Scott Belsky on AI safety layers, consumer AI's untapped potential, and the economics of protecting users

Summary

Scott Belsky argues that the dominant AI safety conversation is misdirected. Rather than fixating on model alignment and reinforcement learning, the more actionable frame is economic: protective AI applications will scale when the financial weight behind them rivals that of the actors causing harm. Belsky points to on-device AI as the immediate opportunity, capable of flagging voice-cloned grandparent scams, filtering phishing attempts, and warning users when algorithmic rabbit holes are pulling them toward fringe content. He notes that screen-time and wellness apps already exist but are structurally disadvantaged because they monetize far less effectively than the doom-scroll products they compete against.

On business models, Belsky draws an analogy to home security. Consumers routinely pay $60 to $120 per month for Ring and similar monitoring services to protect physical property, yet no comparable paid layer exists for digital safety. He sees genuine consumer willingness to pay for an AI-powered equivalent, particularly as long-form social engineering scams become more sophisticated. The logical delivery mechanism is the operating system layer, where iOS and Android hold the persistent, cross-app visibility required to detect patterns over time. Apple's Content Credentials initiative, which Belsky worked on at Adobe, offers one cryptographic approach to flagging AI-generated audio and media at the point of creation.

Consumer AI as the Underdiscussed Frontier

Belsky left a recent major tech conference struck by what was absent from the agenda: new consumer AI applications and AI-era social networks. Every major platform shift, from desktop to mobile, produced a new generation of mainstream consumer products. That wave has not yet materialized for AI, and Belsky believes it represents a significant underexplored investment category.

Early signals he is watching include Tolen, a conversational pet-alien app that has raised from top-tier venture firms. He is also exploring digital-twin simulations where trained AI representations of individuals interact autonomously, and what he calls "wingman as a service" products for dating platforms. A concept he calls Peanut Gallery imagines a social platform where humans post content but only AI personas comment, turning engagement into a voyeuristic exercise in watching curated AI archetypes debate your output.

The Meta Talent War Through an Acqui-Hire Lens

The Meta-OpenAI talent competition is best understood not as a traditional acquisition of network effects but as a series of unauthorized acqui-hires. When Meta bought Instagram and WhatsApp, it was purchasing entrenched user networks. The current talent deals, including the discussion around Alex Wang and Scale AI, are bets on individuals whose value will persist across whatever AI use case emerges next. Scale AI itself has already pivoted from self-driving car data to RLHF for language models, and its durable asset is Wang and the team rather than any specific product surface.

Belsky sees Meta pursuing two parallel goals: making Llama best-in-class so AI capability is effectively free across its existing products, and building net-new consumer experiences centered on companionship and social entertainment. The economics favor both aggressively. A 0.1% improvement in ad efficiency can generate $100 million in incremental revenue at Meta's scale, making even large talent expenditures straightforwardly rational.

LLMs and the Evolution of Human Communication

Belsky flags an underappreciated dynamic: LLMs are now producing more written output than any human in history, and that volume is already feeding back into training data, creating a compounding influence on how people write and communicate. He expects models to personalize more aggressively toward individual communication styles, but draws a parallel to music, where generic production eventually triggers a counter-reaction from artists who do something structurally new.

The limits of AI creative output remain visible in a telling gap: no fully AI-generated social account has built a large, genuinely engaged following through content quality alone. The Velvet Sundown, a reportedly AI-generated music act, has reached roughly one million monthly Spotify listeners, but no equivalent has emerged in text-based social content. Belsky attributes this to taste requiring consistent editorial judgment across a body of work, not just individual outputs, and to the contextual awareness demanded by real-time cultural conversation. He frames the anticipated breakthrough as a "Move 37 moment," referencing AlphaGo's unconventional 2016 play against Lee Sedol, a move initially read as error that proved to be novel genius. That equivalent moment for LLMs, he argues, has not yet arrived.