AI kognitiv avlastning: Produktivitetsvinster kostar expertise – “intuition rust” bevisad

AI kognitiv avlastning skapar augmentation trap visar MIT Sloan dynamisk modell + longitudinal evidence: AI raises productivity short-term MEN sustained use erodes skill on which gains depend. Cancer specialists 1-year study (Ehsan et al. 2026): initial productivity gains kom med gradual dulling expert judgment – “intuition rust”. ChatGPT learners retain significantly less material 45-day follow-up (Barcaui 2025). Programming experiments: participants delegating tasks learned least (Shen & Tamkin 2026). För dig som projektledare betyder detta: Even fully informed adoption rational när front-loaded gains outweigh long-run costs = steady-state loss (worker ends less productive than before). Managers’ short-termism (δ_firm > δ_worker) turns loss into TRAP: worker worse off än if AI never adopted. Model identifies five regimes (non-adoption, augmentation worse/better, automation worse/better) separating beneficial från harmful deployment. Cognitive offloading displaces practice through which expertise develops → permanent skill divergence.

Evidence base: Skill erosion NOT theoretical – empirically validated

Longitudinal studies visa gradual decline:

Cancer specialists (1-year AI decision support):

Initial productivity gains confirmed
Over time: gradual dulling av expert judgment
Term coined: “intuition rust”
Experts begin missing errors de förut caught easily
Source: Ehsan et al. 2026

ChatGPT learning study:

Students using ChatGPT för learning
45-day follow-up: Significantly LESS material retained
Compared to no-AI control group
Cognitive offloading = degraded long-term retention
Source: Barcaui 2025

Programming tasks + neuroimaging:

Degraded performance efter sustained AI use
Lab experiments, programming tasks, brain imaging ALL converge
Source: Lee et al. 2025, Patra et al. 2025

Coding delegation experiment (Shen & Tamkin 2026):

Participants delegating tasks: learned LEAST
Participants staying cognitively engaged: fared BETTER (men still below no-AI)
Critical insight: Offloading drives skill loss EVEN when goal is learning
Production settings (där incentive preserve skill weaker) → unlikely fare better

Experienced developers vs. novices (Sarkar 2026):

Experienced: produce more aligned outputs, accept suggestions higher rates (agents)
Gradient REVERSES för autocompletion: less experienced accept MORE
Seasoned expertise = distinguish good från merely plausible answers
Men expertise built through continuous practice coding/debugging
Deadline pressure → rational rely på passable AI → expert gradually stops exercising skills

För projektledare: This is NOT speculation. Multiple independent studies across domains (medicine, education, programming, neuroimaging) converge på samma pattern. Cognitive offloading erodes expertise gradually but persistently.

Two productivity channels: α (skill-neutral) + β (scales with expertise)

Model decomposes AI productivity into:

Channel 1: α (skill-neutral component) Raw AI output independent av who’s using it

Example: Translation tool – novice + veteran benefit equally
Template-based report drafting – senior partner extracts only marginally more value än first-year

Channel 2: β (knowledge-complementary component) Scales with worker’s judgment

Example: Client strategy work – model alone provides little, veteran consultant extracts significant insight
Coding: Expert evaluates AI-generated code spot mistakes, anticipate technical debt, reject poor suggestions

Three regimes:

β > 1 (Skill complement):

Productivity gain from AI more than compensates displaced human contribution
Higher-skill workers benefit MORE från tool
Usage INCREASES with skill
Feedback: Self-correcting (high-skill uses heavy → loses skill → reduces usage → recovers)

β = 1 (Skill neutral):

AI provides same net benefit regardless expertise
Usage FLAT across skill levels
Translation, basic automation

β < 1 (Skill substitute):

AI partially substitutes för skill
Narrows gap between high/low-skill workers
Lower-skill workers gain MORE at margin → adopt HEAVIER
Usage DECREASES with skill
Feedback: Self-reinforcing (low-skill uses heavy → loses skill → increases usage → deskills faster)

För projektledare: Same language model produces different effective α, β depending på workflow embedding. Design determines whether tool complements eller substitutes expertise.

Steady-state loss: Rational adoption, permanent degradation

Even fully informed decision-maker adopts AI when: Front-loaded productivity gains outweigh discounted long-run skill costs

Result: Worker ends up LESS productive än before adoption at steady state

Three adoption thresholds (skill-neutral case β=1):

α₀ (adoption onset): Below: AI never adopted Above: Productivity boost justifies usage

α₁ (break-even threshold): Between α₀ and α₁: STEADY-STATE LOSS REGION

Adoption raises current productivity ✓
Long-run value < no-AI benchmark ✗
Privately rational för decision-maker
Worker permanently worse off

Above α₁: Adoption improves BOTH short-run + long-run

Productivity gain outweighs skill cost

För projektledare: Loss region EXPANDS monotonically med discount rate δ. Mer impatient decision-maker → wider set parameters där adoption privately rational men long-run harmful.

Augmentation trap: När loss becomes welfare problem

Steady-state loss = informed tradeoff (worker choosing own usage)

Augmentation trap = moral hazard when: Decision-maker + worker misaligned

Two misalignment forms:

1. Managerial short-termism (δ_firm > δ_worker):

Manager evaluated på quarterly output
Worker investing i long career
Firm discounts future heavier → places LESS value preserving skill
Chooses HIGHER AI usage at every skill level

Concrete example (illustrative parameters):

Manager 3-year effective tenure (δ_F=0.33)
Worker 10-year career plan (δ_W=0.10)
Manager sets usage: û=0.26
Worker would choose: û=0.14
Manager’s policy NEARLY TWICE usage
Worker’s steady-state skill: 14% LOWER än self-chosen (Ŝ=0.75 vs 0.86)

2. Worker skill externality (ω): Workers value skill för reasons firm ignores:

Side projects
Intellectual communities
Ability understand independently
Long-term career mobility

Firm’s objective OMITS these returns → externality

För projektledare: Privately rational managerial decisions systematically overuse AI relative worker’s long-term interest. Misalignment NOT accidental – structural feature when horizons differ.

