Hidden_Cognitive_Costs_-_GIDS.pdf

Transcript

1. The Hidden Cognitive Costs of AI-assisted Programming Ragunath Jawahar, Founder Legacy Code Labs • legacycode.com
2. 2019 2017
3. 01 Mechanized Comprehension Understand software systems through tooling that extracts structure, history, meaning, and intent from code at scale. 02 Mechanized Veri fi cation Verify software changes deterministically as they happen. 03 Mechanized Remediation Identify and resolve structural decay and technical debt through tool-assisted intervention over expensive rewrites. 04 Directed Evolution Deliberately evolve software toward new behavior through continuous, traceable, and guided transformation.
4. 01 Mechanized Comprehension Understand software systems through tooling that extracts structure, history, meaning, and intent from code at scale. 02 Mechanized Veri fi cation Verify software changes deterministically as they happen. 03 Mechanized Remediation Identify and resolve structural decay and technical debt through tool-assisted intervention over expensive rewrites. 04 Directed Evolution Deliberately evolve software toward new behavior through continuous, traceable, and guided transformation.
5. 01 Mechanized Comprehension Understand software systems through tooling that extracts structure, history, meaning, and intent from code at scale. 02 Mechanized Veri fi cation Verify software changes deterministically as they happen. 03 Mechanized Remediation Identify and resolve structural decay and technical debt through tool-assisted intervention over expensive rewrites. 04 Directed Evolution Deliberately evolve software toward new behavior through continuous, traceable, and guided transformation.
6. 01 Mechanized Comprehension Understand software systems through tooling that extracts structure, history, meaning, and intent from code at scale. 02 Mechanized Veri fi cation Verify software changes deterministically as they happen. 03 Mechanized Remediation Identify and resolve structural decay and technical debt through tool-assisted intervention over expensive rewrites. 04 Directed Evolution Deliberately evolve software toward new behavior through continuous, traceable, and guided transformation.
7. 01 Mechanized Comprehension Understand software systems through tooling that extracts structure, history, meaning, and intent from code at scale. 02 Mechanized Veri fi cation Verify software changes deterministically as they happen. 03 Mechanized Remediation Identify and resolve structural decay and technical debt through tool-assisted intervention over expensive rewrites. 04 Directed Evolution Deliberately evolve software toward new behavior through continuous, traceable, and guided transformation.
8. My AI adoption journey. Three years. Six checkpoints. H I D D E N C O G N I T I V E C O S T S 02 / 43 S E C T I O N · 1 O F 6
9. April 2023 Three years ago T O O L GitHub Copilot S P E N D $10 / month W O R K F L O W TDD with autocomplete M I N D S E T AI as a threat to evaluate "Turned it off when the suggestions got annoying." J O U R N E Y · 1 O F 6 H I D D E N C O G N I T I V E C O S T S 03 / 43
10. October 2024 Eighteen months ago T O O L Claude / ChatGPT S P E N D $20 / month W O R K F L O W Browser ↔ IDE copy-paste M I N D S E T Delegate solved pieces, keep the thinking “Rethink TDD. Comfortable delegating entire classes." J O U R N E Y · 2 O F 6 H I D D E N C O G N I T I V E C O S T S 04 / 43
11. November 2024 Seventeen months ago T O O L Cursor and (Claude / ChatGPT) S P E N D $40 / month W O R K F L O W Cursor with Composer mode for large chunks of work, IntelliJ for reading and targeted edits M I N D S E T Supervise every delegation, carve focus hours 1,000–3,000 LOC / day · 4–5 hours "Leverage changes in kind, not degree." J O U R N E Y · 3 O F 6 H I D D E N C O G N I T I V E C O S T S 05 / 43
12. January 2026 Three months ago T O O L Claude Code S P E N D $20 / month W O R K F L O W Structure & behavior review, not line review M I N D S E T AI owns well-scoped work 31K LOC · Go "A language I don't know." J O U R N E Y · 4 O F 6 H I D D E N C O G N I T I V E C O S T S 06 / 43
