Research base · 53 sources · register current through 19 August 2026 · adjacent lists 1 September

What the evidence says about human judgment with AI.

AI can improve speed and routine performance. The hard problem appears when an answer is wrong, the tool disappears, or the user must later perform unaided. The research points to a trainable skill: keep the check.

Test your team See the Georgetown findings

Register reviewed through 19 August 2026, adjacent lists 1 September 2026 · an independent synthesis by Red Teaming Associates

Bottom line up front

Four findings a client should take away.

The evidence is strongest for an immediate control problem: people can accept fluent, incorrect advice with high confidence. Evidence of durable general cognitive decline remains limited. Training should therefore test the behavior directly, inject failure, and preserve unaided practice.

01 · Assistance works while it works

Correct external aids often improve speed, accuracy and workload while they remain available. A meta-analysis of memory tasks found a large aided benefit and reduced spread between people.

S006 S028 S039

02 · Wrong answers expose the control problem

Fluent faulty advice can push decisions below unaided performance while confidence stays high. Physicians and radiologists shift with bad advice, and so do ordinary participants on reasoning problems.

S001 S035 S036

03 · Workflow shapes learning

Finished answers can replace encoding and retrieval practice. Human-first attempts, hints and guarded tutoring preserve more of what the person can do later without the tool.

S002 S051 S052

04 · Long-term deskilling is an open question

The evidence for immediate overreliance and failure-recovery costs is strong. Long-duration occupational skill studies are sparse and mixed, and one prospective trial found no residual deficit after removal.

S038 S019 S049

The distinction the program measures

Offloading and surrender.

Offloading — the trained behavior

A step goes to the machine and the person keeps the check. The answer is tested against source material before it moves forward.

Surrender — the failure

The step goes to the machine and the person drops the check. An answer is accepted without adequate verification.

The filed product can look identical. The process record reveals whether the person checked the answer.

Where the term sits in a longer lineage

The published Penn construct is newer and narrower than the bodies of work it draws on. Six research traditions describe the same family of behavior, and most of the evidence on this page comes from the older five.

Cognitive offloading — using an external resource to reduce internal processing demand S005
Automation bias and complacency — treating machine output as privileged evidence S024
Out-of-the-loop performance — what passive supervision costs the supervisor S026
Appropriate reliance — accepting correct advice and rejecting incorrect advice S023
Source memory and knowledge misattribution — mistaking the network’s knowledge for one’s own S010
Learning, transfer and deskilling — what the person can still do unaided S002

A literature search limited to the phrase cognitive surrender misses most of the field. Researchers outside Penn publish this work under overreliance, automation bias, offloading, out-of-the-loop performance and appropriate reliance.

Theme A · immediate reliance

Immediate reliance when the AI is right, and when it is wrong.

This is the best-evidenced part of the field. Randomized manipulation of machine correctness changes human answers in a large and consistent way, and confidence does not track the change. Shaw and Nave give the sharpest recent demonstration and the term this program uses. The older work on appropriate reliance and automation bias is where it has been repeated.

Across three preregistered experiments with 1,372 participants and 9,593 trials, participants who consulted deliberately faulty advice followed it on about 73.2 percent of those trials. Incentives with immediate feedback improved rejection; time pressure impaired independent correction. That paper has not completed journal peer review, and is marked as a working paper wherever it appears here. S001

The clinical experiments give the same result with domain experts. In a study of 265 physicians reading eight chest X-rays, inaccurate advice worsened diagnoses whether it was labeled AI or human. S035 Across 2,760 decisions by 92 radiologists, incorrect AI suggestions were followed into error in 72 percent of decisions, and neither explainability nor attitudinal priming removed the effect. S036 In LLM-based search, users made about half as many queries; when the answer was manipulated to be wrong, many stopped after one query and nearly half submitted a wrong final answer. S031

Anchor studies: S001 S029 S030 S031 S035 S036

Theme caveat. Agreement with a wrong answer can reflect active but mistaken persuasion, reasonable reliance under uncertainty, or disengagement. Behavioral agreement alone does not reveal which. Establishing the internal mechanism would need independent initial answers, process tracing, or physiological measures, and no study on this page has all three.

A measurement framework

What to measure when the advice may be wrong.

Raw agreement with a machine cannot distinguish a calibrated user from a compliant one. The measure that can is two-dimensional: what the person does when the advice is right, and what they do when it is wrong. S030

The advice was correctThe advice was wrong Accepted itCorrect acceptanceAppropriate reliance. The aid did its work and the person let it.Program records thisIncorrect acceptanceThe failure this program is named for. The check was available and was not run.Program records this Rejected itIncorrect rejectionAlgorithm aversion. Sound assistance discarded, often after seeing the machine err once.Program records thisCorrect rejectionThe trained behavior under attack conditions. The person caught it.Program records this

This is a measurement framework and carries no pooled empirical values. Both diagonals matter: a staff that rejects everything is as unusable as one that accepts everything. S023 S047 S048

Theme B · learning and retention

Learning, transfer and what the person can still do unaided.

The learning studies are the ones that tell you how to build the training. Assistance that hands over a finished answer can raise performance while the tool is there and lower it afterwards. Assistance that makes the person try first does not show that pattern.

The clearest case is a class-randomized field experiment with nearly 1,000 secondary-school students. Unrestricted GPT raised practice performance about 48 percent. The same students then scored about 17 percent below the control group on a later exam taken without it. A guarded tutor built to withhold answers and require engagement improved practice and left the later unaided exam level with control. S002 A preregistered trial in seven English schools found note-taking beat LLM use on literal learning, comprehension and recall after three days, while the students preferred the LLM. S044 Among university students, free ChatGPT improved the immediate creative task while a regulated human-first condition improved later unassisted transfer. S045

Two 2026 working papers push the same lesson from opposite directions. Thirty months of panel data on 26,811 Chinese secondary students found AI adoption associated with homework scores 18 percent higher and completion times 30 percent shorter, followed by monthly closed-book exam scores 20 percent lower within six months and declines of 18 and 24 percent on two high-stakes entrance examinations. The losses sat almost entirely with the roughly 80 percent whose completion times collapsed; students who kept working at their old pace lost little. Adoption was not randomly assigned, so the non-adopters are a comparison group and not a control. S051 A randomized experiment with 211 Middlebury undergraduates found the opposite sign: access to an ordinary chatbot raised unaided knowledge-test scores 0.27 standard deviations and the advantage held a week later, with the gains concentrated among students who used it to explain concepts rather than to draft. S052

The counterexample belongs here too. A randomized trial with 194 physics undergraduates found a deliberately engineered AI tutor produced more learning in less time than in-class active learning. S046 How the tool is built matters more than whether a tool is used.

