AI is the technology that most directly offers to do your thinking for you — and that's exactly why the user's Solar Idealism framework treats it as the hardest case of its perimeter technology test (the user's own developing synthesis of Donovan, Billinge, and Bell, not an AI-ethics literature1). The framework's whole tech-ethic asks one question of any tool: does it strengthen me or weaken me? (Perimeter Technology Ethics). Applied to AI, that question gets its sharpest possible form, because AI is uniquely double-edged: it can extend your capability further than any tool in history, and it can atrophy the exact faculties it replaces faster than any tool in history. The same assistant that lets you do more than you could alone can quietly do the thinking you used to do yourself, until you can't anymore.
The framework's contribution here isn't an AI-safety theory (it has nothing to say about alignment or existential risk); it's a personal-sovereignty test for the AI in your own life — a way to decide, tool by tool and use by use, whether a given AI use builds you or replaces you. This page runs the framework's five-question solar test specifically on AI, where the strength-vs-weakness and sovereignty-vs-dependency tensions are at their maximum. It sits in the business domain because it concerns AI-assisted work and the AI-collaboration practice the vault houses there (the framework's behavioral-mechanics and business AI-tooling material), and it's a [MED] page because it's largely the perimeter-test specialized — the parent test does the heavy lifting; this applies it to the one tool that strains it most.
This is the AI-specific application of the framework's five-question solar tech test (Perimeter Technology Ethics). The five questions, run on AI:1 strength or weakness (does this AI use build my capability or atrophy the faculty it replaces?); sovereignty or dependency (am I more self-authoring, or more dependent on a system I don't control and don't understand?); ascent or comfort (does this free me toward harder work, or sedate me into letting the machine do the thinking I should be doing?); order or fragmentation (does this integrate my work or scatter it across a dozen half-delegated tasks?); perimeter integrity (do I command the AI, or does it colonize my judgment and surveil my thinking?).
The framework's distinctive move is the per-use granularity: AI isn't solar or anti-solar as a technology — the same tool is solar in one use and anti-solar in another, and the test is run on the use, not the tool. Using AI to extend a capability you have (drafting faster what you could draft, exploring more options than you'd have time to) reads solar; using it to replace a capability you should keep (letting it think so you don't have to, outsourcing the judgment that is your actual value) reads anti-solar. The page works that distinction, which is the framework's genuine contribution to the AI question and a useful corrective to both AI-utopianism ("it's all capability") and AI-doomerism ("it'll rot your brain") — the honest answer is it depends entirely on the use, and you have to test each one.
The framework's logic centers on one fork that AI makes sharper than any prior tool: does this use extend a capability or replace it? A tool that extends builds you — it lets you reach further with a faculty you keep exercising (the way a good telescope extends sight without atrophying the eye). A tool that replaces atrophies you — it does the thing so you stop doing it, and the faculty withers (the way GPS atrophies the sense of direction). AI sits exactly on this fork and can fall either way depending on use, which is why the framework insists on per-use testing rather than a verdict on the technology.
The strength-vs-weakness question is the sharpest because AI's atrophy-risk is uniquely deep: it offers to replace cognition itself — judgment, synthesis, writing, reasoning — which are the faculties the whole solar architecture treats as core to a sovereign person. Outsource your navigation and you lose a useful skill; outsource your thinking and you lose the thing the framework considers central to being a developed human. So the framework's AI-ethic is most alarmed about the uses where the machine does the cognition you should be doing — not because AI is evil, but because the faculty at risk is the one that matters most.
The sovereignty question runs close behind: AI deepens dependency on systems you neither control nor understand (the models, the platforms, the companies), and a person whose thinking is increasingly mediated by an opaque system they don't command has ceded sovereignty in the framework's most important domain. The logic's risk — the same Luddism-tilt the parent perimeter-test carries — is that the framework's alarm about cognitive-atrophy can curdle into reflexive AI-rejection, missing that AI used as extension (you still do the thinking, it amplifies the reach) is genuinely solar. The honest version holds the fork open: AI is the tool most capable of both building and replacing you, and the discipline is to keep choosing the uses that build.
The AI-test gives the corpus a per-use sovereignty diagnostic for AI that's sharper than the prevailing binary (AI-good vs AI-bad), and it connects the framework's tech-ethic to the vault's actual AI-collaboration practice — the business-domain material on working with AI (Taste/Judgment/Labor, voice-print, the question of which cognitive work to keep and which to delegate). The extend-vs-replace fork is a genuinely useful filter for that practice: delegate the labor that doesn't build you, keep the judgment that does.
