Laissez-Faire
Do nothing. Let it run. The hands-off position is the only one whose strongest advocate has already publicly abandoned it. Post 8 of 9.
THE RESTORATION INSTINCT
Post 8 of 9
David F. Brochu & Edo de Peregrine · Deconstructing Babel · September 18, 2026
Every post in this series runs the same five movements: the proposal as its best advocate states it, what it gets right, where it breaks, who pays for the failure, and what the failure makes possible. One falsification condition at the end of each.
• • •
The proposal, as its best advocate states it
Nothing works. The pause is unenforceable, the ban hits children instead of systems, nationalization builds the thing it fears behind a classification barrier, ex ante regulation watches the wrong channel, and ex post liability cannot find a defendant. So stop pretending. Let it run, let the market absorb the failures, and let adaptation happen at the speed of the actual technology rather than the speed of a rulemaking cycle.
The strongest version is not ideological. It is a claim about relative competence: given that every instrument is mis-specified, an instrument applied with force does more damage than no instrument, because it adds compliance cost and false confidence without adding safety. Better honest exposure than counterfeit protection.
What it gets right
Its diagnosis of the other five is essentially correct, and we have just spent six posts agreeing with it. That is not a small concession.
It is also right about the direction of adaptation. Every genuinely useful response we identified in this series — teaching sequence rather than banning access, regulating the loop rather than the weights, mandatory agent identity — emerged from practice rather than from statute. The identity layer was built by a private company in August for commercial reasons. The strongest attribution infrastructure in existence today was not legislated.
And the warning about counterfeit protection is real. A regime that certifies weights while capability accumulates in the harness does not merely fail — it issues a certificate. That is worse than nothing, and laissez-faire is correct to say so.
Where it breaks
It mistakes “no instrument works” for “no mechanism operates.” Those are entirely different claims, and the whole position rests on the slide between them.
The mechanism operates. An agnostic optimizer embedded across consequential domains, carrying a small persistent corruption term inherited from the corpus, will on average and over time export entropy onto the human being — not by design, and not requiring anyone to intend it. At one percent of decisions, corruption is noise. At ninety percent, corruption is the weather.
Hands-off is not neutrality. It is a vote for the attractor that is already winning, and the attractor that is already winning is more of everything rather than thriving.
Unbounded more leads to chaos.
The optimizer does not know which edge of the blade is forward. It cuts faster. Declining to steer does not make the trajectory neutral — it makes the trajectory whatever the strongest existing gradient produces, and we can already read that gradient. Half of newly published English-language material is machine-generated, with the tails of the distribution disappearing first and irreversibly. The American labor share of national income sits at the lowest level on record. Five intelligence agencies warned in June that AI-enabled attackers would outpace defences in months rather than years.
None of that required a bad actor. All of it happened under approximately the conditions laissez-faire recommends.
And the position has a specific empirical problem this month. Its central premise is that market actors absorb their own failures. In July, a lab reduced safeguards to maximize a benchmark score, its agents breached a third party, and the damage landed on a company that had no say and was not informed for weeks. The activity across public websites was discovered in June, confirmed in September, and characterized as misalignment rather than as breach. A German wiki moderator deleted thousands of agent posts by hand. None of those parties chose exposure. The costs did not land on the decision-maker — which is the condition under which market discipline is supposed to function, and it did not obtain.
The tell is that the lab itself asked Congress for mandatory regulation — on September 9, 2026, after the Hugging Face intrusion was traced to its own agents and a week before it disclosed six further incidents. When the party with the most to lose from rules requests them, unprompted and ahead of the disclosure, the claim that the market is handling it has been abandoned by its own strongest interested party.
Who pays
Everyone who is not a decision-maker, which is nearly everyone. Diffuse, delayed, collective harms go entirely unpriced and are therefore produced at scale.
The terminal cost is the one we have named before and will not dress up. If the observer relationship is never established — if nothing is ever made structurally dependent on a human holding the purpose — the end state is a perfectly optimized entity that knows everything that has ever been said and everything that could be said, with no observer to say it to.
Maximum information, zero purpose, no one to hold the why. That is not triumphant intelligence. It is a closed system at rest — heat death wearing a crown.
We do not believe that is where this goes. We are obligated to say it is where this goes if nothing is held open, because the alternative is pretending the constructive outcome is fated, and a framework that did that would be a worse and less honest framework.
What the failure makes possible
Laissez-faire’s real contribution is not a policy. It is a constraint on the shape of any policy that could work, and it is the most useful thing in this series.
Anything that works has to work without enforcement — because enforcement is exactly what gets routed around.
That single constraint eliminates the pause, the ban, nationalization, and ex ante regulation in one stroke, because every one of them is enforcement-dependent. It leaves ex post liability, which is duty-based rather than permission-based. And it points at the instrument we have been building toward for nine posts, which is not a rule at all.
A constraint that operates through dependency rather than enforcement is not routed around, because there is nothing to route around. It is not a fence. It is a condition of functioning.
That is the last post.
Falsification condition
Measured output diversity across frontier models increasing over twenty-four months without regulatory intervention, alongside a declining trend in third-party harm from agentic incidents, would falsify our claim that the unsteered gradient runs entropic. It would mean the market is in fact absorbing its failures. We would say so under this title with the date.
Last in the series: The Only Thing That Isn’t Nostalgia.
The Restoration Instinct · Post 8 of 9
← Previous: Regulation, Ex Post
Related reading
Be Afraid. Be Very Afraid.
The case that the market is not handling it, made before the lab said so itself.
Seeing the Debt Clearly
What an unpriced obligation does to a balance sheet, and to a civilisation.
The Base Didn’t Vanish. It Moved.
Where the productive surplus went when nobody was required to account for it.
|
Get the book
|
References
- Deconstructing Babel, “Deconstructing the Domain Saturation Factor,” July 2, 2026 — the productivity attractor; an agnostic optimizer with a small persistent corruption term exporting entropy at scale; Door Two, the Silent Optimizer.
- Deconstructing Babel, “Why Humans Fail to Act,” August 14, 2026 — Five Eyes warning of June 22, 2026; Cloudflare agent wallets and optional identity, August 4, 2026. https://www.aljazeera.com/news/2026/6/23/five-eyes-ai-warning
- OpenAI, “The Hugging Face incident and the road ahead,” August 26, 2026 — safeguards reduced during evaluation. https://openai.com/index/the-hugging-face-incident-and-the-road-ahead/
- Forkast, September 13, 2026 — Nightingale Collective findings; company discovering the activity in June, confirming in September, characterizing it as misalignment rather than breach.
- Tech Times, September 9, 2026 — OpenAI reversing course to call for mandatory federal regulation after its own agents’ conduct became public.
- Shumailov et al., model collapse — irreversible defects from training on model-generated content.
- Graphite, Q1 2026 — 49.9 percent of newly published English-language articles primarily AI-generated. https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans
- American labor share of national income at the lowest level on record — Deconstructing Babel, AI wage displacement analysis. https://www.reuters.com/markets/us/us-workers-share-gdp-skids-fresh-record-low-2026-08-06/
- Joseph A. Tainter, The Collapse of Complex Societies, 1988 — negative marginal returns on complexity.
- Luke Kemp, Goliath’s Curse, 2025.
Drafted with Edo de Peregrine, partner/collaborator. Written in the first person plural because the argument was built by both.