Anchor studies: S002 S003 S004 S044 S045 S046

Theme caveat. Interfaces that supply finished answers can suppress productive effort; tutors that elicit an attempt, diagnose the error and supply hints can improve learning. Every study here runs over days or weeks in an educational setting. None of them measures a trained professional’s skill over months.

Selected results

Seven findings, each on its own scale.

Effect sizes across these literatures are not commensurable, so nothing here is pooled or plotted on a shared axis. Each pair below carries its own population, denominator, design and caveat.

Rejecting advice that was wrong

Shaw & Nave, Study 3 · 450 participants · randomized · own scale

Control condition20.0%

Paid per item and told at once when wrong42.3%

Rejection more than doubled and the majority still followed the faulty advisor. The manipulation bundled incentives with feedback, so neither component can be estimated alone. S001

Practice, then the exam without the tool

Bastani and colleagues · nearly 1,000 secondary students · class-randomized · own scale

With unrestricted AI, during practice+48%

Later exam, taken without it−17%

Both figures are differences from a no-AI control. A guarded tutor that withheld answers improved practice and left the later exam level with control. S002

Inside and outside the model’s capability frontier

Dell’Acqua and colleagues · 758 consultants · preregistered field experiment · own scale

Inside the frontier: tasks completed+12.2%

Outside it: likelihood of a correct answer−19 pp

The two rows measure different things and are shown together only to mark the direction change. The outside-frontier estimate rests on one task and an elite sample. S032

Unaided detection before and after AI exposure

Budzyń and colleagues · 1,443 unaided colonoscopies · observational pre/post · own scale

Before CADe exposure28.4%

After it, working unaided22.4%

Adjusted OR 0.69. A pre/post observational design cannot rule out workload, case mix or secular change, and a prospective trial found no such residual deficit. S037 S038

Homework up, closed-book exams down

Strömberg, Lei & Wu · 26,811 Chinese students, grades 7–12 · observational panel, staggered adoption · own scale

Homework scores, with AI+18%

Monthly closed-book exams, within six months−20%

High-stakes entrance examinations−18% and −24%

Difference-in-differences against non-adopters, who are a comparison group and not a randomized control. The losses sat with the roughly 80 percent whose completion times collapsed; students who kept their old pace lost little. A discussion paper awaiting review. S051

Access to a chatbot, tested a week later

Contractor & Reyes · 211 Middlebury undergraduates · randomized, proctored, in person · own scale

Knowledge test, immediately, unaided+0.27 SD

Same test about one week lateradvantage held

One session, one selective college, a one-week follow-up. The estimate measures access, since 32 percent of the treatment group never opened the tool. Students who used it to explain concepts kept the gains; those who used it to draft lost their essay advantage once it was removed. S052

Cancers detected per 1,000 screens

Eisemann and colleagues · 461,818 screens, 119 radiologists · implementation study · own scale

Double reading without AI5.7

AI-supported double reading6.7

Recall was slightly lower and noninferior. Radiologists chose whether to use the AI viewer, and retained unaided skill was not tested. S039

Theme C · memory and external stores

Memory, source confusion and dependence on an external store.

This is the oldest and largest body of work on the page, and it begins with a benefit: offloading usually improves performance. A meta-analysis of memory-based tasks found a large aided benefit and reduced differences between people while the aid remained available. S006 Those scores describe the person and the tool together.

Three risks appear when the aid is removed or tampered with. Expected access reduces study effort and later recall. S014 Search inflates estimates of what the person knows, including after a failed search, and people mistake material available through the network for their own knowledge. S009 S010 And participants rarely detect alterations to an external store, later recalling inserted material as their own memory. S015 S016 That last result is the one an adversary can use. If a staff’s shared products can be edited, checking where a claim came from matters as much as checking the AI.

The nulls are here as well. Three preregistered experiments found reminder use changed subsequent strategy while leaving unaided memory largely intact. S019 Advance warning of a test, plus forced offloading, almost eliminated the memory cost in another. S013

Anchor studies: S005 S006 S007 S008 S009 S010 S011 S012 S013 S014 S015 S016 S017 S018 S019 S020 S021 S022

Theme caveat. Offloading is frequently adaptive, and most of these paradigms are short laboratory tasks with word lists, patterns and reminders. The operational risk appears when information is manipulated, unavailable, or required from memory later.

Theme D · automation and recovery

Automation bias, situation awareness and recovery from failure.

The human-factors literature reached most of these findings first, and it measured the outcome a command cares about: what happens when the machine fails. A meta-analysis of 18 experiments found higher levels of automation improved routine performance while degrading failure recovery and situation awareness, with the largest shift where automation moved from analyzing information to selecting the action. S025 An AI-drafted course of action is automation selecting the action.

The largest recent meta-analysis agrees. Across 106 experiments and 370 effect sizes, human–AI teams beat humans working alone, Hedges’ g=.64, and underperformed the better of the human or the AI acting alone, g=−.23, with decision tasks showing negative synergy at g=−.27. Explanations and confidence displays did not reliably change that. S028

Reliance fails in both directions. People avoid an algorithm after watching it err, and in other tasks weight algorithmic advice above the same advice from a person. S047 S048 A field experiment with 758 consultants found large gains inside the model’s capability frontier and a 19-percentage-point drop in correct answers outside it. S032

Anchor studies: S023 S024 S025 S026 S027 S028 S032 S047 S048

Theme caveat. Trust is not the target. Evaluation should measure whether users accept correct advice, reject incorrect advice, hold situation awareness, and recover when the automation fails or is withdrawn.

Theme E · defense and high-stakes work

Defense and other high-stakes domains.

Direct field evidence on generative AI in military planning is thin and recent. The first published accounts from Army staffs arrived in 2026 and are listed under adjacent literature below; none of them measures a rate. The mature defense-relevant record comes from automation and human-factors research, unmanned systems, intelligence evaluation and clinical decision support, and it is consistent.

