product as proprietary AI. The SEC found that, for a period, all deployed units ran on a third-party's speech recognition. When Presto did build in-house technology, 70% of orders still required human intervention.
Basic automation relabeled as AI. The product performs rule-based operations (if/then logic, keyword matching, template filling) and markets them as "AI-powered insights" or "intelligent automation." Arvind Narayanan and Sayash Kapoor at Princeton call this "AI snake oil": technology that does not work as advertised and probably cannot work as advertised.
The term comes from greenwashing
Greenwashing is making unsubstantiated environmental claims to attract ESG-conscious buyers. AI washing is the same playbook for technology buyers. Both exploit the gap between what a label promises and what verification costs. Both attract regulatory attention when the gap becomes fraud.
Documented Cases
These are not allegations from competitors or anonymous blog posts. Every case below involves formal charges from the SEC, DOJ, or FTC, or public reporting from credible outlets.
| Company | Claim | Reality | Enforcement |
|---|---|---|---|
| Nate Inc. | AI-automated shopping app with 93-97% automation rate | 0% automation. Human workers in the Philippines processed every transaction. | SEC + DOJ fraud charges (April 2025). Up to 20 years prison. $42M+ in investor losses. |
| Presto Automation | Proprietary AI drive-thru ordering | Third-party speech recognition + 70% human intervention rate | SEC cease-and-desist order (January 2025). First AI washing case against a public company. |
| Delphia (USA) Inc. | AI and machine learning to analyze client data for investment advice | Never actually used AI to analyze client data | SEC penalty: $225,000 (March 2024) |
| Global Predictions Inc. | AI-driven investment strategies | Misrepresented the role of AI in its investment process | SEC penalty: $175,000 (March 2024) |
| Air AI | Conversational AI that replaces full-time human sales agents | AI was unavailable or nonfunctional for many customers. Basic calling and scheduling failed. | FTC complaint (August 2025). Third AI washing case under current administration. |
| Amazon Just Walk Out | Automated checkout using AI cameras and sensors | Over 1,000 workers in India manually reviewed 70%+ of transactions | No enforcement action. Amazon phased out the technology from Fresh stores. |
The pattern: humans behind the curtain
Nate Inc. is the clearest example of outright fraud. Albert Saniger told investors his app used AI to complete online purchases with a single tap. The DOJ indictment alleges the actual AI success rate was zero. Every transaction was processed by human workers overseas, and Saniger directed employees to conceal this from investors. The company raised over $42 million between 2019 and 2022 before ceasing operations in January 2023. Investors lost tens of millions.
Presto Automation is the subtler case and the more common pattern in enterprise software. The company did build some AI technology. But its public statements described the product as eliminating human order-taking when it did not. Over 70% of orders processed through the in-house version of Presto Voice required human agents, primarily located in the Philippines and India. The SEC issued a cease-and-desist order in January 2025, making Presto the first public company charged with AI washing.
Amazon's Just Walk Out technology followed the same mechanical turk pattern. Marketed as AI-powered automated checkout, the system relied on over 1,000 workers in India to manually check roughly 70% of transactions. Amazon did not face regulatory action but quietly removed the technology from most Fresh stores.
The pattern: humans behind investment advice
The SEC's first AI washing enforcement actions targeted investment advisors. In March 2024, the SEC charged Delphia and Global Predictions simultaneously. Delphia claimed it used AI and machine learning to analyze client data. It never did. Global Predictions made similar claims about AI-driven investment strategies. The penalties were small ($225,000 and $175,000), but the signal was clear: the SEC considers false AI claims a form of securities fraud.
The API Wrapper Problem
Not all AI washing involves fraud or human labor. The most common form is simpler: a company builds a nice interface on top of a ChatGPT or Claude API call, charges $30-100/month, and describes the result as proprietary AI technology. No custom model. No fine-tuning. No training data. No inference infrastructure.