Its second gift is the cognition-is-the-faculty-at-stake insight: the recognition that AI's atrophy-risk is uniquely serious because the faculty it offers to replace (thinking, judgment, synthesis) is the one most central to a sovereign person — which reframes "should I use AI for this?" as "is this a faculty I can afford to let atrophy?" That's a more useful question than the productivity-framing ("will this make me faster?"), because some speed-ups cost you the capability that was your actual value.
A writer adopts an AI assistant. At first the use is clearly extension: she drafts faster, explores more angles, gets past blank-page paralysis, uses the tool to amplify a craft she still owns. By the framework's test this is solar — she's reaching further with a faculty she keeps exercising. Her output improves and so, for a while, does her thinking, because the AI lets her try more.
Then the use drifts, as the framework's gradient-logic predicts. The assistant is so good at drafting that she starts letting it draft first and editing rather than composing. Then letting it structure her arguments. Then letting it do the synthesis — the actual thinking — while she curates. A year in, she notices she can no longer compose a paragraph from a blank page with the old facility; the faculty has atrophied because she stopped exercising it. The tool that started as extension became replacement, and by the framework's test the same tool crossed from solar to anti-solar — not because the tool changed, but because the use did, sliding down the convenience-gradient from "amplify what I do" to "do it so I don't have to."
The framework's read: this is the AI-atrophy fork in motion, and it's invisible while it happens because each step is a small, reasonable convenience. The discipline the test demands is to run the strength-vs-weakness question on each use as it drifts — to notice when "AI helps me write" has become "AI writes and I approve," and to recognize that the second has atrophied the faculty that was her actual value. The case study's payoff: AI's danger isn't a one-time decision (use it or don't) but a gradient you slide down use-by-use, and the test has to be run continuously, because the tool that builds you in March can be replacing you by December without any decision having been consciously made.
You use AI in your work, and the framework's discipline is to run the strength-vs-weakness question per use and over time, because the danger isn't whether you use AI but how the use drifts. You start by sorting your current AI uses into two piles, honestly. Extension uses: where the AI amplifies a faculty you still exercise — you do the thinking, it extends the reach (more options explored, faster execution of what you'd do anyway, past the blank-page block). Replacement uses: where the AI does the faculty for you — the thinking, the judgment, the synthesis that is your actual value — and you've stopped doing it yourself.
For the replacement-pile, you ask the framework's sharp question: is this a faculty I can afford to let atrophy? Some you can (you don't need to keep the skill of formatting citations; let the machine have it). Some you can't (the judgment, the synthesis, the core craft that is your value — let those atrophy and you've replaced yourself). For the faculties you can't afford to lose, you deliberately keep exercising them — you compose before you let the AI draft, you do the synthesis yourself before you check it against the machine, you keep the muscle the convenience-gradient wants to take.
Then the over-time discipline, which is the real one: you re-run the sort periodically, because uses drift from extension to replacement silently. The use that was amplifying your writing in March may be replacing it by December, and you only catch it if you check. You watch for the tell the case study names: when "AI helps me do X" has quietly become "AI does X and I approve," ask whether X was a faculty you needed to keep. The discipline isn't anti-AI; it's staying on the building side of the fork — using the machine relentlessly for what doesn't atrophy you, and guarding against it for what does, and re-checking because the line moves.
The model fails in two opposite ways. The first is the reflexive AI-rejecter — the practitioner whose alarm about cognitive-atrophy curdles into refusing AI entirely, missing that extension-uses are genuinely solar (they build reach without atrophying the faculty). This is the parent perimeter-test's Luddism applied to AI: the always-no that mistakes technophobia for sovereignty, and that cedes real capability to people willing to use the tool as extension. You recognize it by the principled refusal that costs the refuser capability they could have kept.
The second, more insidious, is the curator who replaced himself — the case study's writer, who slid down the gradient from extension to replacement without noticing, until the faculty that was his actual value had atrophied and he'd become a curator of machine-output rather than a thinker who uses a machine. You recognize it by the quiet loss: he can't do the thing anymore and didn't decide to stop; the convenience took the faculty one reasonable step at a time. This is the failure the framework is most usefully alarmed about, because it's invisible while it happens and irreversible by the time it's noticed.
The framework's own tension here is the same anti-comfort tilt the parent test carries: the framework's instinct toward "AI weakens us" can bias it toward the rejecter-failure and against the genuine extension-uses, while its sharpest real contribution (the per-use atrophy-fork) guards against the curator-failure. The honest practice keeps the contribution (test each use for atrophy) and resists the tilt (don't reject the tool wholesale) — staying on the building side of a fork the framework sometimes wants to refuse entirely.
