Balaji Srinivasan: Prove Correct, Not Just Go Direct

What It Covers

Balaji Srinivasan joins a16z's Eric Torenberg to argue that the era of "going direct" on social media is insufficient in 2026. As AI-generated content floods every communication channel — collapsing trust in resumes, journalism, and sales — the only durable solution is cryptographically verifiable information: on-chain data, signed records, and math-based truth that requires no institutional trust.

Key Questions Answered

• •Prove Correct vs. Go Direct: Publishing your own content on social media was the winning media strategy from 2015–2022, but AI-generated synthetic content has made distribution alone insufficient. The next five-year strategy requires cryptographic proof attached to claims — timestamped, signed, on-chain records that anyone can independently verify without trusting the publisher. The shift is from "I said it" to "here is the unfakeable evidence that it happened," a fundamentally different standard for credibility.
• •The Verification Gap Destroys High-Trust Channels: AI makes writing a resume, cold email, or sales pitch nearly costless, but it raises verification costs exponentially. Recruiting, sales, and marketing channels — built for moderate adversarial load — now receive volumes of synthetic content that defeat probabilistic spam filters entirely. The practical result is that only warm introductions survive as trusted signals. Organizations should rebuild outbound and inbound pipelines around deterministic trust mechanisms rather than volume-based filtering.
• •Optimal AI Usage Is Not 100%: Treating AI as a complete replacement for human output produces detectable "slop" — content that defaults to unchanged model settings and reads as generic filler. The functional framework is a Laffer-curve model: 0% AI is inefficient, but 100% AI degrades signal quality to zero. The actionable standard is disclosed, polished AI use — where output has been prompted aggressively enough that it no longer reads as templated — combined with human verification at every output stage.
• •Crypto Is Deterministic Where AI Is Probabilistic: AI cannot compute the preimage of a cryptographic hash function, cannot forecast chaotic or turbulent systems, and cannot forge a blockchain timestamp. These are provable mathematical constraints, not temporary limitations. This makes cryptography the complementary layer to AI: AI handles probabilistic pattern recognition while cryptographic systems handle unfakeable attestation. Builders should treat on-chain signatures, timestamped records, and verifiable credentials as the hardened factual substrate beneath any AI-generated content layer.
• •On-Chain Media as the Ledger of Record: Financial data already lives on-chain with full auditability — the FTX hack timeline, for example, can be reconstructed entirely from Etherscan records without relying on any news outlet. Extending this model to social data via protocols like Farcaster creates a verifiable, open, non-paywalled record of events. The practical build path is: raw on-chain data feeds