a16zHow Bots, Deepfakes and AI Agents Are Forcing a New Internet Identity Layer | Alex Blania on a16z
CHAPTERS
- 0:00 – 0:53
Why proving you’re human is about to become urgent (agents, bots, and mislabeling)
The hosts and Alex Blania frame “proof of human” as a rapidly escalating need as AI agents become capable of operating accounts at scale. They distinguish between humans, agents acting for humans, and fully autonomous agents—warning that confusion will soon lead to widespread accusations of people being bots.
- •Proof of human is becoming critical as AI capability and scale accelerates
- •Three interaction categories: human, human-controlled agent, and autonomous agent
- •Bots will soon be far more convincing and numerous than today
- •Agents could create accounts and vouch for other agents, undermining social trust
- 0:53 – 4:01
What “proof of human” really means: uniqueness + ongoing control
Alex defines the core requirements for proof of human: each person should have one (or limited) account and maintain control over it over time. He emphasizes that the hardest property isn’t authentication but uniqueness at internet scale, ideally with strong privacy preservation.
- •Goal: uniqueness per individual (one person → one account / limited accounts)
- •Two phases: initial verification vs ongoing authentication
- •System should be anonymous or privacy-preserving by default
- •Current bot mitigation on major platforms is a losing catch-up game
- 4:01 – 6:54
Three early approaches—and why they failed: web-of-trust, government ID, biometrics
Blania recounts three popular approaches considered before ChatGPT and explains why World dismissed the first two quickly. Purely digital reputation graphs can be simulated by AIs, and government IDs are problematic for privacy, free speech, and global interoperability—leaving biometrics as the remaining option despite discomfort.
- •Web-of-trust/reputation graphs are vulnerable once AI can mimic digital behavior
- •Government ID systems risk central control, weak anonymity, and don’t scale globally
- •Global products can’t rely on one nation’s ID infrastructure
- •Biometrics trigger privacy concerns but address the uniqueness constraint
- 6:54 – 8:25
Why Face ID isn’t enough: the jump from 1:1 to 1:N uniqueness
Alex explains the technical reason common biometrics break down at scale: Face ID-style checks are one-to-one (you vs your stored template), while proof of human requires one-to-many comparisons against the entire network. That shift demands far more entropy, making faces and fingerprints insufficient beyond tens of millions of users.
- •Authentication (1:1) differs fundamentally from uniqueness checking (1:N)
- •Information/entropy requirements grow with network size
- •Faces/fingerprints hit scaling limits at large user counts
- •Iris patterns provide higher entropy suitable for global uniqueness
- 8:25 – 9:01
Replay attacks and the verification/authentication split
The conversation moves to security threats like replay attacks and deepfakes. Alex separates the “verification” moment (akin to getting a passport) from ongoing “authentication” (showing it repeatedly), noting that consumer-device reauth is hard when phones can be compromised or spoofed.
- •Replay attacks are a known biometric vulnerability class
- •Verification (enrollment) vs authentication (continuous proof) are distinct problems
- •Deepfakes can spoof camera input, especially on less-secure devices
- •Some users may need periodic re-verification on dedicated hardware
- 9:01 – 11:35
The Orb: hardware defenses, and why iris scanning may normalize via AR/VR
Alex describes the Orb as purpose-built hardware with multiple sensors to detect spoofing (e.g., displays). Ben addresses early misconceptions about ‘having your eyeball,’ while Alex argues iris scanning will become culturally normal as AR/VR devices adopt similar modalities.
- •Orb uses multi-sensor checks to prevent display/deepfake spoofing during verification
- •System sends a signed face image to the phone for later reauthentication
- •Privacy criticism (‘they have my eyeball’) is framed as a misunderstanding of design
- •Iris scanning may become common due to AR/VR headsets (e.g., Vision Pro)
- 11:35 – 15:21
Privacy architecture: multi-party computation + zero-knowledge separation
Alex outlines how World aims to achieve uniqueness without a central biometric database. Iris codes are computed on-device, split across multiple parties via multi-party computation, then later used through zero-knowledge proofs so platforms can verify uniqueness without learning identity.
- •Uniqueness requires some comparison beyond the device; privacy must be engineered
- •Iris code computed on Orb, then split and distributed—no single party holds the full data
- •Multi-party computation enables uniqueness checks without centralizing biometrics
- •Zero-knowledge proofs allow users to prove uniqueness to platforms without revealing identity
- 15:21 – 21:51
Where bots become unbearable: dating, video calls, gaming, and the creator economy
Ben asks where proof-of-human matters beyond social media replies and propaganda. Alex lists human-centered contexts that break when bots and deepfakes proliferate—dating authenticity, high-stakes video conferencing fraud, competitive gaming integrity, and advertiser/creator trust.
- •Dating apps need assurance the other side is a real, matching person (anti-catfish)
- •Deepfakes will enable real-time impersonation in high-value video calls
- •Gaming will demand proof opponents are human, especially in wagered competitions
- •Creator economy and ad markets need to distinguish human creators/viewers from AI farms
- 21:51 – 23:30
The near-future escalation: “less than 1%” of what’s coming + AI persuasion at scale
Alex argues today’s bot problem is only a preview because intelligence is getting cheaper and agents more capable. He cites research showing AIs can outperform humans in persuasion by tailoring messages to individuals’ profiles—making PSYOPs and manipulation far more effective.
- •Current bot impact is a tiny fraction of what will exist in 1–2 years
- •Agent capabilities scale rapidly; cost of intelligence is falling fast
- •AI can personalize persuasion using user profile and behavioral cues
- •Manipulation becomes scarier when you can’t tell if the counterparty is human
- 23:30 – 25:01
Business/product status: three-sided deployment challenge and current traction
Alex gives a project update and frames success as a three-part coordination problem: platform integrations, physical device distribution, and compelling user utility. He shares current adoption numbers and explains the company’s pivot toward solving distribution at scale.
- •Three-sided problem: platforms adopting it, Orbs distributed, and user demand/utility
- •Metric for distribution: average minutes to reach a device
- •Scale estimate: ~50,000 devices to reach <15 minutes across the US
- •Reported traction: 18M verified users; 40M total app users
- 25:01 – 29:21
US go-to-market pivot: regulatory clarity, normalization, and “Orb everywhere”
The conversation turns to why the US now becomes the primary focus after earlier hesitation due to crypto-related regulatory risk. Alex describes a push to normalize verification behavior (e.g., Orbs in everyday retail locations) and notes shifting sentiment from ridicule to execution urgency post-ChatGPT.
- •Prior US underinvestment attributed to regulatory environment around crypto
- •Next-year focus: concentrate company effort on US distribution and adoption
- •Vision: Orbs in ubiquitous locations (e.g., Starbucks) to normalize behavior
- •Market shifted from skepticism to urgent execution after recent AI/bot breakthroughs
- 29:21 – 33:11
From “proof of personhood” to “proof of human,” and why the market flipped
Alex explains the terminology change: ‘personhood’ could apply to AIs, while ‘human’ is the target boundary for now. He also describes two inflection points in interest—ChatGPT making AI real, and later bot surges (e.g., “Claude bots”) turning the issue into an immediate operational priority.
- •Original framing: proof of personhood; changed as AI ‘personhood’ became plausible
- •Post-ChatGPT: more inbound interest, but still treated as future problem by many
- •Later bot waves made it feel urgent and execution-driven rather than thesis-driven
- •Expectation: platforms will try weaker phone-based biometrics first, then hit limits
- 33:11 – 36:43
Proof-of-human as civic infrastructure: payments, benefits fraud, and democratic legitimacy
Ben broadens the discussion to governance: distributing money to real people, preventing fraud, and maintaining election integrity in an AI-enabled impersonation world. They argue society may need cryptographically strong identity infrastructure to preserve trust at national scale.
- •Stimulus and benefits programs show how costly identity fraud can be
- •AI can massively scale fraudulent claims and identity misuse
- •Democracy/elections require confidence voters are real, living people
- •Need for cryptographically strong infrastructure for citizenship and eligibility proofs
- 36:43 – 40:09
Execution plan: platform-driven demand, distribution partnerships, and “Orb on demand”
Alex details how go-to-market changes once large platforms can funnel users into verification. He describes distribution strategies—from big retail partnerships to one-off placements—and introduces an “Orb on demand” concept for dense cities where a mobile Orb can come to users.
- •Large platform integrations could drive massive user demand; initial rollouts may be geo-limited
- •Key engineering priority: make Orb operate reliably at scale with minimal supervision
- •Distribution options: major partners (Walmart/Starbucks), local venues, even DMVs
- •“Orb on demand” pilot: deliver an Orb to users (e.g., Bay Area/NYC) to reduce CapEx burden
- 40:09 – 42:11
Verification gradations: face checks, NFC government IDs, and rate-limiting tradeoffs
Ben asks about tiers of verification, prompting Alex to describe multiple verification levels inside World. Face-based checks and NFC-based government IDs can provide weaker guarantees useful for rate limiting and onboarding, but Alex warns deepfakes will eventually break phone-based face checks—making Orb-level verification the durable path.
- •System supports gradations of assurance (strong Orb vs weaker phone-based checks)
- •“Face check” uses camera + multi-party computation to preserve anonymity
- •Weaker methods can rate-limit account creation (e.g., 10–20 vs 100 accounts)
- •NFC government ID verification exists but faces stigma; deepfakes will erode face-only reliability
