This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
Brand safety technology firms marketed their products as utilizing advanced AI and machine learning to protect advertisers, but in practice, these systems relied on rudimentary keyword lists. This resulted in the systematic misclassification of legitimate news articles as 'unsafe' due to the presence of benign words like 'death' or 'weapons' in news contexts. Consequently, mainstream publishers suffered significant ad revenue losses, while ad budgets were inadvertently diverted to lower-quality sites.
Brand safety tech firms falsely claimed use of AI, blocking ads using simple keyword lists.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
Also classified as
The systematic misclassification and subsequent defunding of high-quality journalism devalues the economic viability of professional human reporting.
The brand safety systems lacked the capability and robustness to accurately analyze context, relying instead on blunt keyword matching that resulted in high error rates and misclassifications.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The financial harm to publishers was an unintended consequence of the brand safety AI systems failing to accurately classify content post-deployment.
Brand safety technologies marketed as advanced AI relied on blunt keyword lists, systematically misclassifying legitimate news articles as unsafe. This resulted in significant ad revenue losses for major publishers and diverted funds to lower-quality sites, presenting a minor long-term risk to the information ecosystem and the economic viability of independent journalism.
The brand safety systems operate as tools that perform the specific, well-defined function of classifying webpages as safe or unsafe based on text content inputs.
Minimal or No Risk: The brand safety classification systems are similar to spam filters or simple content classifiers, which pose low direct risk to users and do not fall under the high-risk or prohibited categories of the EU AI Act.
The AI systems are designed to moderate and verify webpage content to ensure it is safe for brand advertisements, filtering out content deemed unsafe.
The report provides clear, empirical data from the Adalytics study quantifying the misclassification rates (e.g., 21% of Economist, 30% of NYT/WSJ, 52% of Vice articles) and the disagreement rate between vendors (40%). The mechanism of failure (blunt keyword blocking marketed as advanced AI) is explicitly detailed.
None. The failure of the technology and its negative economic impact on publishers is clearly established.
Technology
Blunt Keyword Blocking
Systems block ads based on simple keyword lists rather than semantic context.
Lack of Image Analysis
Detection tech fails to analyze images or text inside images on webpages.
Dynamic Content Misalignment
Crawlers miss dynamic webpage changes occurring after initial safety checks.
Data Inputs
Inaccurate Content Labeling
Incomplete and imprecise classifiers falsely label entire domains like CBS.
Conflicting Vendor Classifications
Different vendors disagree forty percent of the time on article safety.
Keyword Manipulation by Fraudsters
Bad actors easily manipulate page text to bypass simplistic keyword filters.
Human Factors
Overreliance on Automated Tools
Advertisers trust automated brand safety tech without verifying its accuracy.
Neglecting Curated Include Lists
Marketers skip manual inclusion lists in favor of automated safety filters.
Process and Methods
Flawed Classification Process
Categorizing entire sites based on a few pages or single words.
Inadequate Crawler Latency
Crawlers analyze pages in milliseconds, missing post-load dynamic content.
Management
Deceptive Marketing of AI
Vendors claim advanced AI and ML but deliver rudimentary keyword blocking.
Prioritizing Scale Over Accuracy
Advertisers prioritize programmatic reach over ad placement quality.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