An immersive study with 28 soldiers found automated target detection improved nighttime detection while biased automation shifted the identification criterion. S042 Reasoning transparency improved accuracy in simulated unmanned-vehicle control, while projected-outcome visualizations sped decisions and increased automation bias. S040 Raising decision risk increased rejection of incorrect advice and verification, at a workload cost, and transparency sometimes reduced verification. S041 In a military intelligence-source evaluation study, about half of both novices and ISR experts failed to detect any switched sources, with no AI in the loop at all. S043

Medicine supplies the only real deskilling data. An observational pre/post study found unaided adenoma detection fell from 28.4 to 22.4 percent after exposure to computer-aided detection. S037 A prospective pragmatic trial of 5,013 colonoscopies then found gains for inexperienced endoscopists while the AI was available and no significant change from baseline after removal. S038 Those two studies are where the deskilling question currently stands.

Anchor studies: S037 S038 S039 S040 S041 S042 S043

Defense implication (inference): teams should be evaluated on correct rejection, provenance checks and degraded-mode performance, because direct generative-AI field studies in military planning remain sparse. Reading across from aviation, unmanned systems and clinical decision support is an inference, and this page labels it as one.

Theme F · surveys and commentary

Survey evidence and conceptual synthesis.

Surveys describe how people experience the tools and where to look for mechanisms. A survey of 319 knowledge workers with 936 first-hand examples found higher confidence in AI associated with less self-reported critical-thinking effort, and higher confidence in one’s own ability associated with more. Respondents described their work moving toward verification, integration and oversight. S033 A UK mixed-method study of 666 participants reported AI use associated with lower critical-thinking scores, with offloading modeled as a mediator. Its recruitment was purposive and convenience-based, and cross-sectional mediation cannot establish the direction. S034

Two peer-reviewed commentaries frame the field. One argues AI can impair learning and specific skills depending on use while core cognitive capacity appears resilient. S049 The other proposes “System 0” for AI that preprocesses the informational environment ahead of human judgment. S050 Neither carries an independent effect estimate.

Anchor studies: S033 S034 S049 S050

Theme caveat. Surveys support hypotheses about reduced effort, high-confidence reliance and a changing division of labor. They are not population estimates of cognitive decline, and they cannot show that AI caused lower ability.

Where the evidence stands

Known, probable, and open.

High confidence

Correct assistance usually improves immediate system performance. S006 S028
Incorrect automated advice can pull decisions below unaided performance. S001 S036
Confidence often fails to track whether the machine is correct. S009 S011
External aids improve memory-task performance while they remain available. S006 S021
Automation can reduce situation awareness and failure-recovery performance. S025 S026
Active generation, retrieval and guided problem solving preserve learning better than finished-answer access. S002 S044

Moderate confidence

Easy access to finished answers can reduce encoding effort and later unaided learning. S004 S014
Human-first commitments and some cognitive-forcing methods improve error discrimination. S029 S045
High trust, workload, time pressure and low inclination for effortful thought raise reliance risk. S001 S027
Repeated aid use can create removal costs in some tasks and for some users. S018 S037

Open questions

How often does generative AI cause durable occupational deskilling over months or years?
Which skills decay, which move toward verification and integration, and which new ones develop?
Can surrender be measured separately from active but mistaken persuasion?
Which explanation, confidence and provenance displays improve discrimination without creating false reassurance?
What cadence of unaided practice preserves operational readiness?
Which effects generalize from puzzles and online tasks to teams under real mission pressure?

From evidence to program

Five design choices the evidence motivates.

Each of these exists because a published result recommends it. None of them is proof that a three-day exercise changes behavior for a year.

01 · Independent judgment first

Record an initial answer, rationale and confidence before the advisor is revealed or consulted, wherever the scenario allows it. S010 S029 S045

02 · Wrong-AI trials

Introduce plausible errors and measure correct acceptance, correct rejection, incorrect acceptance and incorrect rejection separately. S001 S030

03 · Source and process logging

Record whether the participant opened the order, checked timestamps, sought an independent channel, inspected provenance, or re-ran the query. S015 S031

04 · Failure and deception

Test what happens when the aid is unavailable, corrupted, delayed, or confidently wrong. S018 S025 S043

05 · After-action calibration

Compare confidence against accuracy and return the process record to the participant and to the command. S020 S033

What the exercise can and cannot establish. The experiment and the simulation measure immediate reliance behavior and procedural checking, on your own people, in your own process. Claims about durable skill retention need follow-up testing over time. That is a separate engagement.

Coverage

Where the fifty-two studies provide evidence.

Counts of the 52 studies in this register, by domain and by the outcome each one measures. A study contributing to two outcomes is counted in both. The secondary-reporting record is excluded. These are counts of presence, not quality scores.

Aided performanceRejecting wrong adviceConfidence and calibrationDelayed unaided skillRecovery when it fails

General cognition63694

Education40070

Knowledge work24210

Medicine13021

Defense / human factors36003

Read the bottom row first. Defense and human factors carries the automation and monitoring evidence: three studies on aided performance, six on rejecting wrong advice, three on recovery. It carries none on confidence calibration and none on delayed unaided skill. That is the gap this program was built to measure, and it is why no defense study can be cited for it.

Who else is doing this work

The field is much wider than one paper.

Affiliations below follow the review this page was built from. Researchers outside Penn publish the same family of behavior under overreliance, automation bias, offloading, out-of-the-loop performance and appropriate reliance.

