Research

Randomized controlled trial with 16 experienced open-source developers completing 246 real-world coding tasks. Screen recordings analyzed. Primary tools: Cursor Pro + Claude 3.5/3.7 Sonnet.

Developers predicted AI would make them 24% faster. They were actually 19% slower — a 43-point perception gap.

"Surprisingly, we find that allowing AI actually increases completion time by 19% — AI tooling slowed developers down."

Annual developer survey. Trust in AI coding accuracy dropped from 43% to 33%. Adoption rose from 76% to 84% over the same period.

People use tools they don't trust, perceiving benefits they don't actually get. Suggests a gap between expectation and reality.

Survey of 791 developers. 30% of seniors edit AI output enough to offset time savings (vs 17% of juniors). Despite lower trust, seniors ship 2.5x more AI-generated code to production.

Experience enables judgment about when AI output is trustworthy and when it needs work.

Survey of 319 knowledge workers sharing 936 examples of GenAI use. Strong negative correlation between confidence in AI capabilities and critical thinking behavior.

The more people trust AI to handle tasks, the less they question its output.

Experimental study using EEG to measure brain activity during memory formation. Participants showed significantly reduced memory encoding for content created with AI assistance versus content written themselves.

The AI assistance prevented memory formation at a neurological level — they never learned the content in the first place.

Large-scale RCT with nearly 1,000 Turkish high school students across four 90-minute sessions. Students with unrestricted ChatGPT access scored 17% worse on unassisted final exams compared to control group.

Students with scaffolded AI (guardrails and structure) showed no significant difference from control. How you use it determines whether it helps or hurts.

User studies testing whether explanations help humans evaluate AI recommendations. Explanations increased acceptance of AI suggestions regardless of whether they were correct.

When AI was right, performance improved slightly. When AI was wrong, performance degraded. Explanations created trust without calibration.

Crossover randomized controlled trial with endoscopists. After 3 months of AI-assisted colonoscopy, physicians' adenoma detection rate dropped from 28.4% to 22.4% when the AI was removed.

The skill atrophied measurably. You don't maintain skills you stop exercising.

Analysis of 211 million lines of code committed between 2020-2024. Copy-paste duplication within codebases increased 8-fold. Refactoring work dropped from 25% to 10% of changes.

Suggests developers accepting AI output without understanding or adapting it to existing patterns.

Literature review + 14 interviews + two experiments with 654 professionals (financial services and HR).

Key findings:

• Transparency (show reasoning) — strong positive effect on outcomes
• Process control (let user steer how) — strongest positive effect
• Outcome control (let user steer what) — strong positive effect
• Engagement features (system prompts for feedback) — significant negative effect for experienced users

"Engagement, for example, through asking for user feedback, is perceived by employees as effortful rather than empowering."

The insight: Show your reasoning, give them control, don't interrupt with engagement theater.