Between September 2021 and March 2022, researchers inside the ACCEPT clinical trial tracked experienced endoscopists across four Polish medical centers before and after artificial-intelligence polyp-detection overlays entered daily rooms (Warwick WRAP / Lancet Gastroenterology & Hepatology, Aug 2025; STAT News, Aug 12, 2025). Krzysztof Budzyn and colleagues did not grade doctors while the AI assist was turned on. Instead, they audited 1,443 standard, non-AI colonoscopies: 795 unassisted procedures before AI rollout and 648 unassisted procedures after clinicians grew used to green AI boxes flagging tissue. Once AI became routine, the doctors' unaided adenoma detection rate fell from 28.4% to 22.4% across those 1,443 cases, a 6.0 percentage-point drop (multivariable odds ratio 0.69). Experienced eyes had quietly outsourced the first scan.
AI deskilling is a measurable drop in unaided human judgment after routine reliance on automated drafts or alerts, not permanent "brain rot." Three 2025 studies document real vigilance decay and a 39-point speed illusion across colonoscopies, open-source coding, and student essays, though each study carries strict design limits. Below is the ed.cafe Weekly Cold Rep Anti-Deskilling Protocol for Teams & Apprentices.
Workers already sense the risk even when viral posts exaggerate the neuroscience. In EY's Work Reimagined Survey (n = 15,000 employees and 1,500 employers in 29 countries), 37% of surveyed workers worried over-reliance on AI could erode their skills, while 64% of employees reported heavier workloads (EY, Dec 8, 2025). Globally, KPMG and the University of Melbourne (n > 48,000 respondents across 47 countries) found 66% of employees rely on AI output without evaluating accuracy, and 56% of workers have made job mistakes due to AI (KPMG, Apr 2025).
You leave with four tools: the human-factors mechanism explaining why editing AI drafts dulls error detection, a plain-English audit of all three 2025 studies and their exact design limits, an apprenticeship cost model for a 12-person team, and The ed.cafe Weekly Cold Rep Anti-Deskilling Protocol for Teams & Apprentices. Pair this guide with How Often Is AI Wrong? and our AI Safety at Work Hub.
Why does letting AI write the first draft dull human error detection?
Letting generative AI produce every first draft erodes human sharpness through two documented mechanisms: vigilance decay during passive review and losing the mental scaffolding built while wrestling with a blank page. Skimming fluent machine prose exercises a weaker cognitive loop than structuring an argument, query, or diagnosis yourself.
Engineer Lisanne Bainbridge named this paradox in 1983 in Ironies of Automation: the smoother an automated system looks during routine minutes, the harder it gets for a human monitor to stay sharp and step in during edge-case failures (Automatica, 198390046-8)). Modern desk work repeats that trap daily. In a 2025 Microsoft Research and Carnegie Mellon survey of n = 319 weekly GenAI workers across 936 examples, higher confidence in AI correlated directly with lower critical-thinking effort (The Register, Feb 11, 2025).
First-draft offloading works like driving two hundred miles on a dry highway with lane-centering assist engaged. When your hands make tiny steering corrections every second, your eyes track lane geometry and braking gaps automatically. Rest two fingers passively on a self-centering wheel for hours, and your reaction lags when construction cones bend the lane at dusk.
Where that highway analogy breaks is warning chimes. A car beeps loudly and flashes a red icon the second road paint vanishes. Chatbots never chime when your prompt crosses outside their competence. Worse, as your independent recall of tax rules, SQL joins, or clinical differentials rusts over months, the machine's polished paragraphs start looking more convincing to you rather than less.
What did the colonoscopy, open-source coding, and MIT essay studies actually measure?
Three 2025 studies drive nearly every argument about whether AI makes workers slower or less observant, yet viral threads routinely blur clinical observation, developer randomized trials, and non-peer-reviewed student preprints. Checking each study's sample and design prevents both panic and complacency.
