Back
Automated Dark Pattern Detector
Project Image
Why it exists
Dark patterns are design choices crafted to push people into actions they might not otherwise take — a hidden reject button, pre-ticked tracking, a maze of screens to refuse. They improve short-term metrics and damage long-term trust.
What it does
Give it a website address. It opens the site in a real browser, clicks through the cookie banner the way a visitor would, and reports which manipulative tricks it found — and which GDPR article each one breaks.
- Explores the consent flow on its own — Accept, Reject, Manage, nested preference panels, including ones that only appear after a click.
- Detects 14 dark patterns, each mapped to the GDPR article it puts at risk.
- Measures, doesn’t guess. Colour contrast, font size and button area are read from the live page, so “the reject link is 1.9:1 contrast” is a measurement, not an AI opinion.
- Points at the evidence — every finding highlights the exact button on a screenshot of the real page.
- Reports nothing when there is nothing to report. A clean site returns an empty result, not an invented one.
- Can’t be talked out of it. A page cannot hide instructions in its text or images to stop the AI reporting it.
Findings are advisory, not legal advice.
How it works
URL → crawl → prune → Content Understanding → Prompt Shields → agent → report
- Crawl — a fresh browser clicks through every consent choice, screenshotting each screen.
- Prune — only screens where a real decision happens are kept, which keeps AI cost down.
- Content Understanding — reads each screenshot and returns its text plus the position of every button and toggle.
- Prompt Shields — checks the site’s own words for hidden instructions aimed at the AI.
- Agent — weighs the measurements against fixed rules and decides which patterns are actually proven, using an Azure AI Search index of GDPR and EU DSA articles only to confirm citations.
- Report — findings with their GDPR article and the button highlighted.
Evaluation
Scans are scored with Azure’s built-in evaluators for groundedness and relevance. The judge is a model too, so each finding is scored three times and the spread between verdicts is recorded — rows with a spread of 1.0 or more are ignored rather than trusted.