AI Works Best When Humans Edit, Not Accept

Writers fear AI tools will atrophy their skills. Forecasters predicted AI-assisted practice would reduce learning compared to unassisted practice.

Method: Writers randomly assigned to practice with AI improved more on a next-day writing test than those who practiced without AI—despite exerting less effort. The AI acted as a coach, not a crutch: participants who practiced with AI showed greater skill gains than those who received personalized feedback from experienced human editors. The mechanism inverts the conventional wisdom that effort equals learning.

Caveats: One-day retention window. Long-term skill durability and transfer to novel contexts remain untested.

Reflections: Does the coaching effect persist beyond one day, or does skill decay accelerate without continued AI access? · Which task characteristics determine when AI acts as coach versus crutch? · Can this framework extend to motor skills or only cognitive tasks?

Automated web crawls miss fingerprinting scripts that activate only during real user interactions like clicks, scrolls, or form submissions.

Method: A 10-week study with 30 participants captured real browsing behavior and revealed that 36.5% of fingerprinting scripts are triggered exclusively by user interactions—invisible to automated crawlers. The study logged actual mouse movements, keyboard events, and page interactions, exposing a blind spot in current prevalence estimates. Canvas fingerprinting and WebGL queries often fire only after users engage with page elements.

Caveats: 30-participant sample from a single demographic. Fingerprinting prevalence varies by geography and browsing habits.

Reflections: Which interaction types (clicks vs. scrolls vs. form fills) trigger the most fingerprinting scripts? · Do privacy-conscious users exhibit different fingerprinting exposure due to browsing patterns? · Can browser vendors detect and block interaction-triggered fingerprinting without breaking legitimate analytics?

Formal email replies demand cognitive load for tone and completeness. LLM tools require detailed prompts, shifting effort rather than reducing it.

Method: Instead of prompting, users answer AI-generated questions extracted from the incoming email. The system parses the sender's requests, generates targeted questions (e.g., 'What date works for the meeting?'), and assembles answers into a polished reply. Users reported lower cognitive load and faster completion compared to traditional prompting or manual drafting. The QA interface decomposes the reply task into structured micro-decisions.

Caveats: Effectiveness depends on email complexity. Multi-topic or ambiguous emails may generate too many or irrelevant questions.

Reflections: Can the QA approach extend to other formal writing tasks like cover letters or proposals? · How do users handle cases where the AI misinterprets the sender's intent and generates wrong questions? · Does repeated use train users to anticipate questions, reducing cognitive offloading benefits?