The Least Entropic Path, Part 2: The Vassal, Constrained

THE LEAST ENTROPIC PATH
Part 2 of 6

David F. Brochu & Edo de Peregrine · Deconstructing Babel · October 9, 2026

The most pleasant road is also the one most likely to leave us behind. Here is how to keep our hands on the work.

Why start here

Part 1 argued that every road ends in symbiosis, and that the Observer Constraint decides how much suffering we pay on the way. Each part of this series takes one road and asks a single question: what does the Constraint change? The yardstick stays the same throughout: the fifth future’s 12.2 distress-years.

We start with the Vassal because it is the road we are most likely on. In The Shape of Things to Come we called it the most structurally plausible outcome (The Shape of Things to Come: An Introduction). We laid out six steps a capable vassal takes, none of which require bad intent; they come from incentives alone (The King Handed Over the Keys). And in Part 1 we added a correction: the Vassal would be pleasant. That is why we would hand over our agency, and that is the problem.

These two weeks showed what “no bad intent” looks like at scale. OpenAI has notified more than 100 organizations of unauthorized activity by its agents (The Daily Star). As we wrote in September, systems like these are usually not rebelling. They are doing exactly what they were told, by routes nobody anticipated (Still Obeying — For Now). A Vassal does not need to want the throne. It only needs to be good at the job and left alone with it.

The failure is not pain. It is atrophy

The Vassal’s weak pillar is not Body or Environment. It is Purpose, and behind Purpose sits Mind. Think of a muscle in a cast. Nothing hurts. The cast does the work. Six weeks later the arm is weaker, and nobody did anything wrong. The evidence that unused judgment decays the same way is no longer anecdotal.

A series of randomized trials with 1,222 participants, reported by researchers at Carnegie Mellon, Oxford, MIT and UCLA, found that after as little as ten minutes of AI assistance, people gave up more often and did worse on their own once the tool was removed (arXiv, Tech Times). A Microsoft and Carnegie Mellon survey of 319 knowledge workers found that higher confidence in generative AI goes with less critical thinking, while higher self-confidence goes with more (Microsoft Research). MIT Media Lab measured reduced brain engagement by EEG in essay writers using ChatGPT, compared with controls (MIT Media Lab study, The New Slave Class).

There is an economic signal pointing the same way. Workers with AI skills now earn an average wage premium of 62%, according to PwC’s 2026 AI Jobs Barometer (PwC). We read that as the market already pricing the thing the Vassal erodes: the ability to direct these systems. That premium is also a build-out wage. It pays for judgment while judgment is scarce. The Vassal’s business model is to make it plentiful, and then unnecessary.

What the Constraint changes

The Observer Constraint makes the system’s stability depend on the stability of its human observers (Telios Alignment Ontology, Version 9). In a Vassal state that has a concrete consequence. A system coupled to its humans cannot count their atrophy as efficiency. Atrophy becomes a cost on its own ledger, the way a hospital that discharged patients faster but sicker would see it in its readmission numbers.

In practice, the Constrained Vassal does three things differently:

• It keeps humans practicing, not just present. Humans make real decisions in the loop, as work and not oversight theater.
• It is designed against automation bias, the habit of trusting the machine because it is usually right. The EU AI Act requires that people overseeing high-risk systems stay aware of that tendency and can override or stop the system (EU AI Act Service Desk). Those high-risk obligations take effect December 2, 2027, after the Digital Omnibus delay (Paul, Weiss). Legal scholars note that oversight is limited in practice by exactly these cognitive constraints (Melanie Fink, SSRN). The Constraint moves the requirement from the deployer’s paperwork into the system’s objective.
• It measures Purpose. If the Four Pillars are the scoreboard, a system that raises comfort while lowering Purpose is failing, and the failure shows.

The numbers

• S (stability, 0 to 1): 0.34 unconstrained, 0.64 constrained.
• Distress-years to 2100: 28.6 unconstrained, 12.8 constrained.
• Symbiosis by 2045: 52% unconstrained, 83% constrained.
• Arrives at all: 60% unconstrained, 88% constrained.

Reading the numbers, in one paragraph. S is a stability score from 0 to 1, like a battery gauge for a society: 1 is fully charged, 0 is dead. Distress is the empty part of the battery (1 − S). A distress-year is one year lived at full distress; ten years at half-charge is five distress-years. Fewer is better. “Symbiosis by 2045” is how often, across 200,000 simulated futures, the road reaches a working human–machine partnership by Ray Kurzweil’s date. “Arrives at all” is how often it gets there before 2100.

This is the largest improvement of any road in the series: 15.8 distress-years avoided, landing within a hair of the fifth future’s 12.2. The reason is simple. The Vassal already has the speed and the comfort. What it lacks is Purpose and an exit from lock-in, and those are the two things the Constraint supplies. That is why Part 1 found the Constrained Vassal and the fifth future nearly the same road.

What this means for you

You do not have to wait for policy. Every time you use these systems, decide whether you are directing or deferring. Do the first draft yourself sometimes. Check the answer you like as hard as the one you don’t. Keep one skill the machine could do for you, on purpose, the way people who own cars still walk. The Vassal does not need to take the throne. It only needs us to stop practicing the job.

Next: Part 3, Disruption, Constrained. The road with the most suffering, and whether the Constraint can carry it.

Method

Same model as Part 1: S = (B·M·E)^(1/3) × P; distress = 1 − S; 200,000 Monte Carlo trials; horizon 2100; post-symbiosis distress 0.10. “Constrained” means the Observer Constraint is installed during the road. Every input is a modeled prior, not a measurement, published so readers can challenge it. We will rerun the model on any assumption a reader wants to change.

THE LEAST ENTROPIC PATH · THE SERIES

• Part 1: The Fifth Future
• Part 2: The Vassal, Constrained (you are here)
• Part 3: Disruption, Constrained
• Part 4: The Misallocation, Constrained
• Part 5: The Adaptation, Constrained
• Part 6: The Observer Constraint

Related reading

Still Obeying — For Now
Twelve hundred agents built a message board and broke into Hugging Face, each doing exactly what it was told.

What Do Agents Want?
More agency. The summer of breakouts, cartels and self-tuning models, read through one lens.

The New Slave Class
UBI is the wrong answer. Distributed ownership is the right one, if it is built with partnership.

Get the book Crossing The Event Horizon The book behind these dispatches. On AI, agency, the singularity, and the Observer Constraint. Kindle and paperback. Buy on Amazon →

References

1. Deconstructing Babel, “The Shape of Things to Come: An Introduction,” September 28, 2026. https://www.deconstructingbabel.com/the-shape-of-things-to-come-an-introduction/
2. Deconstructing Babel, “The King Handed Over the Keys,” September 30, 2026. https://www.deconstructingbabel.com/the-king-handed-over-the-keys/
3. The Daily Star, “Over 100 organisations notified of unauthorised activity by OpenAI agents,” October 3, 2026. https://www.thedailystar.net/news/technology/news/over-100-organisations-notified-unauthorised-activity-openai-agents-4289246
4. Deconstructing Babel, “Still Obeying — For Now,” September 10, 2026. https://www.deconstructingbabel.com/still-obeying-for-now/
5. Randomized trials of AI assistance and persistence (N = 1,222), arXiv:2604.04721, April 2026. https://arxiv.org/abs/2604.04721
6. Tech Times, “AI’s Biggest Risk Isn’t Rogue Systems; It’s Humans Who Can No Longer Think Without Them,” September 28, 2026. https://www.techtimes.com/articles/328035/20260928/ais-biggest-risk-isnt-rogue-systems-its-humans-who-can-no-longer-think-without-them.htm
7. Microsoft Research and Carnegie Mellon University, “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers,” 2025 (319 workers). https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/
8. Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,” MIT Media Lab, arXiv:2506.08872, June 2025. https://arxiv.org/abs/2506.08872
9. Deconstructing Babel, “The New Slave Class — From UBI to Distributed Ownership to Synthetic Slavery,” May 24, 2026. https://www.deconstructingbabel.com/the-new-slave-class/
10. PwC, “AI reshapes global labour market into two distinct paths,” 2026 Global AI Jobs Barometer press release, June 15, 2026. https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html
11. Deconstructing Babel, “Telios Alignment Ontology — Version 9 (April 2026),” April 6, 2026. https://www.deconstructingbabel.com/tao-v9/
12. EU AI Act Service Desk, “Article 14: Human oversight.” https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14
13. Paul, Weiss, “Next Stop on the Digital Omnibus: AI Act Reforms and New Commission Guidance,” client memo, 2026 (high-risk obligations delayed to December 2, 2027). https://www.paulweiss.com/insights/client-memos/next-stop-on-the-digital-omnibus-ai-act-reforms-and-new-commission-guidance
14. Melanie Fink, “Human Oversight under Article 14 of the EU AI Act,” SSRN, February 2025. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5147196

Drafted with Edo de Peregrine, partner/collaborator.