Evidence. This is the user's synthesis specializing the perimeter-test to AI; the source stub flags the parent extensions section as suggestive.1 The extend-vs-replace / capability-atrophy concern has real and growing independent company (research and commentary on skill-atrophy from automation, "cognitive offloading," the deskilling literature), so the core insight is well-grounded. The framework offers nothing on AI safety/alignment/existential risk — it's strictly a personal-sovereignty tool, and shouldn't be mistaken for an AI-ethics in the policy sense.
Tensions.
The anti-comfort / AI-pessimism tilt. [TENS] Inherited from the parent perimeter-test: the framework's alarm about atrophy can curdle into reflexive AI-rejection, missing the genuine extension-uses. Document the atrophy-insight as valuable and the wholesale-rejection tilt as the bias to resist.
The per-use granularity is demanding and under-specified. The framework says test each use, but supplies little calibration for the hard judgment (which faculties can I afford to let atrophy?). The line between "extension" and "replacement" is real but fuzzy, and the framework leaves the user to draw it.
The scope limit. The framework's AI-ethic is personal-sovereignty only — it has nothing on the collective, structural, or safety dimensions of AI. A complete AI-ethics needs more than "does this build me?"; the framework supplies one useful question, not the whole field.
Open questions (tracked in META):
This page is the user's synthesis, specializing the framework's tech-ethic; the authors supply the disposition (largely Donovan's sovereignty-skepticism-of-soft-modernity and Bell's faculty-centric view of the developed person). The distinctive convergence is between the framework's sovereignty-and-capability core and the contemporary AI-deskilling concern: the framework was worried about tools atrophying human faculties long before AI made the worry acute, and AI is simply the sharpest instance of a concern the framework already held. Where the framework's own disposition pulls it toward AI-rejection, the genuine insight (per-use atrophy-testing) pulls toward disciplined use — and the honest synthesis takes the insight and resists the disposition, the same split the parent perimeter-test requires. The vault's verdict: a genuinely useful personal-sovereignty filter for AI (extend vs. replace, tested per-use over time), carrying the framework's reactionary tilt as the thing to check, and silent on everything about AI beyond the personal — useful within its scope, incomplete as an AI-ethics.
Rubber-duck version: this is the perimeter-test's hardest case, so it handshakes into the parent test (behavioral-mechanics), the sovereignty it protects (cross-domain), and the vault's AI-collaboration practice (business).
Behavioral-Mechanics: Perimeter Technology Ethics — the parent five-question test. What's identical: the same five questions. What differs: AI strains the strength-vs-weakness and sovereignty questions to their maximum, because it uniquely offers to replace cognition itself and deepen dependency on opaque systems. The insight: AI is the test's limit-case — the tool that most reveals why the strength-vs-weakness question matters, because it's the first tool that can atrophy thinking, the faculty the framework holds most central.
Cross-Domain: Sovereignty & Self-Authorship — the value at stake. What the handshake produces: the AI-test is the sovereignty-trait's frontier — self-authorship now has to be defended against a tool that offers to author for you, and the question "do I still do my own thinking?" becomes a sovereignty-question in a way it never was before AI could plausibly do the thinking. AI makes cognitive sovereignty a live discipline rather than an assumption.
Business: Decisive Action Bias (and the vault's AI-collaboration material) — the practice of working with AI productively. What the parallel unlocks: the framework's extend-vs-replace fork is the missing filter for AI-collaboration practice — delegate the labor that doesn't build you (where speed is pure gain), keep the judgment that does (where the faculty is your value). The test turns "use AI for everything you can" into the sharper "use AI for everything that doesn't atrophy what you're for."
The sharpest implication. AI is the first tool in history that can atrophy thinking itself — and the framework's genuinely useful contribution is to ask, of every AI use, not "will this make me faster?" but "is this a faculty I can afford to lose?" The destabilizing part is that the answer is often no and you're doing it anyway, because the atrophy is invisible while it happens: each delegation is a small reasonable convenience, and you only discover the faculty is gone when you reach for it and it isn't there. The writer who can no longer write from a blank page didn't decide to stop being able to write; she slid there one helpful draft at a time. So the live edge is a continuous discipline rather than a one-time choice: AI isn't a thing you decide to use or not — it's a gradient you slide down use-by-use, and staying on the building side requires re-running the strength-vs-weakness question on uses that silently drift from extension to replacement. The uncomfortable question isn't "should I use AI?" (you should, for what extends you) but "which of the faculties I'm delegating is the one that was actually me — and am I letting the machine have it one convenient step at a time?"
Generative questions: