PaperlayerView exact revision on arXiv

Source-linked AI summary

Auditing AI-Washing in German Startups: A Mixed-Methods Study of Marketing Claims and Perceptions

A Pranav, Luke Hartmann, Anne Lauscher

arXiv:2610.11788v1cs.CY

TL;DR

AI-washing raises the question of how German AI startups market their systems, what drives those claims, and how audiences receive them. The authors audit 100 startups and interview experts, finding capability exaggeration, reduced consumer trust after prior exposure, and substantial ethical concern in the sample.

  • Problem

    The paper asks what AI-washing looks like in German AI-startup marketing, what drives founders and marketers, and how consumers and domain experts receive those claims.

    Source: Introduction · p. 1

  • Method

    The authors scored 100 German AI startups with a seven-dimension codebook and composite index and conducted 63 expert interviews about motivations and audience perceptions.

    Source: Introduction · p. 1, Codebook · p. 2

  • Results

    38% of the sample were flagged for moderate or serious ethical concern, while startups marketed wrapped, rule-based, or absent AI as autonomous or proprietary and consumers with prior exposure trusted later AI marketing less.

    Source: Introduction · p. 2, Discussion · p. 9

  • Takeaways & Limitations

    The authors recommend structured product-page disclosures of user-data collection and system limits, plus independent certification of commercial AI claims.

    Source: Discussion · p. 9

    The interview sample comprises AI-literate professionals and graduate students rather than a probability sample, so it does not establish how German consumers generally receive AI marketing.

    Source: Interviews · p. 4

Abstract

from arXiv · show

AI-washing is the use of 'AI' as a marketing term where products' claims are exaggerated, likely to capture the hype from investors and consumers. In this paper, we examine how startups market AI and how people perceive those claims. We audited the German startup market, applying an annotation codebook of seven dimensions to 100 startups and conducting 63 expert interviews. Startups exaggerated AI capability in three forms: claims without inspectable evidence, performance numbers without methodology, and claims without acknowledged limits. The capabilities they claimed rarely matched the underlying system, and consumers who had encountered exaggerated claims before withdrew trust from later AI marketing. AI-washing has also been linked to the marketing of unethical products: 38% of the sample were flagged for moderate or serious ethical concern, in domains such as surveillance and automated decisions in regulated fields. Together these findings show how AI-washing happens: founders exaggerate capability to attract investors who reward AI branding they cannot verify, and consumers in unfamiliar domains take the marketing at face value. We recommend that startup AI claims be brought under disclosure requirements, with the codebook offered to regulators, investors, and consumer-protection bodies for routine audits.

Introduction

The introduction frames AI-washing as exaggerated or obscured AI claims encouraged by investor interest and consumer assumptions, and reports capability overstatement, reduced trust, and ethically concerning applications in German startup marketing.

Introduction

  • Investor attention rewards AI branding, while consumers may treat AI as near-magical and people with lower AI literacy may overestimate advertised capabilities.The passage reports that founders presenting products as AI-driven raise capital more easily than those who do not.Source: Introduction · p. 1
  • Startups advertised capabilities beyond what current AI can reliably deliver, often using buzzwords without describing the system or overstating AI’s role in the product.Some claims applied AI to tasks that simpler approaches could handle.Source: Introduction · p. 1
  • Marketing deception can involve claiming nonexistent abilities, reselling another company’s system as proprietary, or exaggerating capability while obscuring what the system does; outsiders may be unable to verify claims.The cited account notes that marketing may omit methods and independent testing.Source: Introduction · p. 1
  • 38% of startups were flagged for moderate or serious ethical concern (19% serious), including surveillance and high-risk or prohibited applications; consumers with prior exposure to exaggerated claims reported lower trust in later AI marketing.The flagged concerns included surveillance, autonomous decisions in regulated domains, and behavioural tracking and manipulation.Source: Introduction · p. 2
  • Earlier research links startup exaggeration to investor expectations and describes AI-washing through inflated capability claims or third-party systems presented as proprietary.Source: Introduction · p. 2

Methods

The study combines a codebook-based audit of public-facing claims from 100 German AI startups with interviews about claim motivations and audience perceptions. Its methods define scoring, gather qualitative evidence, and analyze written and oral interviews, while limiting claims to the observed samples and materials.