The product works until the underlying provider changes pricing, rate limits, or model behavior. When OpenAI deprecates an API version, every wrapper built on it breaks simultaneously. When Anthropic changes its safety filters, downstream products change behavior in ways their customers didn't choose. The wrapper company has no control over its core technology because it owns none of it.
A Google VP warned in February 2026 that two types of AI startups face extinction: those wrapping thin intellectual property around foundation models, and those competing directly with the model providers themselves. A 2025 report documented a 2.5x year-over-year increase in Series A shutdowns, with AI wrappers heavily represented.
No model, no moat
If your vendor's product is a UI on top of a ChatGPT API call, the same product can be built in a weekend. When GPT-5 ships with the feature built in, the wrapper has no reason to exist.
Price dependency
A product charging $50/month that costs $0.02/call in API fees has healthy margins. Until the API provider raises prices 3x. The wrapper cannot negotiate, switch models, or optimize costs because it doesn't control the inference layer.
Behavior dependency
When the underlying model updates, the wrapper's behavior changes without warning. Safety filter changes, output format changes, and capability changes all propagate to customers. The wrapper company learns about these changes the same way its customers do.
How to distinguish a wrapper from infrastructure
Wrappers are not inherently bad. Some add genuine value through workflow integration, domain-specific prompting, or user experience. The problem is when they claim to be something they are not. A wrapper that honestly says "we provide a better interface to Claude for legal workflows" is a legitimate product. A wrapper that says "our proprietary AI technology analyzes legal documents with enterprise-grade accuracy" when it's sending prompts to Claude with no fine-tuning or custom training is AI washing.
The test is straightforward: ask the vendor what happens if their upstream API provider disappears tomorrow. If the answer involves rebuilding from scratch, it's a wrapper. If the answer involves switching to a different model because they own the inference, routing, and evaluation layer, it's infrastructure.
Regulatory Enforcement
AI washing enforcement is accelerating across multiple federal agencies. This is not a partisan issue. Cases filed under the previous administration have continued under the current one.
SEC: Securities fraud framing
The SEC treats false AI claims as a species of securities fraud when they affect investor decisions. Its enforcement timeline:
- March 2024: Simultaneous actions against Delphia ($225K penalty) and Global Predictions ($175K penalty) for false AI claims in investment advisory marketing.
- January 2025: Cease-and-desist order against Presto Automation for misleading AI product disclosures. First case against a public company.
- February 2025: Created the Cybersecurity and Emerging Technologies Unit (CETU), a dedicated enforcement unit focused on AI-related misconduct.
- April 2025: Parallel SEC and DOJ charges against Albert Saniger (Nate Inc.) for raising $42M+ on fabricated AI claims. Criminal charges carry up to 20 years.
FTC: Consumer protection framing
The FTC approaches AI washing through consumer protection law. Its Operation AI Comply initiative has produced 12+ enforcement actions since 2024. Cases target companies making deceptive claims about what their AI products can do, particularly around replacing human workers, generating income, and automating business processes.
The Air AI case (August 2025) is representative. The company claimed its conversational AI could replace full-time human sales agents with no ramp-up time. The FTC alleged the AI was either unavailable or couldn't perform basic functions for many customers. The case remains pending.
The FTC also charged Ascend (operators William Basta and Kenneth Leung) with defrauding consumers of more than $25 million through false claims that AI tools could help them earn passive income from online storefronts.
Enforcement survives administration changes
Both the SEC and FTC have continued AI washing enforcement under the current administration. The SEC's CETU unit was created in 2025. The FTC filed new cases through early 2026. The White House AI Action Plan directed agencies to review prior investigations but did not stop new enforcement. If your company makes AI claims, assume regulators are watching regardless of the political cycle.
DOJ: Criminal prosecution
The Nate Inc. case introduced criminal prosecution for AI washing. The DOJ indictment charges Albert Saniger with securities fraud and wire fraud, each carrying a maximum sentence of 20 years. This is the strongest signal yet that fabricated AI claims can result in prison time, not just fines.
How to Evaluate AI Vendor Claims
Whether you are a procurement team evaluating enterprise software, an engineering leader choosing infrastructure, or a buyer evaluating any product with "AI-powered" in the description, these questions separate real AI from marketing.
