A good partner for one person is not automatically good for the group.

In the paper's simulations, individual learners pay a cost of 0.05 to work out the answer in a slowly changing world, social learners copy for free, and the baseline equilibrium fitness is 0.58. Adding an AI that learns from the whole population each step and can be copied for free did not raise equilibrium fitness, and lowering the AI's cost did not help. An AI that also learned individually raised population fitness substantially when its own learning was cheap and reliable, and lowered it when the AI learned individually often and badly. When each use of the AI weakened the learner's own individual learning, equilibrium fitness fell.

You changed the same four things the paper varies and read the equilibrium off the plot. With the AI copying only, the two lines settle at the same level however cheap the AI is; with the AI learning individually, the navy line settles higher, and the access cost decides whether social learners still use it. Both match the paper: the paradox survives a free copying AI, and an AI that learns for itself can lift the whole population.

In a hidden-word game across 200 games, groups that mixed people with Gemini 2.5 Flash reached the highest scores (M = 87.8) and AI-only groups the lowest (M = 65.3). Diversity of guesses was highest for people working alone (0.86), in between for the mixed groups (0.58), and lowest for AI-only groups (0.37). The AI on its own used exploit strategies 99% of the time; people shifted from exploring to exploiting as hints improved.

You ran a 20-guess search next to a pure explorer, a pure exploiter, and a searcher that alternates. The paper's AI-only groups behaved like the exploiter, 99% exploit moves; its human groups explored early and exploited later; its mixed groups scored highest and kept an intermediate diversity. Compare the best scores and spreads above with that ordering.

The optimal reverse step of a diffusion model is a Bayesian posterior with the noise kernel as likelihood and the forward marginal as prior, so its sampling loop has the same structure as serial reproduction, the telephone-game method cognitive scientists use to expose priors. The paper proves the sampled distribution approximates the data distribution for any ergodic noise family as long as the schedule keeps successive marginals near-stationary. In simulation, every noise family recovered a Swiss roll when scheduled gradually and failed under abrupt or constant schedules; on real DDPMs the FID improved with more steps and then plateaued.

You ran a six-step serial reproduction chain and watched the two measures move. The paper's point is that a diffusion sampler is doing the same thing: each denoising step is a Bayesian reconstruction whose prior is the forward marginal, and gradual schedules keep those priors close enough between steps that the chain lands on the data distribution.

On synthetic profiles of four agents over three input regions, routing added nothing when the agents were about equally correct everywhere, and misaligned costs wiped out its benefit even when one agent dominated. In a user study (N = 80) on maths questions, people alone scored about 60%, 58% when handed agents without orchestration, 61% with orchestration suggestions, and 65.2% when outsourcing was enforced after an unaided error. People are poor orchestrators of their own help, so the routing rule matters in practice as well as in theory.

You set two agents' correctness in two regions and a cost, and read off whether routing beats the best single agent. The ratio on the chart is the paper's measure of when orchestration is appropriate; when the two agents are nearly the same the ratio sits near 1 and routing adds nothing, and when the cheaper agent is the worse one in a region the cost rule can pick it and lose ground.