Methods

  • The authors audited 100 German AI startups with a seven-dimension codebook and composite index, then interviewed 63 experts to examine claim motivations and audience perceptions.The interviews comprised 53 written responses and ten oral semi-structured sessions.Source: Methods · p. 2
  • Startups were drawn from two public directories and had to be German-registered, pre-IPO firms with an AI claim and a substantive functional website.The directory search included random subsampling to reduce selection bias.Source: Sample · p. 2
  • The codebook emerged from iterative review of about 30 startups, formalized five recurring patterns into seven dimensions, and achieved macro-averaged Cohen’s κ of 0.82 on an overlap set.The annotators independently coded 50 startups each, with a separate 10-startup overlap set.Source: Codebook · p. 2
  • AI meant a function using learned models rather than fixed rules, while D2 rated whether the model was proprietary, adapted or aggregated, a wrapped API, or undetectable.The 1–5 composite combined dimension scores; ethical concern could raise it, while strong functionality evidence or disclosed limitations could lower it.Source: Codebook · p. 2
  • Annotators reviewed websites, available demos or trials, social media, and external sources, scored only public material, and separately recorded qualitative claims and ethical concerns.Manual observation of AI-generated website text or imagery was recorded but excluded from the composite index.Source: Annotation procedure · p. 3, Codebook · p. 3
  • The startup sample covered varied sectors and product types, but the interview findings describe AI-literate professionals and graduate students, not a probability sample of German consumers.The authors also limit generalization beyond Germany and note that annotations reflect public materials, demos, and trials rather than long-term product use.Source: Sample composition · p. 3, Interviews · p. 4
  • Spearman rank correlations and Mann–Whitney U tests were used for dimension, composite, and group comparisons; all group comparisons were exploratory.Dimension directions were aligned so positive correlation with the composite always indicated greater AI-washing severity.Source: Definitions · pp. 3–4
  • Written interviews asked about AI familiarity, trust in the “AI-powered” label, perceived AI capabilities, and encounters with AI-washing; oral interviews covered organizational drivers and remediation.The written instrument had 53 complete responses; oral sessions lasted 30–60 minutes and were independently open-coded by two authors.Source: Interviews · p. 4

Current AI Cannot Deliver

Startups often marketed AI capabilities that their systems could not reliably deliver, using vague descriptions, exaggerated AI prominence, and claims poorly matched to product function. The examples include non-proprietary systems, relabelled workflows, and applications where simpler deterministic approaches would be more reliable.

Current AI Cannot Deliver

  • 59% described AI only in buzzwords or superficial terms, 35% had no actual AI or relied on an LLM wrapper without proprietary contribution, and 40% overstated AI’s product role.Wrapper-based AI was associated with greater overstatement of AI’s role (ρ(D2, D5) = +0.51, p < 0.001, n = 99).Source: Current AI Cannot Deliver · p. 4
  • 31% had unclear or technically unspecific AI descriptions, and business-only founders were significantly more likely than technical founders to overstate capabilities and AI prominence.The reported differences were U = 287, p = 0.003 for overstatement and U = 76, p = 0.002 for marketing prominence.Source: Current AI Cannot Deliver · pp. 4–5
  • Some startups applied AI to tasks that deterministic systems could handle more reliably, while current LLMs cannot guarantee zero hallucinations.The paper gives deterministic retrieval over the same source corpus as an alternative for no-hallucination, fully traceable results.Source: Current AI Cannot Deliver · p. 5
  • 72% of the high-composite group lacked proprietary AI, while 60% of startups marketed AI prominently without owning a proprietary AI system.The high-composite group included startups with no AI, third-party wrappers, or combinations of existing tools.Source: Current AI Cannot Deliver · p. 5
  • When human work is sold as AI, the workers become invisible to customers and their labour falls outside the wage and protection norms for software development.The paper also describes startups selling third-party models they do not own as AI.Source: Current AI Cannot Deliver · p. 5
  • Examples include a data-aggregation workflow relabelled as AI reporting, voice control that failed on German-accented speech, and “Digital Employees” promised to think.The automotive startup’s claim was coded D1 = 2, D2 = 3, D5 = 2.Source: Current AI Cannot Deliver · pp. 5–6

Finding 2: Startups Exaggerate AI Capabilities And Customers Are Losing Trust

Across the audited startups, capability claims exceeded what products demonstrably did, with exaggeration taking the form of unverifiable claims, unsupported numbers, and undisclosed limits. Interviews trace these claims to investor and internal branding pressures; prior encounters with exaggeration aligned with lower trust in later AI claims.

