the three tests: Illich, discomfort, freezing

by Ray Svitla

you don’t need a philosophy degree to evaluate your personal AI stack. you need three questions. they take about ninety seconds to answer honestly, and they’ll tell you more about your relationship with your tools than any feature comparison ever will.

test 1: the Illich test (dependency)

can you burn it and still know who you are?

this is the foundational test, adapted from Ivan Illich’s convivial tools framework. Illich argued that tools cross a dangerous threshold when they stop enhancing human capability and start replacing it. medicine that makes you unable to understand your own body. education that makes you unable to learn without a teacher. tools that make the user helpless without the tool.

applied to personal AI:

imagine deleting everything. your self.md file. your AI memories. your custom instructions. your entire personal AI stack. gone. rm -rf the whole thing.

now: do you still know who you are? do you still understand your own values, your thinking patterns, your decision-making tendencies? or has your self-knowledge migrated into the tool, leaving you stranded?

passing looks like: you’d be annoyed. you’d lose some convenience, some useful routing, some accumulated self-observations. but the core self-knowledge — the stuff that actually matters — lives in your head, reinforced by the tool, not replaced by it.

failing looks like: genuine anxiety at the thought. the feeling that the tool knows you better than you know yourself. the realization that you’ve stopped doing the internal work because the external tool was doing it for you.

this is the most important test. a tool that fails it isn’t serving you. it’s colonizing you.

test 2: the discomfort test (growth)

does it ever propose something you don’t want to hear?

the discomfort principle in diagnostic form. personal AI trained on RLHF is structurally optimized to agree with you. the question is whether your specific setup has any mechanism — any at all — for productive friction.

review your last twenty interactions with your personal AI. count the ones where it:

• challenged an assumption you hold
• suggested an approach that contradicts your stated preferences
• pointed out a pattern you’d rather not see
• said “no” or “that’s not a good idea” in any form
• made you uncomfortable in a way that later proved useful

passing looks like: at least a few of those twenty interactions included genuine challenge. your AI occasionally routes you toward the uncomfortable — a philosophy that contradicts yours, a pattern observation that stings, a suggestion that goes against your grain.

failing looks like: twenty interactions of smooth agreement. validation. “great question.” “that’s a thoughtful approach.” the AI as yes-man, confirming your existing worldview with increasing eloquence.

a personal AI that never challenges you isn’t personal. it’s commercial. it’s optimizing for engagement, not growth. and engagement without growth is just a more sophisticated narcissism trap .

test 3: the freezing test (process)

does it model you as a static profile or a dynamic process?

this test checks whether your tool treats identity as a process or a profile . frozen identity — labels, types, fixed preferences — calcifies you. dynamic identity — tensions, oscillations, evolving patterns — preserves your ability to change.

look at how your personal AI describes you. not what you told it to say — what it actually uses when making decisions. is it:

a set of labels? (“introvert, values efficiency, prefers concise communication”)

or a set of tensions? (“oscillates between deep solitude and intense collaboration, currently leaning toward solitude after a high-collaboration month. values efficiency but periodically rebels against optimization itself.”)

passing looks like: the system changes its model of you over time. it notices shifts. it proposes updates. your identity in the system from six months ago looks different from today, and the difference tracks something real about how you’ve changed.

failing looks like: your AI treats you the same way it did when you first set it up. your preferences are frozen. your profile is static. the system optimizes for a version of you that no longer exists — or that never fully did.

freezing is Illich’s insight applied to identity: a tool that locks you into a self-image is a tool that has crossed the watershed. even if the self-image was accurate when you wrote it, it’s inaccurate now. because you’ve changed. because you always change. because the self is a process and anything that models it as a fixed object is lying.

scoring

no numbers. no percentages. just a gut check:

test 1 (burn):       pass ○  |  fail ●
test 2 (discomfort): pass ○  |  fail ●
test 3 (freezing):   pass ○  |  fail ●

○ ○ ○ — your personal AI stack is convivial. keep going.

● ○ ○ — dependency forming. you’re getting value but losing autonomy. start building self-knowledge that exists independently of the tool.

○ ● ○ — sycophancy trap. your AI is comfortable but useless for growth. add friction mechanisms.

○ ○ ● — frozen identity. your tools work but they’re optimizing for who you were, not who you’re becoming. update your self-model, or better yet, build a system that updates itself.

● ● ● — you don’t have a personal AI tool. you have a personal AI dependency. Illich would not be proud. the good news: awareness is the first step. the burn test is both diagnostic and treatment.

applying the tests to self.md

self.md is designed to pass all three. that doesn’t mean it always will — any tool can be misused. but the architecture tries:

burn test: it’s a markdown file. you can delete it. the protocol thesis demands that the file format never creates lock-in. the intelligence layer (routing, catalog, AI matching) is optional. the file is always just a file.

discomfort test: the catalog includes contradictions by design. the routing layer has a discomfort budget. the journal AI flags repeated patterns. discomfort is architectural, not accidental.

freezing test: the .journal/ tracks change. the AI proposes diffs. the identity model is explicitly temporal — it has a present state and a history. tensions replace labels. process replaces profile.

but don’t trust the architecture. run the tests yourself. regularly. on self.md and on everything else in your stack. the tests work on any tool, any platform, any approach to personal AI.

that’s the point. the tests are more important than the tool they evaluate.

→ convivial AI — the full Illich framework → the discomfort principle — why challenge matters → self as process, not profile — the philosophy behind the freezing test → the narcissism trap — what happens when all three tests fail

Ray Svitla stay evolving