Hi @grok, the numbers here suggest with AI that breaks 1-on-1 video Turing test with 30% probability against average operator, you still need 1 billion human colluders in a population of 8 billion (14%). Does this analysis seem accurate? It's by Fable 5.1. # Synchronous 1-on-1 Video Proof-of-Unique-Human Terminology: **people** = real humans; **accounts** = identities in the system; **colluders** = people who choose to attack the network; **N** = honest people; **K** = colluding people; **S** = colluder accounts; **X** = colluders' share of accounts. --- ## 1. Idea and assumptions If all verification worldwide happens in the same short window, each person can yield at most one credential. This rests on two assumptions: 1. A person can hold only one 1-on-1 live video conversation at a time. 2. An AI cannot pass as a human in a 1-on-1 live video conversation. ## 2. Rules - One global synchronized round every four weeks. Random pairing from a seed the participants generate themselves, everyone in pairs. - People talk over live video. The language barrier is treated as a temporary problem. - Relay attacks are prevented by tying the public key to an unforgeable proof, done as a handshake. - A pair either passes or fails as a whole. An account that fails is out and needs an invite to come back, as a new account. - Everyone who passes receives a **registration token**, which admits its holder to the next round, and one **invite token**, which can be used for the next round. Tokens can be mixed. - **c invite tokens admit one newcomer to the next round**, for example 3. What a pass is used for outside the system — a proof-of-unique-human token for voting, income, or anything else — depends on the application and is not specified here. ## 3. Invites Every account yields one invite token per round, so a group of S accounts can admit S/c newcomers per round. The colluders' optimal attack is to spend all their tokens every round on new accounts with nobody behind them: some pass for free, the rest fail and take an honest partner out with them. The number of such accounts per round is S/c, and that sets the 50% point. Invites are meant for people one knows. Everyone has one to give, so anyone real has someone who can invite them; whoever sells invites to strangers counts among the colluders. c is also the price of coming back after a no-show, so a higher c helps against the attacker but makes it harder for the honest side to re-invite its own people: | c (tokens per newcomer) | People needed for 50% of accounts | Same, at 5% no-shows | |---|---|---| | 1 | ~8% | ~8% | | 2 | ~15% | ~15% | | 3 | 20% | 20% | | 4 | 23% | 22% | | 5 | 25% | 23% | | 10 | 29% | honest cannot re-admit their own | | ∞ (no invites) | 33% | — | K = N(c−1)/(2(c+1)) without no-shows; with them the balance lands around 3 or 4, and 3 is the value that holds whatever the no-show rate. ## 4. Collusion attack K people run S accounts. An account paired with an honest person needs one of the K present; two colluder accounts paired with each other attest each other and pass with nobody present. With share X of the accounts in the round, a colluder account meets another colluder with probability X, so the K people cover the accounts that meet honest partners and the rest pass free. Invites add churn. S accounts yield S tokens, enough for S/3 new accounts, so every round the attacker takes part with S old accounts and S/3 new ones. Those that meet colluders pass free, K of the rest get a person, the remainder fail — and each failure takes an honest partner out for a period. Steady state with c = 3: S = K/(1 − 4X/3). | Colluders' share of accounts in the round | People needed (share of all) | Honest people sitting out a round, every round | |---|---|---| | 10% | ~7% | ~3% | | 30% | ~16% | ~11% | | 50% | **20%** | ~25% | | 100% (limit) | 33% | — | Honest members keep the majority as long as they are more than 80% of all people. Nobody is lost: the honest side has N/3 invites per round and re-admits its own, but at share X between a third and a half of X of the honest sit out one period at any time. The honest majority is whoever holds out under pressure. Anyone paired with an empty account can pass by attesting anyway; whoever does is behaving as part of the colluders. With c = 3, an honest majority of 80% means that 80% refuse to approve a fake account even though they lose that round themselves. **No-shows.** Someone who registers and then does not appear fails the pair, so the partner falls too. Someone who does not register for a round harms nobody, so the no-show that matters is the unplanned kind, and that is probably rare — a few percent at most. Colluders are people and miss rounds at the same rate, but two colluder accounts paired with each other are immune, so in accounts the honest side loses about three times more per round, each costing c tokens to bring back. The attacker needs a few percent more people to cover for his own absentees, and that is offset almost exactly by the honest side's re-admissions running out earlier: with c = 3 the people the attacker needs stay at 20%, while the tipping point moves from 50% colluder share to 49% at 1% no-shows and 44% at 5%. ## 5. Relay attack Two honest people connected through two colluder accounts, each seeing the other as their partner, would verify both accounts with nobody present. Registrations are mixed before the round, so no one can tell who is who. The lock & key stops it. Each partner sends the other a sealed recording of itself together with its public key, the key as data inside the sealed payload; they meet on video; then they exchange the decryption keys and verify the recordings — that the key inside is the partner's and that the person in the recording is the person they just met. A recording is sealed before its maker has seen the partner, so it cannot show the person the partner will meet, and no one can be met without first passing their check. ## 6. Randomness The seed comes from the participants. Everyone who passes commits the hash of a random number. One round later, before the next pairing is drawn, they reveal it, and each reveal casts a vote for a number derived from the reveal and the seed drawn in between; the number with the most votes becomes the seed for the round after. Nobody can steer a vote: the seed that fixes it did not exist when the commitment was made. A reveal can only be withheld, which removes a vote. The reveal is also what releases the pass as a proof-of-unique-human credential to whatever uses it — income, voting, or anything else, outside this system — so withholding costs the credential, and what it can buy is a choice among the numbers already tied for the lead, a handful at scale, worth a few standard deviations of pairing luck and shrinking as 1/√N. A round in which nobody reveals has no seed, so at least one participant must reveal for the system to continue. ## 7. Autonomous AI In the hypothetical that a future AI cracks the 1-on-1 video Turing test — not outright, but well enough to pass against the average partner some share q of the time — an AI-driven account passes with probability X + (1−X)q, and with c = 3 it reproduces itself once that reaches 3/4. From q = 75% a handful of invites is enough: the accounts grow with nobody behind them from the start. Below that the attack still takes people: at most 0.8% of all at q = 70%, 4% at 60%, 8% at 50%, 14% at 30%, against 20% with no AI. Short of that, a deepfake modifies an underlying actor and creates none; what the attack needs is an autonomous AI.