Why AI makes deep work slower, not faster

Executive overview

A 2025 study by METR found experienced programmers were 20% slower on tasks when using AI tools — the opposite of what everyone predicted. The culprit is cybernetic collaboration: offloading cognitive load to AI reduces the intensity of focus, and focus intensity is precisely what drives output quality in knowledge work.

Deep work scales with focus intensity multiplied by duration. Anything that breaks or softens that focus — even something pleasant — degrades results. AI-assisted coding creates a comfortable back-and-forth rhythm that feels productive but runs the brain at a lower gear.

The brain focusing hard is an incredibly powerful tool. Be wary of anything that reduces that intensity unless you can truly outsource the work entirely.

The METR study: what happened

• METR (nonprofit AI evaluator) recruited 16 experienced open-source developers
• Tasks were randomly assigned: AI allowed or disallowed per issue, averaging ~2 hours each
• AI tools used: Cursor Pro with Claude 3.5/3.7 Sonnet — frontier models at the time
• Economic experts predicted ~40% speed-up; developers themselves predicted 20–30%
• Observed result: AI-assisted tasks took ~20% longer than unassisted tasks
• Developers rated the AI-assisted experience as more pleasant despite being slower

Why cybernetic collaboration backfires

• Cybernetic collaboration: splitting cognitive effort between human and AI in an interactive loop
• The loop — prompt, review, prompt again — creates frequent low-effort breaks
• Breaks reduce focus intensity; lower intensity means slower, lower-quality output
• Programmers spent less time actively coding and more time prompting, reviewing, and waiting
• Some idle time appeared in screen recordings — no activity at all
• Reviewing and correcting AI output often took longer than writing the code from scratch
• The machine can't yet take over deep work entirely, so the human still has to drive — just at a lower gear

The whiteboard effect vs cybernetic collaboration

• Productive human collaboration works by increasing focus intensity, not reducing it
• At a whiteboard with peers, social pressure prevents attention from wandering
• When a collaborator explains an insight, you must lock in to follow — focus deepens
• Formula: intensity of focus × duration = output for cognitively demanding work
• The whiteboard effect uses other people to squeeze more focus out of you
• Cybernetic collaboration does the opposite — it gives you more breaks and reduces strain
• Deep work doesn't benefit from pleasant; it benefits from hard

Takeaway and limits of the finding

• Adding anything to a workflow that reduces focus intensity will likely reduce productivity
• AI is better suited to automating shallow tasks or accelerating non-deep-work activities
• Speeding up information lookup, for example, frees more time for actual deep work — that's compatible
• Using AI to generate project notes as a shutdown ritual is lower-risk than ongoing interactive loops
• Cybernetic collaboration for non-trivial creative or analytical work is likely fool's gold
• Smaller, specialized AI systems (not frontier models) may be where real productivity gains emerge

Q&A highlights

• ChatGPT as project diary: useful for shutdown markers and restart notes, but risks turning work into an ongoing cybernetic loop — the same trap as the METR programmers
• Active recall spacing: revisit material within ~2 weeks if not applying it; using knowledge is the strongest form of retention
• AI environmental footprint: frontier models are computationally expensive, but not economically sustainable at scale; future AI will likely run as smaller, specialized systems on local hardware with a footprint similar to current computing
• Lifestyle-centric planning (Sevin's case study): fixating on one radical change (work in nature) improved one dimension while worsening others (commute, family time, income); reskilling in web development, achieving a 2x salary increase, and planning Fridays off for nature produced a better overall life — without needing to work there

Green Bank investigative segment

• Green Bank, WV has no Wi-Fi due to a nearby radio telescope; its school posted the lowest scores in Pocahontas County
• Washington Post op-ed attributed low scores to lack of internet-connected Chromebooks
• County-level NEAP time-series data: math scores rose through ~2017, then declined — compatible with the Wi-Fi hypothesis but not conclusive
• Comparable West Virginia counties (similar demographics, Wi-Fi available) saw larger score drops over the same period and smaller post-pandemic recovery
• County-level data does not support the claim that Wi-Fi absence caused the underperformance; differences may reflect pre-existing disadvantages or gifted-program concentration in the comparison school
• Broader lesson: intuitive data points are easy to overextend; school-by-school longitudinal data would be needed to confirm the claim