Regulate Real-Time Personalized GenAI to Prevent "Digital Heroin" [NeurIPS '25]

In this position paper accepted at NeurIPS 2025, we argue that in the near term, real-time feedback loops + on-the-fly personalization + perverse engagement incentives could turn GenAI into a compulsion engine akin to 'digital heroin,' which may especially affect young people.

Why we think this matters now: - Existing feeds already optimize every swipe and have been shown to be addictive (just look at short-form platforms like TikTok). - AI video is here (think Sora-style quality and speed) and improving fast. - Major platforms are rolling out new AI video-creation experiences that tighten the loop. - The paradigm is shifting from ranking a finite set of human videos to generating effectively infinite, objective-function-optimized clips. Because AI video is now directly optimizable, platforms can, in principle, tune content to maximize engagement (or even micro-signals like pauses, rewinds, gaze, ...).

Here's an interesting post from Karpathy on the topic as well: https://x.com/karpathy/status/1929634696474120576

What we propose: a label for Designated Addictive Systems, audits & transparency, friction-by-design, age-based protections, and a proactive research agenda from the machine learning community.

Oral presentation and poster session details: https://neurips.cc/virtual/2025/poster/121952