  • Startups exaggerated capability through claims without inspectable evidence, numerical claims without methodology, and claims that acknowledged no limits.Disproportionate AI marketing weight was the strongest correlate of the composite AI-washing score (ρ(D5, composite) = +0.57, p < 0.001).Source: Finding 2: Startups Exaggerate AI… · p. 6
  • 43% of startups offered no publicly inspectable evidence for AI capabilities, while only 2% provided evidence at the codebook’s highest tier.Commonly offered logos, anonymised case studies, and controlled demos were seller-controlled and did not let outsiders audit the claimed capability.Source: Finding 2: Startups Exaggerate AI… · p. 6
  • 79% of quantitative claimants (49 of 62) showed no substantiation, none were fully transparent, and weaker substantiation accompanied higher AI-washing severity (ρ(D4, composite) = +0.25, p = 0.05, n = 62).Examples included a 96% saving claim without a disclosed comparison baseline and an “industry average” baseline without a defined industry, measurement, or time window.Source: Finding 2: Startups Exaggerate AI… · p. 6
  • 76% of startups named no edge cases, accuracy bounds, data dependencies, or human-review requirements, and only 7% reached the top disclosure tier.No limitations were disclosed by 72% outside the high-composite group and 82% within it; one ostensibly fully automated product relied on a daily human review team.Source: Finding 2: Startups Exaggerate AI… · pp. 6–7
  • 77% of sites flagged for apparent AI-generated marketing material (10 of 13) were in the high-composite group, compared with 33% of unflagged sites (ρ(AI-gen flag, composite) = +0.25, p = 0.011).The paper suggests that superlative and infallibility framing in LLM-generated text may explain this correlation.Source: Finding 2: Startups Exaggerate AI… · p. 7
  • Participants traced exaggeration to investor and public-funder pressure and to management or marketing urging AI branding, while technical developers often did not decide how products were presented.One firm’s website described it as “AI First” while its small development team was still in early planning meetings for an AI feature.Source: Finding 2: Startups Exaggerate AI… · p. 7
  • Participants who had encountered AI-washing reported less trust in “AI-powered” claims than those who had not, despite no significant difference in self-rated AI familiarity.Trust ratings were lower among those reporting encounters (median 1, mean 1.71) than among those without encounters (median 3, mean 2.48; U = 190, p = 0.003).Source: Finding 2: Startups Exaggerate AI… · p. 7

Finding 3: AI-washing Leads To Ethical Risks

AI-washing appeared in products raising ethical concerns, including surveillance, regulated decisions, and behavioural tracking. The study connects these applications to high-risk or prohibited categories and reports practical risks for people subject to them.

  • 38% of the 100 startups marketed applications raising ethical concerns: 19% serious and 19% moderate, all in categories the EU AI Act classifies as high-risk or prohibits outright.The largest subcategories were surveillance, autonomous decisions in regulated domains, and behavioural tracking and manipulation.Source: Finding 3: AI-washing Leads To Ethical… · p. 7
  • 70% of interview participants rated AI’s surveillance capability in the lower two scores of a five-point scale (mean 2.17), and technically trained participants were more sceptical.A participant who worked inside a surveillance startup said its cameras could not support the capabilities the firm tried to sell investors.Source: Finding 3: AI-washing Leads To Ethical… · p. 8
  • Participants were sceptical of AI making regulated decisions: 57% rated it incapable and 32% capable, while 81% rated AI therapy in the lower two scores.Examples included embryo selection and child autism screening paired with an LLM therapy platform; current language models lack the contextual and situational grounding clinical practice requires.Source: Finding 3: AI-washing Leads To Ethical… · p. 8
  • Startups marketed behavioural tracking, manipulation, and outputs based on personal data collected without consent as product features, including undisclosed AI-generated calls and real-time webpage changes.At a trade fair, a startup offered a biometric “psychological assessment” without explaining what would happen to the collected data.Source: Finding 3: AI-washing Leads To Ethical… · p. 9
  • Two workplace AI systems sold as efficiency tools underperformed against their claims, making shifts harder through misallocated staffing and false positives.The systems also placed workers under algorithmic decisions and monitoring during each shift and collected tracking data about them.Source: Finding 3: AI-washing Leads To Ethical… · p. 9
  • Across the three cases discussed, marketing concealed whether applications belonged to high-risk or prohibited regulatory categories, while the people watched by these systems could bear the harm.Source: Finding 3: AI-washing Leads To Ethical… · p. 9

Discussion

The paper documents AI-washing in German startups through marketing claims, expert accounts, and consumer responses. It recommends disclosure and routine audits, using the codebook to screen public marketing before escalating flagged cases for system-level review.

Discussion

  • German AI startups marketed surveillance, autonomous decisions in regulated domains, and behavioural tracking and manipulation as ordinary AI products, without mentioning their EU AI Act high-risk or prohibited status.Source: Discussion · p. 9
  • Consumers who had encountered exaggerated claims trusted later AI marketing less, while investors may fund products whose AI they cannot easily verify.The conclusion describes consumers in unfamiliar domains taking marketing claims at face value.Source: Discussion · pp. 9–10
  • The authors recommend product-page disclosures about collected data and system limitations, independent certification of marketing claims, and codebook-based screening to identify cases for escalation.The proposed independent certification would inspect technical documentation and performance benchmarks under confidentiality.Source: Discussion · p. 9
  • The study combines an annotation codebook applied to 100 startups with 63 expert interviews to document AI-washing across marketing, expert accounts, and consumer responses.The interviews comprised 53 written and ten oral interviews.Source: Discussion · p. 9
  • The authors recommend disclosure requirements and routine audits, offering the codebook as a low-cost instrument for population-scale review by regulators, investors, and consumer-protection bodies.They frame the audit target as startup marketing claims rather than the systems themselves.Source: Discussion · p. 10

Ethics Statements

The study reports ethics approval, participant protections, and author positionality. Its public-materials audit supports claims about patterns across the sample, not findings that a named firm misled a particular person.

