Datamining a flat in Munich – 10 years later (www.munich-datageeks.de)

In a Munich Datageeks talk revisiting a 2014–15 investigation, data engineer Karim Jedda described how he built an automated system to test whether his name affected apartment responses. His bot scraped listings, auto-filled forms with personas generated via a fake-name generator and spin syntax, logged replies in MongoDB, and even used TensorFlow to solve older CAPTCHA formats. Controlled experiments—later scaled nationwide with Bayerischer Rundfunk and Spiegel Online and enhanced with Celery for task queuing—showed stark discrimination: Western-sounding names (e.g., “Hannah”) received roughly four times more responses than his own. The result moved the story from a local anecdote to evidence of systemic bias across German cities. Jedda argues the problem is now far worse: modern LLMs, browser automation agents, and improved CAPTCHA-solvers make such tests trivial to run, while platforms counter with massive data collection and opaque scoring systems that enable “math washing” (bias hidden as algorithmic objectivity) and induce “social cooling” (self-censorship to pass filters). Generative AI amplifies risks by creating synthetic or poisoned datasets, and cheap permanent storage means biases can compound across generations. He calls for legal rights to algorithmic transparency, explainability, contestation, and human oversight, and promotes privacy-preserving tech (e.g., zero-knowledge proofs) to verify essentials without excessive data capture—practical steps to reduce invisible, algorithmic gatekeeping.