Graduated AI Access and Judgment Atrophy
Question. Does graduated AI access sharpen or atrophy independent judgment, and where does it flip?
Analogy. It is the autopilot problem for judgment: some automation frees the pilot for the harder work, but fly everything on the autopilot and you land fine right up until the day it disengages and your hands have forgotten how. The study asks where that line sits for AI and judgment, how much help sharpens the operator and at what point leaning on it quietly hollows out the judgment it was meant to support.
What's at stake. Operators are going to use frontier AI. In a competitive field opting out is not realistic, and the economic pressure runs one way, so the design question is not whether AI is in the operator's work but where it is allowed to sit. Missionloops makes two deliberate choices about that: it keeps AI out of the operator's strategic thinking, and it keeps AI out of the emotional-support role, leaving both to the human guide. Both choices rest on one empirical question: does AI involvement degrade human judgment, and at what level of involvement does it start. This study takes the first axis, AI inside the strategic work. The same question is already live for DRDC: AI assistance is being standardised for analysts and commanders on the assumption that more help means better decisions, and nobody has tested whether that holds. If leaning on the AI quietly erodes the judgment it is meant to support, the institutions standardising it are weakening the exact capability they are trying to strengthen, and they would not notice, because each individual decision still looks fine in the moment.
The two answers it decides between. Either the capability lives in the tool, so judgment falls steadily the more the operator leans on the AI; or it lives in the person, in which case a moderate amount of AI sharpens judgment and only heavy reliance erodes it. Those are different curves, and the study tells them apart.
What a null result would mean. If access makes no difference in either direction, then the premise that AI reliance atrophies judgment, the premise the whole content-blind design rests on, needs revising. That is a finding about the theory, not a sign the platform was built wrong.
Why this matters to defence. This sits in DRDC Objective 3 (critical information for situational awareness and decision-making) through the decentralise-and-semi-automate focus area, and it maps onto the live IDEaS Cognition-and-Trust challenge. It changes a concrete decision: how much AI assistance to standardise for analysts and commanders, and whether human-autonomy-teaming doctrine should cap access to protect judgment rather than maximise it.
How we would run it. Three groups of operators work the same kind of real, high-pressure problems but with different amounts of AI help: none, some, and full. After a set number of missions we cut the AI off and test each operator's own judgment under pressure with a tool-denial wargame, against a fourth group that trained with no platform at all. The platform's content-blind design is what lets us dial the AI help up and down. We skip a calm, no-pressure group on purpose: how someone decides with nothing at stake is not the skill in question, and a trained operator's edge only shows when the pressure is real. The no-platform group separates the effect of the AI from the effect of just practising. We measure stress from the outside, not by asking, because a trained operator hides it: body signals like heart rate, ratings from an observer watching, and how well they actually perform.
Earliest start. Stage 10: the study needs the full frontier-AI access level wired in, so the AI dose can be dialled from none to full.