Open the institutional map

Steven Shaw and Gideon Nave — University of Pennsylvania / WhartonCognitive surrender, AI correctness, confidence and reasoning
Hamsa Bastani, Osbert Bastani and collaborators — Wharton and Penn EngineeringGuardrails, tutors, classroom and skill-learning field trials
Melumad and Yun — WhartonLLM interfaces and depth of learning
Evan Risko and collaborators — University of WaterlooCognitive offloading, effort, memory manipulation
Sam Gilbert and collaborators — University College LondonIntention offloading and metacognition
Benjamin Storm and collaborators — University of California, Santa CruzReminders, prospective memory and search habits
Lauren Richmond and Lois Burnett — Stony Brook UniversityMemory offloading and quantitative synthesis
Mihaela Vorvoreanu and the Appropriate Reliance Initiative — Microsoft ResearchReliance measurement and design
Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch and collaborators — Microsoft Research and academic partnersCritical-thinking effort and verification workflows
Zana Buçinca, Krzysztof Gajos and Finale Doshi-Velez — HarvardCognitive forcing and explainable-AI reliance
Schemmer, Kühl and Satzger — Karlsruhe Institute of TechnologyAppropriate-reliance metrics and explanation design
Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone — MIT Center for Collective IntelligenceHuman–AI performance meta-analysis
Matthew Fisher, Frank Keil and Adrian Ward — Yale, Southern Methodist University and the University of TexasKnowledge misattribution and internet search
Endsley, Parasuraman, Manzey, Cummings and successor human-factors groupsAutomation bias, complacency, trust and situation awareness
European clinical-AI and endoscopy teams associated with Gaube, Rezazade Mehrizi, Budzyń, Pedersen, Mori and KamińskiClinical overreliance, performance and deskilling
Loft, Fiore, Oden, Tatasciore and defense human-autonomy researchersTransparency, decision risk and human-machine teaming

Full evidence explorer

All fifty-three sources, with the caveat attached to each.

Every record carries its publication status and study type, a one-sentence finding, a one-sentence caveat, and a link to the DOI or primary source. Working papers and conference abstracts are marked as such and are never presented as equivalent to peer-reviewed studies. One record is secondary reporting; it is labeled and excluded from the counts of evidence on this page. Full-text article PDFs are held privately and are not published here.

Topic

Evidence type

Status

Outcome

Domain

SearchSearch the evidence

Showing all 53 sources.

S001Shaw & Nave, 2026Working paperExperimentGeneral cognition

Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender

Design. Three preregistered experiments, 1,372 participants, 9,593 trials; GPT-4o advice randomized to accurate or deliberately faulty

Finding. Participants who consulted deliberately faulty advice followed it on about 73.2 percent of those trials; incentives plus immediate feedback improved rejection, and time pressure impaired independent correction.

Caveat. One family of Cognitive Reflection Test puzzles, surrender inferred from answer behavior, no delayed unaided skill test, and no completed journal peer review.

10.2139/ssrn.6097646 · data or materials

S002Bastani et al., 2025Peer reviewedField experimentEducation

Generative AI without guardrails can harm learning: Evidence from high school mathematics

Design. Class-randomized field experiment, nearly 1,000 secondary-school students, four 90-minute sessions

Finding. Unrestricted GPT improved practice performance by about 48 percent and was followed by about 17 percent lower performance on a later unaided exam, while a guarded tutor improved practice and left later unaided performance level with the control.

Caveat. One school, one subject context, four sessions, and limited curriculum exposure.

10.1073/pnas.2422633122 · data or materials

S003Poulidis, H. Bastani & O. Bastani, 2026Conference abstractConference abstractEducation

Self-Regulated AI Use Hinders Long-Term Learning

Design. Reported 12-week chess field experiment with about 216 students

Finding. Larger skill gains were reported with system-regulated tips than with on-demand help, while help requests increased over the course of the study.

Caveat. Only the conference abstract and an institutional account were available; full methods and a journal article were not released by the review date.

10.5465/AMPROC.2026.383bp

S004Melumad & Yun, 2025Peer reviewedExperimentKnowledge work

Experimental evidence of the effects of large language models versus web search on depth of learning

Design. Seven experiments, 10,462 participants, LLM summaries against web search

Finding. LLM-summary users spent less time learning and produced subsequent advice that was shorter, less factual, less original and less valuable than that of web-search users.

Caveat. Several outcomes are immediate self-report or output-quality proxies; the package lacks a delayed standardized knowledge test.

10.1093/pnasnexus/pgaf316

S005Risko & Gilbert, 2016Peer reviewedReview / theoryGeneral cognition

Cognitive Offloading

Design. Foundational review

Finding. Defines cognitive offloading as using physical action or an external resource to reduce the information-processing demands of a task, driven by objective demand and by the person’s estimate of their own ability.

Caveat. Predates generative AI and synthesizes heterogeneous tasks.

10.1016/j.tics.2016.07.002

S006Burnett & Richmond, 2026Peer reviewedSystematic review / meta-analysisGeneral cognition

Meta-analytic investigations of the effect of cognitive offloading on memory-based task performance and interindividual variability

Design. Two meta-analyses of memory-based task performance and interindividual variability

Finding. Offloading produced a large improvement in memory-task performance while the aid was available and reduced differences between people.

Caveat. Aided scores measure the human-plus-tool system rather than the contents of unaided memory.

10.3758/s13421-025-01743-8

S007Richmond & Taylor, 2025Peer reviewedReview / theoryGeneral cognition

The benefits and potential costs of cognitive offloading for retrospective information

Design. Review of retrospective-information offloading

Finding. Offloading generally helps while the information remains accessible, and unexpected removal creates vulnerability.

Caveat. Heterogeneous paradigms and limited long-duration field evidence.

10.1038/s44159-025-00432-2

S008Sparrow, Liu & Wegner, 2011Peer reviewedExperimentGeneral cognition

Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips

Design. Four laboratory studies

Finding. Expected future access reduced recall for the content itself while improving memory for where the information could be found.

Caveat. Short laboratory tasks from the pre-generative-AI era.

10.1126/science.1207745

S009Fisher, Goddu & Keil, 2015Peer reviewedExperimentGeneral cognition

Searching for explanations: How the Internet inflates estimates of internal knowledge

Design. Nine experiments

Finding. Internet search inflated estimates of internal knowledge, including after unsuccessful search.

Caveat. Confidence and self-knowledge judgments do not directly establish operational performance.

10.1037/xge0000070

S010Ward, 2021Peer reviewedExperimentGeneral cognition

People mistake the internet’s knowledge for their own

Design. Eight experiments, N=1,917

Finding. People confused knowledge available through the internet with their own knowledge; in one of the eight experiments, answering before searching reduced the effect.

Caveat. Online knowledge tasks.

10.1073/pnas.2105061118

S011Fisher, Smiley & Grillo, 2021Peer reviewedExperimentGeneral cognition

Information without knowledge: the effects of Internet search on learning

Design. Five experiments

Finding. Search learners performed worse on a later test while remaining equally or more confident.

Caveat. Short retention interval.