Ethics Statements

  • The University of Hamburg Business School ethics review board approved the interview protocols before recruitment began.Source: Ethics Statements · p. 10
  • The authors note that their startup and marketing experience could have led them to score ambiguous claims more harshly, while anchored scales and interviews independently constrain that concern.Source: Ethics Statements · p. 10
  • The audit used publicly accessible materials from registered German firms and collected no non-public company or employee data.Source: Ethics Statements · p. 10
  • The anonymized results describe prevalence across the sample, not whether a named firm misled a particular investor or customer; the protocol cannot settle that legal question.Source: Ethics Statements · p. 10
  • Oral interview participants consented to recording, transcription, anonymized quotation, and withdrawal before manuscript submission; recordings were deleted after transcription and an accuracy check.Source: Ethics Statements · p. 10
  • Written interview participation was voluntary and uncompensated, recruitment used open social-media posts, and identifying information was not collected.German-language interview material was translated into English for analysis.Source: Ethics Statements · p. 10

Appendix A: Codebook Design

The appendix specifies seven codebook dimensions and the composite judgment used to assess AI-washing in startup marketing. It also describes the score directions and the distribution figure’s scope.

Appendix A: Codebook Design

  • Higher scores indicate stronger performance for D1–D5 and D7, but greater ethical concern for D6; choosing the lower-concern rating between adjacent anchors makes the composite a lower bound on AI-washing severity.Source: Appendix A · p. 13
  • D1–D3 assess claim specificity, whether the product uses genuine or repackaged AI or no AI, and what outsiders can verify about AI functionality.D1 ranges from branding without functional content to disclosure of the model and method.Source: Appendix A · p. 13
  • D4 assesses whether numerical performance claims have a source or methodology, while D5 compares AI’s prominence in marketing with its actual product role.A D4 score of zero means no quantitative claim was made, not a low or high degree of washing.Source: Appendix A · p. 13
  • D6 codes the ethical risk of the application independently of marketing behaviour, while D7 records disclosure of limitations such as accuracy bounds and human-in-the-loop dependencies.Source: Appendix A · p. 13
  • The composite AI-washing score is a holistic judgment integrating the seven dimensions, not an arithmetic mean, with ethical concern raising and strong evidence or disclosed limitations mitigating the score.Source: Appendix A · p. 13
  • Figure 1 reports score distributions for the seven dimensions and the full-sample composite, which is distinguished by a different color.Source: Appendix A · pp. 13–14

Appendix B: Written Interview Details

The appendix reproduces the survey instrument, omitting consent and optional contact-detail sections; all questions were optional.

Appendix B: Written Interview Details

  • The appendix reproduces the full survey instrument, apart from the consent and closing contact-detail sections.Source: Appendix B · p. 14
  • Questions in the survey instrument were optional.Source: Appendix B · p. 14
  • The omitted closing section concerned optional contact details for respondents willing to be interviewed.Source: Appendix B · p. 14

1. What is your age?

The age-response options span from below 18 years to 65 years and older.

  • The youngest response options are below 18 years and 18–24 years.Source: §1 · p. 14
  • The middle response options are 25–34, 35–44, and 45–54 years.Source: §1 · p. 14
  • The oldest response options are 55–64 years and 65 years and older.Source: §1 · p. 14

2. What is your gender?

The survey asks about participant backgrounds, AI familiarity and experience, perceptions of startup AI marketing, and perceived AI capabilities. The interview appendix provides consent language and a semi-structured guide.

2. What is your gender?

  • The questionnaire asks about participants’ gender, educational qualifications, field of study, and current work or study.Source: §2 · p. 14
  • It asks respondents to rate AI familiarity, report AI use, and describe their experience, including tools, context, and frequency.Source: §2 · pp. 14–15
  • Questions probe reactions to AI-powered advertising and experiences of startup AI claims or AI-washing, including claimed-versus-actual reality and consequences.Source: §2 · p. 15
  • Respondents are asked whether AI can reliably and ethically deploy surveillance, make complex autonomous decisions, or take over therapy roles.Source: §2 · p. 15

Appendix C: Oral Interview Details

  • The interview consent script describes a 20–30-minute recorded interview, anonymity, optional questions, and discarding recordings after notes are finalized.Source: Appendix C · p. 17
  • The semi-structured interview guide covers background, AI experience, startup claims, capability judgments, and marketing pressure; question order follows conversation flow.Source: Appendix C · p. 17
Loading 2610.11788v1…