A behavioral self-assessment of how much of your thinking, deciding, and remembering has shifted onto AI tools. Inspired by cognitive-offloading research (Risko & Gilbert, 2016, Trends in Cognitive Sciences), the 2025 MIT Media Lab 'Your Brain on ChatGPT' EEG study (Kosmyna et al.), and the emerging literature on AI dependence and critical thinking (e.g., Gerlich, 2025), this test measures five dimensions: cognitive offloading — letting AI do the thinking before you've tried; skill atrophy — abilities that have softened from disuse; unverified trust — passing AI output onward without checking it; life creep — AI spreading from work tasks into feelings, relationships, and health decisions; and cognitive self-doubt — losing confidence in your own unassisted mind. Items are behavioral rather than abstract — they ask about what you actually do, not what you believe about technology. Higher scores indicate more dependence in that dimension. Dependence is a spectrum, not a diagnosis, and every pattern this test measures responds to deliberate practice. For self-reflection and education only, not a clinical diagnosis or a substitute for professional support.
Your result is broken down across 5 distinct dimensions rather than collapsing everything into a single number. Each dimension is scored from its own set of items, so you can see which specific pattern is driving your overall result.
Scores fall into one of 5 interpretive bands. Here is what each one means, what tends to sit underneath it, and what usually helps.
Sovereign Mind
AI is a tool in your hand, not a tenant in your head. Your profile suggests you still do your own first-pass thinking, form your own drafts and judgments, and bring AI in selectively — if at all. The core cognitive loops this test measures — starting from nothing, verifying claims, deciding, remembering — are still running on your own hardware. Whatever assistance you use, you remain the author of your own conclusions.
What it suggests
Low scores here usually mean the 'struggle phase' of thinking is still intact for you — you can tolerate not-knowing long enough to work things out, which is exactly the phase where learning and skill consolidation happen. Cognitive-offloading research (Risko & Gilbert, 2016) suggests offloading itself is not harmful — it frees capacity — but only when the underlying skill is maintained and the offloading is a choice rather than a reflex. Your subscale scores can still show a leading edge: whichever dimension is highest is where a habit may be quietly forming.
What helps
There's nothing to fix, but there is something to protect. Keep deliberately doing some hard things unassisted — first drafts, mental arithmetic, navigating from memory — the way athletes keep training movements that machines could do for them. If you adopt new AI tools, decide in advance what they're for and what stays manual. And keep your verification habit: it is the single strongest differentiator between low and high scorers on this test.
Tool User
You use AI regularly and it's genuinely woven into your work, but the thinking still starts and ends with you. This is the most common profile among people who have adopted AI seriously — real integration, with the important checkpoints still manned. Your scores suggest you still draft, decide, and verify for yourself more often than not, with a few specific areas where the balance has started to shift.
What it suggests
In this range, the pattern to watch is not overall usage but which subscale is climbing. A high Cognitive Offloading score with low Unverified Trust means you delegate work but still audit it — a healthy division of labor. The reverse — trusting without checking — is the pattern the AI-dependence literature associates with declining critical-thinking engagement (Gerlich, 2025). Dependence isn't a switch; it's a slope, and this is the part of the slope where habits are still cheap to change.
What helps
Pick your highest subscale and run one deliberate 'manual mode' experiment against it this week: write one email from scratch, sit with one hard problem for fifteen minutes before asking, or verify one AI claim against a primary source. Keep AI for what it is genuinely better at — breadth, boilerplate, recall — and keep the judgment calls, the first drafts of things that matter, and the final verification for yourself. The lever at this stage is small and cheap: verification plus occasional unassisted reps.
Co-Pilot Mode
AI has moved from a tool you pick up to a partner you think with — and part of the thinking now happens on its side of the table. Your profile suggests significant portions of your drafting, deciding, or verifying have shifted to AI, and that working without it has started to feel noticeably harder. This is not a failure state; it's the natural equilibrium of tools this good. The question your scores raise is whether the delegation is still deliberate — or whether it has become the default.
What it suggests
The 2025 MIT Media Lab 'Your Brain on ChatGPT' study (Kosmyna et al.) found that people who wrote essays with LLM assistance showed weaker neural connectivity on EEG and struggled to recall or quote their own text minutes later — the work got done, but less of it left a trace in the writer. That is the mechanism to watch in this band: output quality stays high while internal ownership quietly thins. Your subscale breakdown shows where — a high Skill Atrophy score points to capacities already softening, while high Unverified Trust means errors may be flowing through you to other people, unchecked.