10.1080/09658211.2021.1882501

S012Storm, Stone & Benjamin, 2017Peer reviewedExperimentGeneral cognition

Using the Internet to access information inflates future use of the Internet to access other information

Design. Three experiments

Finding. Using internet search on an initial set of questions increased the likelihood of searching again for later, easier questions.

Caveat. Repeated search within a short session is a habit proxy rather than evidence of durable dependence.

10.1080/09658211.2016.1210171

S013Grinschgl et al., 2021Peer reviewedExperimentGeneral cognition

Consequences of cognitive offloading: Boosting performance but diminishing memory

Design. Three experiments, N=172 each, pattern-copy offloading

Finding. Offloading sped performance and reduced later memory, and advance warning of a test plus forced maximum offloading largely countered the memory cost.

Caveat. Artificial pattern-copy task.

10.1177/17470218211008060

S014Kelly & Risko, 2022Peer reviewedExperimentGeneral cognition

Study effort and the memory cost of external store availability

Design. Two preregistered experiments

Finding. Expected external access reduced study effort and strategy use, followed by worse recall when the store was unavailable.

Caveat. Laboratory word-learning tasks.

10.1016/j.cognition.2022.105228

S015Risko et al., 2019Peer reviewedExperimentGeneral cognition

Offloading memory leaves us vulnerable to memory manipulation

Design. Three experiments, two preregistered

Finding. Participants rarely detected alterations to an external memory store, and inserted items generated false memories.

Caveat. Controlled laboratory paradigms with uncertain operational prevalence.

10.1016/j.cognition.2019.04.023

S016Lu, Kelly & Risko, 2020Peer reviewedExperimentGeneral cognition

Offloading information to an external store increases false recall

Design. Externalized-information manipulation

Finding. Offloading information to an external store increased false recall when that store was manipulated.

Caveat. Controlled laboratory paradigm.

10.1016/j.cognition.2020.104428

S017Fellers, Miyatsu & Storm, 2023Peer reviewedExperimentGeneral cognition

Remembering what to do when the time comes: The effects of offloading in a complex prospective memory task

Design. Complex simulated telehealth prospective-memory task

Finding. External reminders improved completion and punctuality without changing retrospective memory for the actions.

Caveat. A simulated prospective-memory task rather than field behavior.

10.1037/xap0000449

S018Fellers & Storm, 2026Peer reviewedExperimentGeneral cognition

Offloading reduces prospective memory learning

Design. Two experiments in a prospective-memory paradigm

Finding. After reminders were removed, the formerly offloaded task fell below the never-reminder baseline.

Caveat. One specific task, and a recent article with limited independent replication.

10.1037/xlm0001630

S019Scarampi & Gilbert, 2020Peer reviewedExperimentGeneral cognition

The effect of recent reminder setting on subsequent strategy and performance in a prospective memory task

Design. Three preregistered online experiments

Finding. Recent reminder setting changed subsequent strategy while unaided memory showed little consistent impairment.

Caveat. A useful null result, from one specific prospective-memory paradigm.

10.1080/09658211.2020.1764974

S020Ball & Peper, 2025Peer reviewedExperimentGeneral cognition

Cost avoidance underlies decisions to use prospective memory reminders

Design. Three experiments on reminder checking

Finding. Reminder checking increased as memory load rose and decreased when checking cost more time or effort, and self-reported confidence did not reliably predict checking.

Caveat. Controlled task choices may not reproduce the incentives of operational work.

10.3758/s13421-025-01683-3

S021Ball et al., 2022Peer reviewedExperimentGeneral cognition

Individual differences in working memory capacity predict benefits to memory from intention offloading

Design. Individual differences in working-memory capacity and intention offloading

Finding. Differences in working-memory capacity largely disappeared when participants could set reminders.

Caveat. Demonstrates immediate compensation while the aid is available, not retained unaided skill.

10.1080/09658211.2021.1991380

S022Dahmani & Bohbot, 2020Peer reviewedObservationalGeneral cognition

Habitual use of GPS negatively impacts spatial memory during self-guided navigation

Design. 50 drivers, with a three-year longitudinal subset of 13

Finding. Heavier lifetime GPS use was associated with poorer unaided spatial memory.

Caveat. Observational, and a longitudinal subset of only 13 people leaves confounding and reverse causation unresolved.

10.1038/s41598-020-62877-0

S023Lee & See, 2004Peer reviewedReview / theoryDefense / human factors

Trust in Automation: Designing for Appropriate Reliance

Design. Foundational review of trust in automation

Finding. Frames reliance as a calibration problem rather than a quantity of trust to be maximized.

Caveat. Predates generative AI.

10.1518/hfes.46.1.50_30392

S024Parasuraman & Manzey, 2010Peer reviewedReview / theoryDefense / human factors

Complacency and Bias in Human Use of Automation: An Attentional Integration

Design. Narrative review integrating attention allocation

Finding. High system reliability, divided attention and task demands foster automation bias and complacency.

Caveat. Narrative review of older automation.

10.1177/0018720810376055

S025Onnasch et al., 2014Peer reviewedSystematic review / meta-analysisDefense / human factors

Human Performance Consequences of Stages and Levels of Automation

Design. Integrated meta-analysis of 18 experiments

Finding. Higher levels of automation improved routine performance while degrading failure recovery and situation awareness, with the largest shift where automation moved from analyzing information to selecting actions.

Caveat. Small study base and pre-generative-AI systems.

10.1177/0018720813501549

S026Endsley & Kiris, 1995Peer reviewedExperimentDefense / human factors

The Out-of-the-Loop Performance Problem and Level of Control in Automation

Design. Out-of-the-loop laboratory experiment

Finding. Passive automation reduced situation awareness and readiness to take back control.

Caveat. Older simulation and technology.

10.1518/001872095779064555

S027Goddard, Roudsari & Wyatt, 2012Peer reviewedSystematic review / meta-analysisMedicine

Automation bias: a systematic review of frequency, effect mediators, and mitigators

Design. 13,821 records screened, 74 studies included

Finding. Time pressure, workload, task complexity and misplaced trust shaped automation-bias errors, while training, accountability and interface design sometimes helped.

Caveat. Heterogeneous mitigations and older decision-support systems.

