AI Washing

AI washing is the practice of exaggerating or misrepresenting how much artificial intelligence a product, service, or company uses, or what that technology can accomplish. It creates a gap between advertised capabilities and actual operation: customers and investors are led to expect intelligence, autonomy, or performance that the underlying system does not deliver.

The term borrows from greenwashing, which describes misleading environmental claims. Gartner analyst Jim Hare used "AI washing" in his July 6, 2017 analysis of enterprise software providers exploiting AI disruption, describing vendors applying the AI label indiscriminately.

How AI Washing Works#

AI washing turns technical complexity into a marketing advantage. Buyers encounter terms such as artificial intelligence, machine learning, and "autonomous" without enough information to connect those descriptions to an operational feature. The misrepresentation can concern the technology's existence, its contribution to a workflow, or the strength of its results.

Common patterns include:

• Capability inflation: Advertising a system as performing expert-level work when its outputs lack evidence supporting that claim.
• Human involvement concealment: Presenting a service as autonomous while people perform substantial parts of the advertised work.
• Performance overstatement: Extending results from selected conditions to a broader audience or task than the evaluation supports.
• Timeline acceleration: Describing planned AI functionality as an existing product capability.

These patterns affect purchasing, investment, and staffing decisions. They also obscure operational requirements: customers need to understand which outputs require review, which tasks the system handles, and how errors affect downstream decisions.

AI Washing and Related Concepts#

Robotic process automation automates repetitive workflows, often through predefined rules. Machine learning learns patterns from data, while symbolic AI uses explicit knowledge and reasoning rules. Accurate product descriptions identify the mechanism and its role. Calling a rule-based system AI becomes misleading when the accompanying claims promise learning or adaptability that the system lacks.

Human-in-the-loop workflows assign people defined review or decision-making responsibilities. Transparent descriptions make that contribution visible. For example, Ultralytics Platform applies AI-assisted data annotation: its YOLO smart annotation feature generates draft labels from model predictions, which annotators can review, adjust, or remove. This identifies both the automated task and the human role.

Hallucination is a model producing incorrect or unsupported content. AI washing concerns representations made about the system. Explainable AI helps interpret model behavior; transparency in AI communicates capabilities, evaluation conditions, and responsibilities. Both support evidence-based descriptions of AI products.

Real-World Examples and Consequences#

Regulators adopted the term in enforcement. SEC Chair Gary Gensler warned companies against "AI washing" in December 2023, and the FTC brought the DoNotPay case below as part of Operation AI Comply, its September 2024 sweep against deceptive AI claims.

Investment adviser Delphia claimed that client data strengthened its AI-driven investment predictions. The SEC found that, from 2019 through 2023, the company lacked the AI and machine learning capabilities it advertised. In a March 18, 2024 settlement, Delphia agreed to a $225,000 civil penalty. Global Predictions separately agreed to a $175,000 penalty over misleading claims that included AI-driven forecasts and its status as the first regulated AI financial adviser.

DoNotPay advertised an AI service as the "world's first robot lawyer," claiming capabilities that could substitute for human legal expertise. The DoNotPay enforcement settlement required $193,000 in monetary relief and notices to eligible subscribers. The FTC finalized the order in January 2025, restricting unsupported claims about substituting for professional services.

Evidence That Supports AI Claims#

A substantiated claim connects a named feature to its inputs, outputs, evaluation, and operating conditions. A model card records intended uses, development details, and measured limitations. Data provenance establishes where evaluation material came from, while independent test data supports checking performance beyond development examples.

In computer vision, the evidence for a claim comes from the advertised task itself. Object detection results can include precision, recall, and mean average precision, together with dataset composition and per-class results. Speed claims require hardware, image size, batch size, and measurement scope.

AI governance connects those records to product approvals and marketing review, so each published claim traces back to a model card, an evaluation, or a documented human role.