1. Budzyn et al. (Lancet Gastroenterology & Hepatology, Aug 2025): unaided clinical drop
Analyzing 1,443 non-AI colonoscopies across four Polish centers in the ACCEPT study (795 pre-AI and 648 post-AI procedures, Sept 2021 to Mar 2022), Krzysztof Budzyn and colleagues found endoscopists' unaided adenoma detection rate dropped from 28.4% to 22.4% (-6.0 percentage points; 95% CI -10.5 to -1.6; multivariable odds ratio 0.69: Warwick WRAP; STAT News). Design limit: Retrospective observational comparison rather than a randomized deskilling trial; patient mix or fatigue over time can contribute, and visual polyp scanning differs from analytical office writing.
2. Becker et al. / METR (arXiv:2507.09089, Jul 2025): 19% slower while feeling 20% faster
Model Evaluation and Threat Research (METR) ran a randomized controlled trial with n = 16 experienced open-source maintainers completing 246 real tasks in mature repositories averaging ten years old and over 1.1 million lines of code (arXiv:2507.09089; Simon Willison). Allowed frontier AI coding tools (mainly Cursor Pro with Claude 3.5/3.7 Sonnet), developers posted an acceptance rate under 44% across AI examples and took 19% longer in that 246-task study than when coding unaided. Before starting, they forecast a 24% speedup in completion rate; after finishing 19% slower in the study, they still believed AI made them 20% faster. Design limit: Small expert sample (n = 16) in massive familiar repos; results do not prove AI slows beginners or greenfield scripts, though the 39-point perception gap proves gut feelings misjudge time saved.
3. Kosmyna et al. / MIT Media Lab (arXiv:2506.08872, Jun 2025): small EEG preprint
Nataliya Kosmyna and colleagues tested n = 54 Boston-area student participants writing 20-minute SAT essays across LLM, search-engine, and brain-only groups, with only n = 18 students in a fourth crossover session (arXiv:2506.08872; The Decoder). Over 80% of Session 1 essays by LLM students failed accurate quote recall, and electroencephalography (EEG) showed weaker task connectivity than unaided writers. Explicit scientific caveat: arXiv:2506.08872 is a non-peer-reviewed preprint on a small student sample (n = 54, n = 18 crossover). Lower EEG connectivity during an offloaded 20-minute school essay reflects lower momentary effort, not permanent brain damage or "brain rot."
| Study & Date | Sample & Design | Measured Result | Viral Distortion | Scientific Boundary |
|---|---|---|---|---|
| Budzyn et al. (Aug 2025) | 1,443 non-AI cases at 4 Polish centers. | Unaided detection rate fell from 28.4% to 22.4% (-6.0 pts). | "AI ruins doctors' eyesight." | Observational pre/post; visual scan reflex, not office prose. |
| METR / Becker et al. (Jul 2025) | n = 16 OSS devs across 246 tasks; RCT. | 19% slower study rate (<44% prompt rate accepted); felt 20% faster. | "AI slows all coding." | Small n = 16 expert sample in 1.1M-line repos. |
| Kosmyna et al. / MIT (Jun 2025) | n = 54 students (18 crossover); EEG essays. | >80% of LLM students missed quotes; lower EEG connectivity. | "ChatGPT causes brain rot." | Non-peer-reviewed preprint, small n = 54; measures task effort only. |
| Microsoft & CMU (Feb 2025) | n = 319 workers (936 examples); survey. | Higher AI trust links to lower critical-thinking effort. | "Staff stop thinking." | Self-report survey; six of seven authors at Microsoft. |
Read soberly, all four studies point to one practical rule: keep deliberate manual reps in your week and measure speed with a clock rather than vibes.
Why does the junior apprenticeship ladder break before senior output drops?
Unchecked first-draft automation taxes teams unevenly: senior staff burn quiet hours catching subtle AI errors today, while apprentices and new hires miss the manual reps that turn juniors into senior reviewers tomorrow.