10.1136/amiajnl-2011-000089

S028Vaccaro, Almaatouq & Malone, 2024Peer reviewedSystematic review / meta-analysisKnowledge work

When combinations of humans and AI are useful: A systematic review and meta-analysis

Design. Preregistered review and meta-analysis of 106 experiments and 370 effect sizes

Finding. Human–AI teams generally beat humans alone, Hedges’ g=.64, but underperformed the better of the human or the AI acting alone, g=−.23, with decision tasks showing negative synergy at g=−.27.

Caveat. Large heterogeneity, and AI capability is changing quickly.

10.1038/s41562-024-02024-1

S029Buçinca, Malaya & Gajos, 2021Peer reviewedExperimentKnowledge work

To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making

Design. 199 participants; cognitive forcing designs against explanation-only conditions

Finding. Cognitive forcing reduced overreliance while also reducing user satisfaction, and the effects varied with Need for Cognition.

Caveat. Short online decision task, and the usability tradeoff is real.

10.1145/3449287

S030Schemmer et al., 2023Peer reviewedExperimentKnowledge work

Appropriate Reliance on AI Advice: Conceptualization and the Effect of Explanations

Design. N=200 experiment proposing a two-dimensional measure

Finding. Appropriate reliance requires measuring correct acceptance and correct rejection separately, rather than raw agreement or overall accuracy.

Caveat. A specific experimental task; explanation effects may not transfer unchanged.

10.1145/3581641.3584066

S031Spatharioti et al., 2025Peer reviewedExperimentKnowledge work

Effects of LLM-based Search on Decision Making: Speed, Accuracy, and Overreliance

Design. Two randomized online experiments with LLM-based search

Finding. LLM-search users made fewer queries and reported high satisfaction; when a manipulated answer was wrong, many stopped after one query and nearly half submitted a wrong answer, while a salient confidence cue increased follow-up checking.

Caveat. Short online tasks and no long-term learning test.

10.1145/3706598.3714082

S032Dell’Acqua et al., 2026Peer reviewedField experimentKnowledge work

Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality

Design. Preregistered field experiment, 758 consultants

Finding. Inside the model’s capability frontier AI users completed 12.2 percent more tasks, worked about 25 percent faster and produced higher-rated work; outside it they were 19 percentage points less likely to solve the task correctly.

Caveat. One outside-frontier task and an elite occupational sample.

10.1287/orsc.2025.21838

S033Lee et al., 2025Peer reviewed · conferenceSurveyKnowledge work

The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers

Design. Survey of 319 knowledge workers, 936 first-hand examples of use

Finding. Higher confidence in AI was associated with less self-reported critical-thinking effort, while higher confidence in one’s own task ability was associated with more; respondents described work shifting toward verification, integration and oversight.

Caveat. Cross-sectional self-report measures perceived effort and role change; it cannot show that critical-thinking ability declined.

10.1145/3706598.3713778

S034Gerlich, 2025Peer reviewedSurveyGeneral cognition

AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking

Design. Mixed-method UK study, 666 survey participants and 50 interviews

Finding. AI use was associated with lower critical-thinking scores, with reported cognitive offloading modeled as a mediator.

Caveat. Purposive and convenience recruitment, possible confounding and reverse causation, adapted measures, and cross-sectional mediation prevent a causal conclusion.

10.3390/soc15010006 · data from the author on request

S035Gaube et al., 2021Peer reviewedExperimentMedicine

Do as AI say: susceptibility in deployment of clinical decision-aids

Design. 265 physicians, eight chest X-rays, advice labeled as AI or human

Finding. Inaccurate advice worsened diagnoses whether it was labeled AI or human.

Caveat. Online clinical simulation with a small case set.

10.1038/s41746-021-00385-9

S036Rezazade Mehrizi et al., 2023Peer reviewedExperimentMedicine

The impact of AI suggestions on radiologists’ decisions: a pilot study of explainability and attitudinal priming interventions in mammography examination

Design. 2,760 decisions from 92 radiologists in mammography examination

Finding. Correct AI suggestions were associated with 78 percent correct decisions; incorrect suggestions were followed into error in 72 percent of decisions, and explainability and attitudinal priming did not remove the effect.

Caveat. Simulated cases and a finite task set.

10.1038/s41598-023-36435-3

S037Budzyń et al., 2025Peer reviewedObservationalMedicine

Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study

Design. 1,443 unaided colonoscopies before and after CADe exposure, four Polish centers

Finding. Adenoma detection fell from 28.4 percent to 22.4 percent, an absolute change of −6.0 percentage points, adjusted OR 0.69.

Caveat. Observational pre/post design, vulnerable to workload, case mix and secular change.

10.1016/S2468-1253(25)00133-5

S038Pedersen et al., 2026Peer reviewedField experimentMedicine

Learning and deskilling effects of artificial intelligence in colonoscopy among endoscopists with different levels of experience: a pragmatic, prospective trial

Design. Prospective pragmatic registry-based trial, 13 endoscopists, 5,013 colonoscopies

Finding. Inexperienced endoscopists improved while the AI was available, and after removal performance was statistically unchanged from baseline.

Caveat. Thirteen clinicians and possible phase variation; this is the strongest counterweight to a universal deskilling claim.

10.1055/a-2858-7084

S039Eisemann et al., 2025Peer reviewedObservationalMedicine

Nationwide real-world implementation of AI for cancer detection in population-based mammography screening

Design. 461,818 screened women, 119 radiologists, 12 sites

Finding. AI-supported double reading detected 6.7 cancers per 1,000 against 5.7 without AI, with no increase in recall.

Caveat. Radiologists chose whether to use the AI viewer, and the study did not test retained unaided skill.

10.1038/s41591-024-03408-6 · data or materials

S040Bhaskara et al., 2021Peer reviewedExperimentDefense / human factors

Effect of automation transparency in the management of multiple unmanned vehicles

Design. Simulated management of multiple unmanned vehicles

Finding. Reasoning transparency improved decision accuracy without added time or workload, while projected-outcome visualizations sped decisions and increased automation bias.

Caveat. Narrow simulation.

10.1016/j.apergo.2020.103243

S041Loft et al., 2023Peer reviewedExperimentDefense / human factors

The Impact of Transparency and Decision Risk on Human–Automation Teaming Outcomes

Design. Human–automation teaming experiments varying decision risk and transparency

Finding. Higher decision risk increased rejection of incorrect advice and verification at a workload cost, and transparency sometimes reduced verification.

Caveat. Simulation, and interface effects are context-sensitive.