The four questions
1. What happens if OpenAI raises prices 3x?
This tests API dependency. A vendor with its own inference infrastructure can switch models, optimize routing, or absorb the change. A wrapper passes the cost to you or shuts down. The answer reveals whether the vendor controls its own technology stack.
2. Show me a benchmark on a public dataset
This tests for measurable results. Real AI products have quantifiable performance: F1 scores, accuracy percentages, throughput numbers, latency measurements. 'Our AI provides intelligent insights' is not a benchmark. '0.73 F1 on SWE-Bench' is.
3. What is your inference infrastructure?
This tests for real engineering. Does the vendor run its own models? Use fine-tuned models? Route between models based on task difficulty? Or does it send every request to a single third-party API? The depth of the answer tells you the depth of the technology.
4. Can I see latency and cost metrics?
This tests for transparency. Vendors with real infrastructure know their numbers: p50 and p99 latency, cost per request, throughput under load. Vendors without infrastructure can't answer because they don't control the variables.
What good answers sound like
A vendor with real AI infrastructure will give you specific numbers. "Our router classifies prompt difficulty in ~430ms and routes to the cheapest model that can handle it. Hard tasks go to Opus, easy tasks go to Haiku. Cost savings are 40-70% depending on your workload mix." That is a verifiable claim. You can test it.
A vendor without real infrastructure will give you adjectives. "Our proprietary AI engine leverages advanced machine learning to deliver intelligent insights." That is not testable. There is no benchmark, no metric, no architecture to verify.
| Real claim | AI washing equivalent | |
|---|---|---|
| Performance | 0.73 F1 on SWE-Bench search tasks | AI-powered intelligent code search |
| Speed | 10,500 tok/s apply speed, p50 latency 180ms | Ultra-fast AI processing |
| Cost | $0.001/request routing, 40-70% cost reduction vs single-model | Cost-effective AI solution |
| Architecture | Classifier trained on 1M+ coding prompts, routes between 4 model tiers | Proprietary AI engine |
| Transparency | Published benchmarks, open API, metered billing | Enterprise-grade AI platform |
Red Flags in Vendor Conversations
Beyond the four questions, watch for these patterns during vendor evaluations.
'Proprietary AI' with no published benchmarks
If the AI is genuinely differentiated, the vendor benefits from publishing benchmark results. Refusal to benchmark against public datasets usually means the results would be unimpressive or the 'AI' is an API call they don't control.
Refusal to discuss architecture
'We can't share details for competitive reasons' is sometimes legitimate. But if a vendor won't say whether they run their own models, use fine-tuned models, or call a third-party API, the most likely answer is the third option.
Vague claims about 'AI-powered insights'
'Insights' is the word vendors use when they can't describe what the AI actually does. A real product says 'classifies support tickets by intent with 94% accuracy.' An AI-washed product says 'provides AI-powered insights into your support data.'
Demo that won't go off-script
A controlled demo with pre-selected inputs and pre-cached outputs proves nothing. Ask to run the product on your data, with your edge cases, in real time. If the vendor refuses or says it 'needs to be configured first,' the demo may be theater.
No technical team in the sales process
If every question about architecture, benchmarks, or infrastructure is deferred to a follow-up call that never happens, the technical depth may not exist. Real AI companies put engineers in the room because they want to talk about their technology.
Pricing disconnected from usage
Flat-rate pricing on an 'AI product' can mask that the AI component is minimal. If the AI is doing real work (inference, training, search), usage-based pricing is natural because the vendor has real compute costs. Flat pricing on an 'AI' product might mean the AI cost is near zero because the product is mostly UI.
The Measurable Alternative
Real AI infrastructure publishes numbers because the numbers are the product. Not adjectives, not demos, not promises. Metrics that buyers can verify independently.
Morph builds AI infrastructure for coding agents. Every product ships with published benchmarks.