METR's n = 16 maintainers avoided shipping bad patches across 246 tasks by rejecting over 56% of AI code examples, drawing on ten years of manual debugging (arXiv:2507.09089). First-year paralegals, junior accountants, and new ops analysts lack that ten-year mental filter. When juniors skip manual first drafts, they pass three times as much unverified volume upstream (pair handoffs with Before You Send AI-Assisted Work and our Team Workslop Agreement Script) and stall whenever a vendor updates model behavior overnight (see our 10-Prompt Canary Set).
US Department of Labor TEN 07-25 (February 13, 2026) addresses this directly under Delivery Principle 3: build complementary human skills alongside AI use (DOL TEN 07-25). At No Worker Left Behind (Boston 501(c)(3), EIN 88-3268312), ed.cafe Boston Circle #1 at ed.cafe/start/ pairs Massachusetts's free Google AI certificate with human verification labs across 30 free chairs; note openly that Circle #1 is pre-launch and forming as of October 2026, closes December 1, 2026, and has zero completed cohorts yet.
What do weekly cold reps cost a 12-person team compared with perception-gap waste?
One 45-minute unaided "Cold Rep" per week across a 12-person team adds roughly $168 per month in net drafting time while catching negative-ROI AI habits that quietly burn over $1,600 monthly.
Assume a 12-person team (base pay plus a 25% benefits rate per worker): 9 specialists at $48.00/hour loaded, 2 apprentices at $32.00/hour loaded, and 1 lead at $64.00/hour loaded (blended rate $46.67/hour). Because a 45-minute Weekly Cold Rep completes real Tuesday work unaided, even a 10% manual speed delta (4.5 mins/week per worker = 3.6 team hours/month) costs just $168.01/month, plus $140.01/month for a 15-minute monthly stopwatch check ($308.02/month total).
Compare $308/month against two measured drags: 6 specialists losing METR's 19% slowdown rate across 5 legacy hours weekly (22.8 hrs/mo = $1,094.40/mo), plus 6 monthly "workslop" handoffs at BetterUp and Stanford's survey average of 111 minutes each (n = 1,004 / 1,150 US desk workers, $186/employee/month survey-based; 6 x 111m = 11.1 hrs = ~$518/mo: BetterUp, Sep 2025).
| Monthly Item (12 Staff) | Unchecked AI Default | ed.cafe Weekly Cold Rep Protocol |
|---|---|---|
| Unaided Cold Reps | 0 hrs (skills rust silently). | 36 hrs real output (~$168 speed delta). |
| Stopwatch Check | 0 hrs (trusts gut feelings). | 3 hrs/mo ($140 loaded; clocks real speed). |
| METR Legacy Drag | 22.8 hrs lost ($1,094/mo at 19% study rate). | Caught in Week 2; AI turned off where slow. |
| Workslop Rework (6x) | 11.1 hrs lost ($518/mo at 111m each). | Near zero (skeleton-first rule keeps ownership). |
For deeper vendor benchmarking, add our 50-Example Spreadsheet Method.
Copy-ready protocol: The ed.cafe Weekly Cold Rep Anti-Deskilling Protocol
Pin The ed.cafe Weekly Cold Rep Anti-Deskilling Protocol for Teams & Apprentices in your team channel. It needs zero budget and runs inside normal work.
THE ED.CAFE WEEKLY COLD REP ANTI-DESKILLING PROTOCOL FOR TEAMS & APPRENTICES
Grounded in Budzyn et al. (2025), METR RCT (2025) & DOL TEN 07-25 Principle 3 | ed.cafe
Cadence: Ongoing Weekly Habit | Added Non-Work Overhead: <20 Mins/Month per Person
RULE 1: THE SKELETON-FIRST GATE (EVERY DAY, 3 MINUTES BEFORE ANY PROMPT)
Never open an AI chat window on a blank screen. Before prompting on any memo, code
patch, case note, or analysis, write 3 to 5 raw lines in your own words first:
1. Core thesis, diagnosis, or technical approach I expect: _________________________
2. Two primary facts, edge cases, or constraints that MUST survive: ________________
3. One failure mode or missing number I must verify manually: ______________________
Why it works: Preserves generation-effect encoding (counters MIT Session-1 recall drop).