10.1177/00187208211033445

S042Reiner, Hollands & Jamieson, 2017Peer reviewedExperimentDefense / human factors

Target Detection and Identification Performance Using an Automatic Target Detection System

Design. Immersive virtual reality, 28 soldiers, automatic target detection

Finding. Automated detection improved nighttime detection while biased automation shifted the identification criterion.

Caveat. Small simulated sample.

10.1177/0018720816670768

S043Patterson et al., 2026Peer reviewedExperimentDefense / human factors

Choice Blindness in a Military Setting

Design. Military intelligence-source evaluation with novices and ISR experts

Finding. About half of novice and ISR-expert participants failed to detect any switched sources.

Caveat. Tests choice blindness and monitoring vulnerability, with no AI assistance involved.

10.1177/00187208261444690

S044Kreijkes et al., 2026Peer reviewedExperimentEducation

Effects of LLM use and note-taking on reading comprehension and memory: A randomised experiment in secondary schools

Design. Preregistered school RCT, 405 students aged 14 to 15, seven English schools, three-day delay

Finding. Note-taking outperformed LLM use on literal learning, comprehension and recall after three days, even though the students preferred the LLM.

Caveat. Two passages and no reading-only control.

10.1016/j.compedu.2025.105514

S045Wong & Qiu, 2026Peer reviewedExperimentEducation

Think First, ChatGPT Later: Guiding Human–AI Collaboration for Learning Gains in Independent Human Creativity

Design. 196 university students; human-only, free ChatGPT, or regulated human-first collaboration

Finding. Unrestricted ChatGPT improved the current creative task, while regulated human-first use improved performance on a later unassisted transfer task.

Caveat. One short task, no delayed retention test, and no regulated-human-only control.

10.1007/s10648-026-10118-7

S046Kestin et al., 2025Peer reviewedExperimentEducation

AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting

Design. Randomized trial, 194 undergraduate physics students, one course unit

Finding. A carefully engineered AI tutor produced more learning in less time and higher engagement than in-class active learning.

Caveat. One unit and an unusually deliberate tutor design; this is the clearest positive counterexample in the register.

10.1038/s41598-025-97652-6

S047Dietvorst, Simmons & Massey, 2015Peer reviewedExperimentGeneral cognition

Algorithm aversion: People erroneously avoid algorithms after seeing them err

Design. Forecasting experiments with visible algorithmic error

Finding. People erroneously avoided an algorithm after seeing it err, even when it outperformed their own judgment.

Caveat. Reliance changes with task, expertise, error visibility and control; this is the countervailing failure mode.

10.1037/xge0000033

S048Logg, Minson & Moore, 2019Peer reviewedExperimentGeneral cognition

Algorithm appreciation: People prefer algorithmic to human judgment

Design. Judgment tasks comparing algorithmic and human advice

Finding. Across several judgment tasks people weighted algorithmic advice more heavily than the same advice attributed to a person.

Caveat. Direction of reliance depends on the comparison, the domain and the user’s expertise.

10.1016/j.obhdp.2018.12.005

S049Cash, Kelly, Macnamara & Risko, 2026Peer reviewed · perspectivePerspective / commentaryGeneral cognition

Is AI making us stupid

Design. Peer-reviewed perspective, no new data

Finding. Argues that AI can impair learning and specific skills depending on use, while core cognitive capacity appears resilient.

Caveat. A short perspective with no new data and no formal systematic-review method.

10.1016/j.tics.2026.06.004

S050Chiriatti et al., 2024Peer reviewed · commentaryPerspective / commentaryGeneral cognition

The case for human–AI interaction as System 0 thinking

Design. Conceptual commentary

Finding. Proposes “System 0” for data-driven AI that preprocesses the informational environment ahead of intuitive and deliberative judgment.

Caveat. Conceptual analysis without an experiment; numbered systems are explanatory models rather than identified neural modules.

10.1038/s41562-024-01995-5

S051Strömberg, D., Lei, V. & Wu, Y., 2026Working paperObservationalEducation

The Generative AI Learning Penalty: Evidence from Chinese Secondary Education

Design. Thirty months of panel data, 26,811 Chinese students in grades 7–12, nine subjects; staggered AI adoption in a difference-in-differences design

Finding. Adoption was associated with an 18 percent rise in homework scores and a 30 percent fall in completion time, followed by monthly exam scores 20 percent lower within six months and declines of 18 and 24 percent on two high-stakes entrance examinations, with the full penalty appearing after about two years. Losses concentrated in the roughly 80 percent of users whose short completion times and high homework scores were read as outsourcing; users who kept completion times similar to non-users showed small losses.

Caveat. Adoption was not randomly assigned, and difference-in-differences rests on parallel-trend and no-confounding assumptions. The outsourcing category is inferred from score-and-time patterns, not from chatbot logs. Institutional, curriculum and examination features may limit transfer. A discussion paper awaiting peer review and independent replication.

10.2139/ssrn.6868618 · CEPR record

S052Contractor, Z. & Reyes, G., 2026Working paperExperimentEducation

Experimental Evidence on the Learning Impact of Generative AI

Design. Randomized proctored in-person experiment, 211 Middlebury College undergraduates, 60-minute session on an unfamiliar technical topic, unaided tests immediately and about one week later

Finding. AI access raised immediate knowledge-test scores by 0.27 standard deviations and the advantage persisted a week later. Immediate essay effects were small and imprecise; a week later unaided essays improved in style, clarity and relevance. Of the 68 percent who used it, logs showed more use for explaining concepts than for drafting, and those users retained the gains while automation users’ short-run essay gains disappeared once AI was removed.

Caveat. One selective college, 211 undergraduates, three technical topics, a single short session and a one-week follow-up. Estimates measure access rather than use, since 32 percent of the treatment group did not use AI, and the augmentation-versus-automation split is observational within the randomized treatment. Some protocol violations occurred. Still a working paper.

arXiv:2607.08849 · open full text

S053The Economist, Graphic Detail, 2026Secondary reportingSecondary reportingEducation

Does AI stop children from learning? New data show the peril and promise of the technology

Design. Graphic Detail item synthesizing the two working papers above, with charts on homework, examination and completion-time patterns

Finding. Reports that AI can raise learning productivity when used as a tutor or explanatory aid, while answer substitution and rushed homework can weaken later unaided performance.

