What Is "AI Washing"? How to Identify Vendors Who Are Overstating Their AI Capabilities

The economic incentive is straightforward: companies that can credibly claim AI capability command higher valuations, win more enterprise deals, and attract better talent. In a market where buyers and investors are enthusiastic about AI, the commercial premium for AI capability is significant.

The definitional ambiguity of 'AI' makes washing easy. Almost any software that uses statistics, rules-based logic, or data-driven decision-making can be described as 'using AI' without being technically false. A scoring model from 2015 can be relabelled an 'AI-powered risk engine.' A keyword search system can become 'AI-driven discovery.' A rule-based chatbot becomes an 'intelligent virtual assistant.'

The consequences of AI washing for enterprise buyers are real: purchasing products on the basis of AI capabilities that do not exist, setting expectations for automation or accuracy that the product cannot meet, and making strategic decisions based on a vendor ecosystem that is less AI-capable than assumed.

Several indicators should raise scepticism in AI product evaluations.

Vague capability claims without specifics: 'powered by AI,' 'AI-driven insights,' 'intelligent automation' without specification of what the AI actually does, what model it uses, or what data it was trained on. Genuine AI capabilities can be described specifically.

Performance claims without methodology: accuracy rates, efficiency improvements, and other performance metrics without explaining how they were measured, on what data, and compared to what baseline. As discussed in the context of overfitting, metrics from training data or cherry-picked scenarios are not reliable indicators of real-world performance.

Resistance to technical due diligence: vendors who cannot answer specific questions about their AI architecture, training data, evaluation methodology, and failure modes may not have the capability they claim. Legitimate AI providers answer these questions readily because they have done the work.

Recent 'AI transformations': companies that added 'AI' to their product name or marketing in 2022-2024 without corresponding changes to their underlying product deserve additional scrutiny.

A small set of specific questions reliably distinguishes genuine AI capability from washing.

What is the underlying model? A specific answer (GPT-4 via Azure OpenAI, fine-tuned LLama 3, a proprietary model trained on specific datasets) indicates genuine AI implementation. A vague answer ('our proprietary AI') without technical specifics warrants follow-up.

What data was it trained on, and how is it evaluated? For custom models, this question should produce specific answers about training dataset size, composition, and evaluation methodology. For products built on top of foundation models, the question shifts to how the product is configured, grounded, and constrained.

What happens when it is wrong? Every AI system produces incorrect outputs. A vendor who cannot describe their failure modes, error rates, and mitigation approach either has not characterised their system adequately or is avoiding the question.

Can we run a pilot on our data before committing? Genuine AI capability holds up under pilot evaluation. Washing is exposed when the product performs significantly worse on a buyer's actual data than on the vendor's demonstration.

Regulatory pressure on AI washing is increasing. The EU AI Act includes provisions around transparency and accuracy of AI claims, and advertising regulators in the UK (ASA) have begun applying existing misleading advertising rules to AI capability claims.

The Financial Conduct Authority has been clear that regulated firms cannot outsource regulatory responsibility to AI vendors: if you deploy an AI system in a regulated context, you are responsible for its performance and compliance regardless of what the vendor claimed. This regulatory accountability creates an additional incentive for rigorous AI procurement due diligence.

Boards approving significant AI vendor contracts should ensure that procurement processes include technical due diligence proportionate to the strategic and regulatory significance of the deployment, and that vendor performance commitments are contractually backed.