TL;DR
This paper presents a mixed-methods audit of "AI-washing"—the deceptive exaggeration of AI capabilities—within the German startup ecosystem. By analyzing 100 startups and interviewing 63 experts, the authors reveal that nearly 40% of these companies market ethically concerning applications under the guise of "ordinary" AI, while simultaneously eroding consumer trust through unsubstantiated performance claims and a lack of transparency.
What Problem Does This Paper Address?
In the current "gold rush" of artificial intelligence, the term "AI" has become a powerful magnet for venture capital and consumer attention. However, this hype has birthed a pervasive issue: AI-washing. Much like "greenwashing" in environmentalism, AI-washing involves companies exaggerating, misrepresenting, or outright fabricating the role of AI in their products to capture market value.
The authors identify a critical gap: while we know AI-washing exists, we lack empirical data on how it manifests in marketing copy, why founders feel pressured to do it, and how it impacts audience perception. This is particularly vital in the context of new regulations like the EU AI Act, where deceptive marketing might be used to camouflage high-risk applications (like surveillance or autonomous medical decisions) as benign consumer tools.
Methods
The researchers employed a two-pronged mixed-methods approach:
- Scoping Study: They selected 100 German AI startups and audited their public-facing marketing materials (websites, social media, press releases). To do this systematically, they developed a seven-dimension codebook ($D_1$–$D_7$) to score everything from claim specificity to ethical risk.
- Expert Interviews: They conducted 63 interviews (53 written, 10 oral) with a diverse group of stakeholders, including AI researchers, marketers, investors, and startup employees. This provided an "inside look" at the organizational pressures—such as investor demands for "AI-first" branding—that drive exaggeration.
The study is notable for its "bottom-up" codebook development, which moved from qualitative observations of recurring patterns to a formalized scoring system.
Datasets
The study utilized a custom-built dataset of 100 German startups sampled from the Startup-Atlas and the AI Startup Landscape (maintained by the appliedAI Institute). The sample was filtered for IT/Communication domains and pre-IPO status to capture the segment of the market most susceptible to funding-related pressures.
Additionally, the researchers gathered qualitative data from 63 expert participants, primarily AI-literate professionals (47% in technical roles like ML or data analytics), to gauge how the "most informed" audience perceives these claims.
Key Insights and Ideas
- Technological Solutionism: Many startups are "shoehorning" AI into tasks that could be handled more reliably by simple, deterministic rules. This often results in higher failure rates and "hallucinations" in contexts where accuracy is paramount.
- The "Wrapper" Trap: A significant portion of the market consists of "wrappers"—products that simply provide a UI for third-party APIs (like OpenAI) but market themselves as proprietary, "groundbreaking" technology.
- Regulatory Camouflage: Perhaps the most concerning finding is that AI-washing is often used to reframe high-risk surveillance or biometric tracking as "smart" features or "personalization," effectively bypassing the scrutiny these technologies would otherwise attract from regulators and the public.
- Persuasion Knowledge: The study found that once consumers encounter one instance of AI-washing, they develop a "skepticism reflex," leading them to distrust all AI-related marketing. This suggests that AI-washing creates a "market for lemons" that harms honest actors in the field.
Key Results
- Vague Claims: 59% of startups described their AI only in buzzwords or superficial terms ($D_1$).
- Lack of Evidence: 43% of startups made AI capability claims with no publicly inspectable evidence ($D_3$).
- The "Infallibility" Myth: 76% of startups acknowledged no limits, edge cases, or accuracy bounds in their marketing ($D_7$).
- Ethical Red Flags: 38% of the sample was flagged for moderate or serious ethical concerns ($D_6$). Notably, there was a positive correlation ($\rho = +0.41$) between high AI-washing scores and ethical risks.
- Trust Erosion: Participants who had previously encountered AI-washing were significantly more skeptical of the label "AI-powered" (median trust score of 1 vs. 3 on a 5-point scale).
Limitations and Future Work
The authors acknowledge several limitations:
- Geographic Scope: The audit was limited to German startups; AI-washing patterns may differ in other regulatory environments (e.g., the US or China).
- Public Record Only: The scores are based on what firms choose to show publicly, not a deep-dive audit of their internal codebases or private investor decks.
- Sample Size: While the interview pool was expert-heavy, it was relatively small and may not represent the "average" consumer's level of AI literacy.
Future Work: The authors propose that their codebook be adopted by regulators and investors as a low-cost "first-pass" screening tool. They also suggest that future research should look at how mandatory disclosure requirements (similar to nutrition labels or financial disclosures) might mitigate the prevalence of AI-washing.
Summary updated