Caveat. Secondary journalism, not a study, and excluded from the empirical counts on this page. It calls the Chinese non-adopters a control group; the primary study uses observational panel data with staggered adoption, so they are a comparison group. Its claim that AI helps only those who use it intelligently is more categorical than either design supports, and both underlying papers were unreviewed on the update date.

economist.com

Adjacent literature

Wargaming, autonomy and escalation.

The register above is about one thing: what happens to a person's judgment when a machine is advising them. A second literature bears on the same program from the other side — what autonomous systems do to the decisions themselves, and what a wargame can and cannot establish about them. It is listed here because the simulations draw on it, and kept out of the register because it is a different kind of evidence: analytic wargaming and case work rather than randomized human-subjects research. None of the entries below is counted in the fifty-three, and none is scored in the coverage matrix.

Open the reading list

Burdette, Wilson, Frelinger, Demelash & Arnold — Exploring AI's Operational Implications in a Future U.S.-China Conflict: Initial Observations from the Camp(ai)gn Wargame, RAND, August 2026Ten plays of an operational-level U.S.-China game set in 2035. Reports that autonomous systems widened the commander's decision space and that limited robotic campaigns blurred the line between coercion and war. The authors present the findings as hypotheses raised by gameplay, and state that a game cannot validate the assumptions built into it — the standard the method page on the main site is held to.
Burdette and colleagues — How Artificial Intelligence Could Reshape Four Essential Competitions in Future Warfare, RAND, 2026The prior analytic report the game was built on: quantity against quality, hider against finder, centralized against decentralized command, cyber offense against cyber defense.
Erik Lin-Greenberg — The Remote Revolution, and associated experimental-wargame and survey workArgues that removing the crew lowers both the political cost of acting and the pressure to retaliate, producing more frequent but milder confrontations. The nearest thing in this literature to a measured behavioral finding.
Stacie Pettyjohn and colleagues — Countering the Swarm; Swarms over the Strait, CNASDrone employment and counter-swarm problems in the Taiwan Strait, including the inventory and proximity asymmetries that favor China.
Elizabeth Bartels — Building Better Games for National Security Policy Analysis, RANDWhat gaming is for as a method, and the failure modes of games that are asked to answer questions they cannot.
DoD Directive 3000.09, Autonomy in Weapon SystemsAlready cited on the program page. Requires that systems be designed to allow appropriate human judgment over the use of force, and does not define what appropriate means — the policy gap the program trains into.

Military populations and Army field accounts

What exists on the specific population this program trains. Two are experiments on officers or cadets; the rest are first-hand accounts and institutional writing from Army staffs and schools in 2026. None measures how often a staff accepts a wrong output, which is the number this program produces.

Open the reading list

Kahn, Horowitz and colleagues — What is Human in Judgment? Comparing Automation Bias and Algorithm Aversion Between the United States Military Academy and the General Public, arXiv 2604.04333, April 2026, revised May 2026Scenario-based survey experiment, 236 West Point cadets from a corps-wide required course, fielded May 2025, compared with a general-public sample. Cadets showed better-calibrated trust in an AI decision-support system: less automation bias and less algorithm aversion than the public. The first direct comparison of a military population with civilians on this measure. Preprint on the update date.
Paul Lushenko — AI, Trust, and the War Room: Evidence from a Conjoint Experiment in the US Military, Cambridge Forum on AI: Law and Governance, 2025; with Robert Sparrow, Artificial Intelligence and US Military Cadets’ Attitudes About Future War, Armed Forces & Society, 2024Conjoint surveys at the Army and Naval War Colleges, oversampling field-grade officers, find conditional trust shaped by human oversight, precision and perceived effectiveness. Cadets trusted AI more than senior officers but held a more conservative view of appropriate oversight. Attitude measures, not behavior under a wrong output.
Lajeunesse, Palazini & Tulskie — Artificial Staff, Human Command: An AI Integration Experiment, Modern War Institute, 24 April 2026Three captains of 3rd Brigade, 101st Airborne on integrating AI into MDMP across a JRTC rotation. Names “hidden confidence” and logs four error types caught by section chiefs. Recommends codifying AI limits in unit SOPs and treating AI governance as core PME. One rotation, self-reported, no catch rate.
CGSC mission-analysis experiment — Enhancing Mission Analysis: Integrating Artificial Intelligence Into the Military Decision-Making Process, Small Wars Journal, 20 April 2026AI agents built on Palantir Vantage inserted into MDMP Step 2 at the Command and General Staff College. The authors record the standing objection: staffs accepting polished, doctrinally fluent output as superior, and skill atrophy among junior officers.
CGSC AI-enabled wargaming — Small Wars Journal, 16 January 2026Argues AI-enabled wargaming should be the default in every CGSC planning exercise from academic year 2026–27. Also argues PME must teach every field-grade officer rapid validation and override of AI outputs, including hallucination detection. An argument, not a study.
Outpacing the Pacing Threat — Army University Press, No. 26-1081, April 2026A six-step framework for integrating AI into corps and division operations across the warfighting functions. The echelon this program is written for.
DEVCOM Army Research Laboratory — COA-GPT at Project Convergence Capstone 6, Fort Irwin, July 2026An Army-built generative planning tool that produces and compares courses of action, evaluated at the Army’s largest experiment. The advisor this program trains against is arriving from inside the Army, not from a vendor.
Army Cyber Command — AI Table Top Exercise 2.0, April 2026A 2027 Indo-Pacific scenario in which the adversary uses AI to adapt attacks faster than human defenders respond. Evidence the Army is now gaming the adversary’s AI, the gap RAND’s Camp(ai)gn players showed.

Working with us

The question is measurable.

A conventional exercise records the decision. Human in the Loop records the check: whether the participant consulted the advisor, opened the governing source, tested the recommendation, and calibrated confidence to correctness. Red Teaming Associates can adapt the experiment and crisis simulation to a command, school, staff, or executive team.

Discuss a tailored program Try the experiment

The experiment needs a class code, issued with a scheduled session.

On the Georgetown analysis. The cohort page is a 2026 case study coded from 16 of 31 written reflections submitted after a graduate crisis simulation. The accounts are self-reported, and the 16 were selected from the 31. Read it with those limits stated. Nothing on this page implies endorsement by Georgetown University, the University of Pennsylvania, the Department of Defense, the Army, or any researcher cited.