Five deployment regimes: Beneficial vs harmful adoption

Region I (Non-adoption):

α too small justify adoption
Examples: Plumbing, complex negotiations
AI adds too little value

Region II (Augmentation, worse-off):

STEADY-STATE LOSS
Adoption rational short-run
Long-run skill cost exceeds gain
Examples: Entry-level financial analysis med LLM drafting
Worker uses AI, productivity initially up, skill erodes, ends worse off

Region III (Automation, worse-off):

Full automation (u=1)
Skill erodes to zero
Raw AI output < worker potential
Examples: Customer service scripting
Worker replaced by insufficient AI

Region IV (Augmentation, better-off):

Productivity gain outweighs costs
Adoption improves long-run position
Examples: Experienced doctors med diagnostic AI
High β (skill complement) → skill retains value under AI usage

Region V (Automation, better-off):

Full automation (u=1)
Raw AI output > worker potential (α – γ > Ŝ)
Examples: Data entry, templated correspondence
Automation legitimately superior

För projektledare: Job’s region depends på HOW tool embedded i workflow, NOT tool itself. Same ChatGPT deployment can fall Region II (harmful) eller Region IV (beneficial) beroende på α, β parameters determined by usage practice.

Permanent skill divergence: β<1 triggers bifurcation

When AI productivity depends LESS på worker expertise (β<1):

Positive feedback loop:

Low-skill worker uses AI heavily
Loses skill
Usage increases further (because AI substitutes skill)
Deskills faster
Eventually: Skill → 0

Negative feedback loop:

High-skill worker avoids AI
Builds skill
Usage decreases further
Skill grows faster
Eventually: Achieves full potential Ŝ

Result: PERMANENT DIVERGENCE

Experienced workers realize full potential
Novices deskill to zero
Small differences i managerial incentives determine which path worker takes

Example scenario: Two junior analysts, identical potential, slightly different managers

Manager A: 5% more short-term focused
Analyst under Manager A: deskills to zero over 2 years
Analyst under Manager B: achieves 85% potential

För projektledare: β<1 deployments EXTREMELY DANGEROUS för workforce development. Initial small differences (manager tenure, quarterly pressure, promotion timing) → massive long-term divergence. Organizations risk creating two-tier workforce: veterans WHO started before AI vs. novices WHO never developed expertise.

Kursinfo: AI för projektledare

Practical protection mechanisms

1. Monitor usage practice, not just tool: Track effective α, β parameters

High α, low β = danger zone (skill substitute)
Low α, high β = safer (skill complement) Don’t ask “do we use ChatGPT?” Ask “HOW do teams interact med it?”

2. Design för β>1 (skill complement): Keep human in loop där judgment shapes AI quality Examples:

Code review (not blind acceptance)
AI drafts, human edits (not copy-paste)
Collaborative workflows (not full delegation)

3. Align incentives across horizons: Manager evaluation metrics include long-term skill preservation

NOT just quarterly productivity
Track team capability trajectory
Penalize deskilling patterns

4. Protect learning opportunities: Mandate AI-free practice zones Junior developers: 40% tasks WITHOUT AI (build foundational “muscles”) Experienced workers: Periodic AI detox (maintain sharp evaluation skills)

5. Measure skill directly, repeatedly: Don’t assume productivity = capability Track:

Error detection rates (cancer specialists catching mistakes)
Retention tests (learning material without AI crutch)
Performance degradation when AI unavailable

6. Recognize externality: Workers value skill beyond firm’s immediate productivity

Career mobility
Professional identity
Intellectual autonomy Firm-optimal ≠ worker-optimal → need governance

7. Regime classification før deployment: Before rolling out AI tool, estimate α, β för intended usage

Region II, III → redesign workflow eller don’t deploy
Region IV, V → proceed with monitoring Use pilot studies measure actual skill trajectories

Bottom line

AI kognitiv avlastning skapar augmentation trap: productivity vinster short-term cost expertise long-term. Evidence robust (cancer specialists “intuition rust” 1-year, ChatGPT learners 45-day retention loss, programming deskilling). Model decomposes productivity α (skill-neutral) + β (expertise-scaling). β>1 complements skill (safer), β<1 substitutes (dangerous divergence). Steady-state loss: Even informed adoption rational när front-loaded gains outweigh costs → worker ends less productive. Augmentation trap: Managerial short-termism (δ_firm > δ_worker) + skill externality turn loss into welfare problem. Example: 3-year manager sets usage 2x higher än 10-year worker would choose → 14% lower steady-state skill. Five regimes separate beneficial (IV, V) från harmful (II, III) adoption. Permanent divergence när β<1: Small managerial differences → some workers achieve potential, others deskill zero. Protection: Monitor α, β not tool, design för complement (β>1), align horizons, protect learning zones, measure skill directly, recognize externality, classify regime før deployment.

Källa: “The Augmentation Trap: AI Productivity and the Cost of Cognitive Offloading” av Michael Cao & Sinan Aral, MIT Sloan School of Management, publicerad 10 april 2026.

Tillbaka till artiklarna