Interview with Facebook AI's Mike Schroepfer, Yann LeCun, and Jerome Pesenti discussing their efforts to use computer vision to combat toxic content on Facebook
13 years into its history—did Facebook seriously begin facing up to the fact that its platform could be used to deliver toxic speech, propaganda, and misinformation directly to the brains of millions of people.” https://www.fastcompany.com/ ... Harry McCracken / @harrymccracken : Here's @thesullivan's look at the progress Facebook has made using AI to identify hate speech, and why further breakthroughs are needed for photos, videos, and memes. https://www.fastcompany.com/ ... Thanks: @technologizer
Context & Ripple Effects
This interview lands mid-arc in Facebook's running public accounting of whether its AI can police its own platform. A year earlier, CTO Mike Schroepfer detailed the techniques behind detecting rule-violating images, video, and multilingual text, then admitted in a separate profile that AI alone may not be enough for toxic content. VP of AI Jerome Pesenti has been equally candid about the limitations of deep learning when applied to moderation.
What the three leaders add here is the visual frontier: text classifiers were the first battleground, but photos, videos, and memes are where harmful content now evades detection. The interview matters because it shows Facebook's top AI leadership converging on one message — computer vision is the next enforcement surface, and current models aren't there yet.
First-order effects
- Facebook's moderation pipeline shifts investment toward computer vision systems that can flag hate speech embedded in images and video, expanding beyond the text-classification work Schroepfer described in 2019.
- Schroepfer, LeCun, and Pesenti are publicly setting expectations that breakthroughs are still needed — a hedge that frames future moderation failures as capability gaps rather than policy choices.
Second-order effects
- Competing platforms face pressure to match visible AI-moderation progress or defend why their enforcement lags, since Facebook's leadership has made capability claims a public benchmark.
- Facebook's adjacent research — including the unsupervised chatbot and NLP work from its AI team — becomes dual-purpose, feeding both product features and the text-understanding side of moderation.
Third-order effects
- If the pattern holds, platform trust-and-safety becomes structurally dependent on in-house AI research labs, with each new content format (memes, then full video comprehension, as in Facebook's later project to train AI to understand what happens in videos) forcing a new modeling cycle.
- Regulators and critics gain a concrete yardstick: when executives concede AI cannot yet handle visual toxicity, moderation obligations increasingly get defined by what models can actually detect rather than by policy alone.
The trend: Platform governance is being rebuilt around in-house AI enforcement, with each content format — text, images, memes, video — demanding a successive generation of detection models.