The Router classifies prompt difficulty in ~430ms and routes to the cheapest model that can handle the task. It was trained on millions of coding prompts. Cost savings are 40-70% depending on workload mix. The classifier, the routing logic, and the evaluation methodology are documented.
Flash Compact reduces context by 50-70% at 33,000 tok/s with zero hallucination. Every surviving sentence is verbatim from the input. No paraphrasing, no invented details. The compaction is a strict subset of the original, character for character.
WarpGrep achieves 0.73 F1 on SWE-Bench search tasks. It runs 8 parallel tool calls per turn across 4 turns in under 6 seconds. The benchmark methodology and results are public.
Fast Apply runs at 10,500 tok/s for code application. The model is purpose-built for applying diffs, not a general-purpose LLM repurposed for the task.
None of these are "AI-powered insights." They are infrastructure with published specifications. If the numbers don't hold up, you can measure that yourself.
The vendor test applied to Morph
What happens if Anthropic raises prices 3x? The router shifts traffic to cheaper model tiers automatically, absorbing the change. Show a benchmark on a public dataset: 0.73 F1 on SWE-Bench Pro search tasks. What is the inference infrastructure? Custom models for apply and compaction, a classifier trained on 1M+ prompts for routing, WarpGrep running parallel agentic search. Latency and cost metrics? ~430ms routing, 10,500 tok/s apply, $0.001/request. Published on the benchmarks page.
Frequently Asked Questions
What is AI washing?
AI washing is marketing a product as AI-powered when the AI is nonexistent, trivial, or misrepresented. It ranges from outright fraud (using human labor behind an AI interface) to deceptive marketing (relabeling basic automation as artificial intelligence). The SEC and FTC both treat it as an enforcement priority, with penalties ranging from fines to criminal prosecution.
What are examples of AI washing?
Nate Inc. raised $42 million claiming AI-automated shopping while using human workers overseas (SEC/DOJ fraud charges). Presto Automation claimed proprietary AI drive-thru ordering but needed human intervention for 70% of orders (SEC cease-and-desist). Amazon's Just Walk Out relied on 1,000+ workers in India. Delphia and Global Predictions claimed AI-driven investment strategies that didn't exist (SEC penalties). Air AI claimed its conversational AI could replace sales agents when it couldn't perform basic functions (FTC complaint).
How can I tell if a vendor's AI claims are real?
Ask four questions: (1) What happens if your upstream AI provider raises prices 3x? (2) Show me a benchmark on a public dataset. (3) What is your inference infrastructure? (4) Can I see latency and cost metrics? Vendors with real AI infrastructure give specific numbers. Vendors without it give adjectives.
What is the difference between AI washing and an AI wrapper?
A wrapper builds a product on top of a third-party AI API. This can be legitimate if the wrapper is honest about its architecture and adds real value through workflow integration or domain-specific design. AI washing is deceptive: claiming a wrapper is proprietary technology, claiming basic automation is AI, or hiding human labor behind an AI interface. The difference is honesty about what the product is and how it works.
What regulatory enforcement exists for AI washing?
The SEC has brought enforcement actions since March 2024 and created a dedicated unit (CETU) for AI-related misconduct. The FTC has filed 12+ cases through Operation AI Comply. The DOJ has brought criminal fraud charges against AI washing (Nate Inc., carrying up to 20 years). Enforcement has continued under the current administration, with new cases filed in 2025 and 2026.
How do I evaluate AI infrastructure vendors?
Look for published benchmarks on public datasets, specific throughput and latency numbers, transparent pricing tied to usage, and willingness to discuss architecture. Real infrastructure vendors know their numbers because the numbers are the product. Morph publishes all of them: 10,500 tok/s apply, 0.73 F1 search, 33,000 tok/s compaction, $0.001/request routing. See benchmarks.
Related Resources
AI Infrastructure That Publishes Its Numbers
10,500 tok/s apply speed. 0.73 F1 search accuracy. 33,000 tok/s compaction. $0.001/request routing. Every number is verifiable. No proprietary magic, no vague claims, no adjectives where metrics should be.