Sheep's clothing, wolfish intent: Automated detection and evaluation of problematic 'allowed' advertisements

The digital advertising ecosystem sustains the free web and drives global innovation, but often at the cost of user privacy through intrusive tracking and non-compliant ads, especially harmful to under-age users. This has led to widespread adoption of privacy tools like adblockers and anti-trackers, which, while disrupting ad revenues, expose users to alternate forms of tracking and fingerprinting. To address this, many adblockers now allow 'non-intrusive' ads by default. In this study, we evaluate Adblock Plus's Acceptable Ads feature and find a 13.6% increase in problematic ads compared to no adblocker use—challenging claims of improved user experience. We also find that ad exchanges on allowlists are more likely to serve problematic content, underscoring the hidden cost privacy-aware users pay when relying on such technologies. While prior work in the domain has been limited by their practical viability, we further propose a methodology to automate the detection of problematic ads using LLMs with zero-shot prompting, achieving substantial agreement with human annotators (IAA score: 0.79). This establishes the efficacy of LLMs in problematic content detection under well-defined environments. As In-browser LLMs emerge, adversaries may exploit problematic ad content to fingerprint privacy-conscious ABP users. At the same time, these advances present new opportunities for adblockers to develop robust defenses, detect malicious exchanges, and uphold both user privacy and the sustainability of the ad-supported web.