RULE 2: ONE 45-MINUTE "COLD REP" EVERY TUESDAY OR FRIDAY (ALL STAFF, SENIORS INCLUDED)
[ ] Block 45 minutes once per week on a real, non-trivial job task with ALL AI OFF:
- Close ChatGPT, Copilot, Claude, Gemini, Cursor autocomplete, and AI summaries.
- Allowed: Primary docs, statutory text, man pages, calculators, coworker check.
- Complete the draft, query, patch, or chart review end to end by hand.
[ ] Log 1 line after finishing:
Date: ______ | Task: ______________________ | Rusty spot noticed: _______________
Why it works: Maintains unaided detection sharpness (counters Budzyn 28.4%->22.4% drop).
RULE 3: THE APPRENTICE & NEW-SKILL 50/50 FLOOR (FIRST 60 DAYS ON ANY DOMAIN)
For interns, apprentices, Year-1 hires, or seniors learning an unfamiliar tool/law:
[ ] Keep at least 50% of weekly tasks completely unaided until passing 3 consecutive
manual reviews without structural errors.
[ ] On the AI-assisted 50%, require a 60-second spoken walkthrough explaining WHY key
lines or claims work without reading off the screen. Rewrite any unexplained line.
RULE 4: THE BIWEEKLY METR STOPWATCH CHECK (COUNTERING THE 39-POINT SPEED ILLUSION)
Before mandating AI on a recurring workflow, clock 6 real tickets over 2 weeks
(3 tickets AI-ON including all prompt/debug/verify minutes vs 3 matched tickets AI-OFF):
- Perceived speed estimate before checking clock: [ ] Faster [ ] Same [ ] Slower
- Median total minutes AI-ON (prompt + wait + verify + fix): ______ mins
- Median total minutes AI-OFF (manual start to verified finish): ______ mins
- Defect count caught by peer reviewer: AI-ON ______ vs AI-OFF ______
DECISION GATE: If AI-ON saves <15% real clock time OR raises peer-caught defects,
move that task to your "Manual First / AI Off" list and re-test in 90 days.Where this stops being true
Four boundaries separate skill-retention habits from accessibility, clinical, and legal rules:
- None of the three 2025 studies supports banning assistive AI: Budzyn et al. (
1,443cases) is observational, METR (n = 16developers) studied veteran maintainers in million-line repos, and Kosmyna et al. (n = 54students) is a non-peer-reviewed preprint. Scoped AI assistance measurably helps bounded formatting, translation, and accessibility tasks when verified. - Disability accommodations override blanket "AI-off" rules: Dictation, reading, or structuring tools used for dyslexia, motor disabilities, visual impairments, or neurodivergent executive-function support may be protected accommodations under the Americans with Disabilities Act; see AI Mandates & Workplace Accommodations.
- Licensed clinical, engineering, and legal rules govern sign-off: This guide provides general operational information, not medical, engineering, or legal advice.
- Skip cold reps on pure drudgery: Spend your 45-minute unaided rep on core reasoning (SQL logic, diagnosis, issue-spotting), not retyping zip codes.
Questions people ask
Did the MIT ChatGPT study really prove that AI causes permanent brain rot?
No. The June 2025 MIT Media Lab paper by Kosmyna et al. (arXiv:2506.08872) is a non-peer-reviewed preprint tracking n = 54 students (n = 18 in crossover) writing 20-minute SAT essays. Lower EEG connectivity and weak quote recall (>80% of essays missed in Session 1) show lower mental effort on an offloaded task, not permanent brain damage.
What happened to doctors' accuracy in the 2025 Lancet colonoscopy AI study?