What helps
Reintroduce deliberate friction where it matters most. Choose two or three task types — important messages, decisions about people, anything you'd be embarrassed to get wrong — and make them manual-first: your draft or your judgment exists before AI sees the problem. Restore verification as a non-negotiable step for anything you'll repeat to another person. And schedule regular unassisted reps for whatever your highest Skill Atrophy area is; skills respond to practice at any stage, and the discomfort of the first few sessions is the signal that it's working, not a reason to stop.
Outsourced Thinking
Most of the cognitive work this test measures — starting, drafting, deciding, verifying, remembering — now routes through AI first, and your own unassisted capacity has become the backup system rather than the primary one. Your profile suggests that working without AI doesn't just feel slower; it feels genuinely difficult, and that AI's judgment frequently overrides your own. This level of reliance usually builds gradually and rationally — each individual delegation made sense at the time — which is exactly why it rarely feels like a decision anyone actually made.
What it suggests
At this level, the two dimensions most worth separating are capability and confidence. A high Skill Atrophy score describes skills that have genuinely softened from disuse — and disuse, not damage, is the driver, which means they respond to practice. A high Cognitive Self-Doubt score describes something different: a loss of trust in your own thinking that often outpaces any real loss of ability. The cognitive-offloading literature (Risko & Gilbert, 2016) is consistent on this point — offloaded functions fade from lack of use, and the capacity to rebuild them stays intact. Unverified Trust is the most urgent subscale if it's high, because its costs land on other people, not just on you.
What helps
Treat this like rehab, not abstinence — the goal is redistribution, not renunciation. Start with verification, the cheapest and highest-leverage fix: no AI claim gets repeated, submitted, or acted on without one independent check. Then carve out a daily 'manual mode' block — thirty minutes of real work with AI closed: a first draft, a calculation, a decision worked through on paper. Expect the first two weeks to feel genuinely unpleasant; that friction is the retraining itself, not a sign it isn't working. And if AI has also become your main confidant for emotional or health decisions, deliberately re-involve humans in those loops — a friend, a doctor, a therapist — not because AI's answers are always wrong, but because those decisions deserve counsel with stakes in you.
AI Runs the Show
Your answers describe a working life — and possibly an inner life — in which AI sits upstream of nearly everything: what you write, what you decide, what you believe, and how you feel about your own mind. Unassisted work may feel close to impossible, AI's answers may function as settled truth, and your confidence in your own thinking has likely taken real damage. Scores this high deserve to be taken seriously — not as a verdict on your intelligence, but as an accurate reading of how completely one set of tools has been handed the wheel.
What it suggests
The most important thing your profile suggests is that the dependence and the self-doubt are now feeding each other: the less you do unassisted, the less you trust yourself, and the less you trust yourself, the less you attempt unassisted. That loop — not any single skill — is the real finding. The research base here is genuinely hopeful: offloaded skills fade from disuse rather than being destroyed (Risko & Gilbert, 2016), and confidence in your own thinking rebuilds through small, successful unassisted attempts. Nothing this test measures is a fixed trait. If your Life Creep score is also high — AI mediating your relationships, health choices, and emotional processing — that is the dimension to address first, because it carries the most real-world consequence.
What helps
Start smaller than feels necessary and be more consistent than feels reasonable: one AI-free task a day, chosen to be easy enough that you will actually succeed — a short message written alone, a sum done in your head, one fact checked by hand. Rebuild verification before anything else: for one week, check every AI claim you intend to repeat; you will likely catch errors, and every error you catch is direct evidence that your judgment still works. If AI has become your primary counsel for emotional, relational, or medical decisions, bring humans back into those specific loops now — and if the anxiety of being without AI feels unmanageable, that is a legitimate thing to bring to a therapist, who will treat it as the workload-and-confidence issue it is, not a character flaw. The capacity you're worried you've lost is recoverable; disuse is not damage, and the first unassisted wins arrive faster than the current level of self-doubt predicts.
AI Dependency Test contains 33 items and takes about 7 minutes. Responses are combined using a subscale weighted method across the dimensions listed above, then mapped to the interpretive bands above. Your percentile is an interim estimate that sharpens as more people take the test, and is labelled as provisional in your result.
This assessment also includes hidden validity items that detect inattentive or socially-desirable answering. When they trigger, your result carries a lowered confidence rating and we tell you — rather than presenting a skewed score as fact.
The test ships in 3 parallel forms that rotate between takes, so repeat visitors do not simply memorise the same items.
This is an educational, self-reflective assessment — not a diagnosis. A questionnaire can describe a pattern and point you toward language for it, but it cannot account for your history, context, or circumstances, and it is not a substitute for evaluation by a qualified professional. If a result concerns you, treat it as a reason to talk to someone rather than as a verdict.
Your answers are scored on your device. Taking this test requires no account, no email address, and no payment, and your full result — every band, every dimension — is shown to you immediately and in full.