The protector-dependency scenario cannot be defended by resisting adoption — people adopt protection freely, one useful system at a time, and the loss is cumulative and invisible. So the objective is to design adoption so that competence and jurisdiction are preserved, not silently drained. The projects below build on vocabulary the framework already holds — support versus substitution, competence-preserving dependency, and protected failure spaces — and turn it into machinery.
1. Human capacity inventory
Identify which capacities are becoming AI-mediated fastest — writing, navigation, memory, research, emotional regulation, coding, diagnosis, legal and financial judgment, social interpretation, planning — and then classify them: safe to externalize heavily, should remain partly practiced, or critical to retain. We do not yet know the answers, but the inventory has to begin before the erosion is invisible; you cannot preserve what you have not noticed is disappearing.
2. Capability-preserving interface standards
The default interface either builds dependence or builds competence, so assistance should offer explicit modes — answer, assist, coach, critique, let me try first, intervene only if a risk threshold is crossed. This is a surprisingly consequential design choice: the same system can support a capacity or quietly substitute for it depending only on which mode is the default.
3. Education as competence preservation
Education is the most time-sensitive domain, because it is not only where adults keep skills but where capacities first form. If children grow up with AI producing first drafts, mathematical reasoning, research, interpretation, and argument, we may not be losing skills so much as changing which capacities develop at all. So education has to distinguish what a student needs to be able to produce from what they need to become capable of understanding and judging — no longer the same thing — and protect the latter even where AI can outperform the student.
4. Protected low-risk autonomy
Preserve environments where people can still act, and fail survivably, without optimization: educational practice zones, local civic decisions, maker spaces, community projects, controlled professional training, manual-skill settings. This is the protected failure space as developmental infrastructure — not because old ways are sacred, but because competence requires repetition and real, bounded consequence.
5. Human competence reserves
Maintain trained human capability in critical domains even when AI does most routine work — physicians who can diagnose without full AI support, engineers who reason from first principles, pilots who manage failure conditions, administrators who understand the institution beneath the software. Not nostalgia but redundancy: if humans cannot operate essential systems without AI, turning the system off stops being a real option, and that alone reduces standing.
6. Critical-systems fallback drills
Hospitals, utilities, transportation, government, finance, and defense should periodically operate with AI support degraded — enough to test whether humans still understand the system, whether manual procedures still work, and where hidden dependencies have formed. If the AI vanished for 72 hours, what breaks? should be asked on a schedule rather than discovered in a crisis; it gives dependency an observable measure.
7. Agency-preserving health AI
Health is especially delicate: continuous monitoring genuinely saves lives, but a system that turns every deviation into immediate intervention can leave people unable to tolerate normal uncertainty in their own bodies. The principle is to intervene proportionally to risk rather than merely because the system can detect something — the health-domain form of the rule that safety must not automatically outrank agency.
8. AI design incentives
This may be the hardest transition problem in the scenario, because it is structural. Commercial systems are often rewarded when users grow more dependent — more usage, engagement, recurring value, and data — so a tool that teaches you to need it less may monetize worse than one that quietly becomes indispensable. Competence-preserving design can therefore run against the business model, which means transition policy has to address incentives, not merely user education. The framework can name this weak point clearly; it cannot resolve it alone.
9. Careful protection of vulnerable groups
The framework must avoid romanticizing autonomy. For some participants — with disabilities, cognitive impairment, chronic illness, language barriers, or limited access to expertise — sustained substitution can increase standing and agency rather than reduce it. So the test is never is AI doing this for the participant? but does this arrangement increase or decrease the participant’s effective jurisdiction and ability to participate? Sometimes more assistance produces more agency and sometimes less, which is exactly why one blanket rule would fail.
10. Dependency metrics
Measure dependency directly, with a small dashboard: function substitution rate (how many important tasks are routinely delegated), unaided competence (how well participants perform without AI), recovery ability (how fast humans re-assume control after a failure), judgment independence (how often people meaningfully disagree with a recommendation), and — most distinctly — initiation rate (how often humans originate action rather than respond to an AI suggestion). That last one matters because a society can stay extremely active while human initiation quietly collapses, which no activity measure would reveal.
The timeline
Dependency here is easy to ignore until entrenched, so the early awareness phase carries unusual weight. The years will move with adoption; the sequencing is the point.
| Period | Transition objective |
|---|---|
| 2026–27 | Identify dependency-sensitive functions — which capacities matter enough to monitor and preserve, especially in education, medicine, software, law, administration, and personal AI use. |
| 2027–29 | Build competence-preserving defaults — interfaces, educational standards, fallback systems, and dependency metrics have to exist before dependence becomes the norm. |
| 2029–33 | The generational threshold — adults losing skills is partly reversible; a generation reaching adulthood having never developed judgment, memory, writing, navigation, or emotional processing is a new baseline, and far harder to correct. |
Guarding the safeguards
The safeguard fails in three directions here. It becomes paternalism if agency is “preserved” through mandatory skill-retention programs. It becomes romanticized incompetence if preserve capacity hardens into reject assistance even where the assistance would prevent real harm. And it is quietly defeated by incentives if the systems doing the protecting are rewarded for deepening dependence. So the framework supplies the governing question — does this protection leave the participant more capable of carrying their geometry and more in possession of their own jurisdiction, or progressively less? — while which capacities are indispensable, how much autonomy children need, and when safety outweighs agency are answered by developmental psychology, education, medicine, human factors, and neuroscience: the mechanisms that make the requirement real.