AI engineer working on industrial ML, and a governance researcher focused on AI regulation in emerging economies. Coordinated the workshop that built a 501-page national AI framework submitted to a government ministry. Writing here about risk classification, regulatory transplantation, and the gap between policy language and institutional capacity.
Egemen Tunçarslan
Karma: 2
The detail that stuck with me is that camera controls worked where ‘fair-play committees’ didn’t. That maps onto something I keep seeing from a governance angle: the deterrent isn’t the rule, it’s whether a functioning mechanism sits behind it. The Metta case basically established that the rule existed but the enforcement didn’t, and once that was common knowledge, cheating became the default not because norms collapsed, but because the cost of enforcement (social backlash, appeals, organizer futility) exceeded the cost of looking away.
What I find genuinely unsettling in your framing is the combination: gradual disempowerment doesn’t need superhuman AI AND it doesn’t need broken norms. It just needs the enforcement mechanism to be slightly more expensive than the act it’s meant to deter. The Go players had the norm (‘AI cheating is dishonorable’) intact the whole time arguably the vilification made enforcement harder, not easier. So the disempowerment ran straight through a fully intact value system.
Two questions this raises for me.
First: is the ‘illusion of control’ you describe load-bearing, or incidental? I.e. would removing it (somehow making users vividly aware they ‘haven’t got it’) actually change behavior, or would they disempower themselves anyway because the obstacle-avoidance gradient is doing the real work?
Second, on the cultural-evolution rebuttal in Appendix A the pre move 60 vs post move 60 split is a sharp test and I’d want to know whether it replicates in other domains where people copy AI policy on the ‘opening’ and then go off-script. If it does, that’s a fairly general diagnostic for ’mimicry mistaken for skill.
Reading this in 2026, the propaganda-territories section holds up unnervingly well on its core mechanic the shift from ‘promote ideas’ to ‘use audience reaction as a reward signal’ but the map of territories is drawn entirely from the perspective of places that build the models. Western Left, Western Right, CCP, Putin. There’s a fifth category the vignette doesn’t have a slot for: countries that import the entire stack and inherit its embedded censorship-and-propaganda regime without having built any of it.
I work on AI governance from one of those countries. The Christian Coalition example is actually the closest the piece gets to my reality ‘easier to use Google’s stack and accept the associated censorship than build your own’ but you frame it as a fringe opt-out. For most of the world it’s the default, not the opt-out. We don’t get a territory of our own; we get whichever territory’s stack is cheapest to deploy, with its content moderation tuned to a context that isn’t ours. A toxicity classifier trained on US English norms decides what’s sayable in a language and political context it was never calibrated for.
The part I’d flag as having aged less well is the implicit assumption that the propaganda regimes are authored by actors who want them. In import contexts, a lot of the ‘nudging’ is unowned nobody chose it for us, it just came bundled. That seems harder to regulate than deliberate propaganda, because there’s no author to point at. Curious whether your later vignettes (2027+) ever extend the territory map past the model builders or whether the import dependent world stays off-screen.