In an August 2025 Lancet Gastroenterology & Hepatology observational study of 1,443 unassisted colonoscopies across four Polish centers, experienced endoscopists' unaided adenoma detection rate fell from 28.4% (795 pre-AI cases) to 22.4% (648 post-AI cases) after routine AI exposure, a 6.0-point drop showing visual vigilance softens without unaided reps.
Why did experienced programmers get 19% slower with AI in the METR trial?
In METR's July 2025 randomized trial (arXiv:2507.09089) of n = 16 veteran open-source maintainers across 246 real tasks in 1.1-million-line codebases, developers spent extra minutes prompting, reviewing, and fixing subtle repo errors while posting an acceptance rate under 44% across AI examples, yet mistakenly felt 20% faster.
How often should workers practice without AI to keep core job skills sharp?
One scheduled 45-minute unaided "Cold Rep" per week on a core analytical, coding, or diagnostic task, paired with writing a three-line outline before opening any daily prompt, keeps independent recall sharp at minimal cost. Apprentices and new hires in their first 60 days should complete at least 50% of tasks unaided.
How can a manager tell if AI is actually saving time or slowing the team down?
Stop relying on self-reported survey feelings, which missed clock reality by 39 percentage points in METR's developer study. Time six matched real tickets over two weeks (three with AI including all verification and fix minutes versus three unaided) and turn AI off wherever net end-to-end clock savings fall below a 15% threshold.
Should junior employees use AI less than senior employees during onboarding?
Yes, on core judgment tasks. Senior staff already hold the error filter built from years of manual troubleshooting, whereas juniors who skip manual first drafts struggle to audit machine mistakes. Keep a 50% unaided threshold rule during a learner's first 60 days plus 60-second spoken walkthroughs of AI-assisted work.
Key terms
- Adenoma detection rate (ADR): The share of screening colonoscopies in which a clinician spots at least one precancerous polyp, a core quality benchmark.
- Automation vigilance decay: The human-factors drop in manual attention and error-catching speed after sustained reliance on mostly accurate automated alerts.
- Cold rep: A scheduled 45-minute block completing a real core job task with all generative AI tools turned off to preserve unaided skill.
- Generation effect: The recall and error-spotting advantage built by actively constructing an outline or solution yourself rather than reading generated text.
- METR perception gap: The 39-point split in Becker et al. (
n = 16developers,246tasks) between feeling 20% faster with AI and clocking 19% slower. - Skeleton-first rule: Writing a three-to-five-line human thesis, constraint list, and verification target before opening any AI prompt window.
Sources
- K. Budzyn et al., "Endoscopist Deskilling After Exposure to AI in Colonoscopy" (Lancet Gastro Hep, Aug 2025, n = 1,443): Warwick WRAP
- STAT News, "AI Deskilling in Doctors: Colonoscopy Study" (Aug 12, 2025): STAT News
- J. Becker et al. (METR), "Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (Jul 2025, n = 16, 246 tasks): arXiv:2507.09089
- Simon Willison, "Notes on METR AI Productivity RCT" (Jul 12, 2025): Simon Willison
- N. Kosmyna et al. (MIT Media Lab), "Your Brain on ChatGPT" (Jun 2025 preprint, n = 54): arXiv:2506.08872
- The Decoder, "MIT Preprint Shows Cognitive Debt Through ChatGPT" (Jun 2025): The Decoder
- The Register on Microsoft/CMU, "Confidence in AI & Critical Thinking" (Feb 11, 2025, n = 319): The Register
- L. Bainbridge, "Ironies of Automation" (Automatica, 1983): DOI90046-8)
- EY, "Work Reimagined Survey" (Dec 8, 2025, n = 15,000): EY
- KPMG & Univ. of Melbourne, "Trust in AI Global Study" (Apr 2025, n > 48,000): KPMG
- US DOL ETA, "TEN 07-25: AI Literacy Framework" (Feb 13, 2026): DOL
- BetterUp & Stanford, "Hidden Costs of Workslop" (Sep 2025, n = 1,004 / 1,150): BetterUp