13. February 2026 Two months ago T O O L Claude Code + Codex App S P E N D $200 / month W O R K F L O W Stack chosen by ecosystem fit M I N D S E T Frontier progress is optional now Go · Rust · TypeScript "Today's models are already enough." J O U R N E Y · 5 O F 6 H I D D E N C O G N I T I V E C O S T S 07 / 43
14. March 2026 One month ago T O O L Claude Code + Codex App S P E N D $297 / month W O R K F L O W Long-horizon, self-verifying runs M I N D S E T Structure and safety nets are the ceiling 6 hours unattended Burnout. J O U R N E Y · 6 O F 6 H I D D E N C O G N I T I V E C O S T S 08 / 43
15. R E W A R D C H E M I S T R Y The mechanism underneath. Before we get to what it cost me, a short detour into why. S E C T I O N · 2 O F 6 H I D D E N C O G N I T I V E C O S T S 09 / 43
16. Dopamine is anticipation, not pleasure. The common misconception, corrected. T H E C L A I M Dopamine is the anticipation chemical. Not the pleasure chemical. P E A K F I R I N G At ~50% reward probability — the sweet spot of a slot machine. "Neurons fire hardest when reward might arrive — not when it does." Fiorillo, Tobler & Schultz, 2003 — Science. P A R T 0 1 · R E W A R D C H E M I S T R Y H I D D E N C O G N I T I V E C O S T S 10 / 43
17. Same chemistry, three machines. Variable ratio reinforcement — the most addictive reinforcement schedule in behavioral psychology. Slot machine · unpredictable payout Social media · unpredictable feed Agentic coding · unpredictable prompt output “Agentic coding and doomscrolling share the same highly addictive pattern. You're pulling the slot machine lever.” Tao Bojlén, 2026 P A R T 0 1 · R E W A R D C H E M I S T R Y H I D D E N C O G N I T I V E C O S T S 11 / 43
18. The loop. Why the cycle hooks. 01 Prompt. (action) 02 Wait a few seconds. (uncertainty window) 03 Outcome arrives — brilliant, mediocre, or broken. (variable reward) 04 Dopamine fires hardest during step 02, not step 03. (neurochemistry) The tighter the loop, the stronger the hook. “Each cycle takes seconds. Each one delivers a tiny hit of novelty.” — Tao Bojlén, via BoingBoing, March 2026. P A R T 0 1 · R E W A R D C H E M I S T R Y H I D D E N C O G N I T I V E C O S T S 12 / 43
19. The treadmill. Why novelty has to keep rising. C O N C E P T Hedonic adaptation. The brain returns to baseline after any reward level. C O N S E Q U E N C E What felt exciting last month feels neutral this month. D R I V E The circuit demands more surface, more stakes, more novelty — to produce the same hit. "You're not chasing more. You're chasing the same feeling from a higher baseline." P A R T 0 1 · R E W A R D C H E M I S T R Y H I D D E N C O G N I T I V E C O S T S 13 / 43
20. My spend curve is the receipt. Two honest readings of the same number. $10 · $20 · $40 · $20 · $200 · $297 3 years. The shape of my own reward circuit. T W O H O N E S T R E A D I N G S Productivity economics: I got better at using the tools. Spend followed leverage. Reward chemistry: Once the mechanism clicked, spend accelerated 15× in two months. Both are true. The talk is about what the second one costs. P A R T 0 1 · R E W A R D C H E M I S T R Y H I D D E N C O G N I T I V E C O S T S 14 / 43
21. April 2023 Three years ago · the loop barely exists C I R C U I T Anticipation onset. Mild, not yet reinforcing. F E L T A low nudge of excitement, mostly drowned out by fighting the tool. S C H E D U L E Too sparse and too unreliable to lock in. Autocomplete missed more than it hit. "The loop was slow and broken. The baseline drifted anyway." D O P A M I N E · 1 O F 6 H I D D E N C O G N I T I V E C O S T S 15 / 43
22. October 2024 Eighteen months ago · the loop arrives, with friction C I R C U I T Reinforcement starts working. Outputs are rewarding enough to anticipate. F E L T Genuinely excited. Delegating entire classes feels like a magic trick. S C H E D U L E Variable-reward, but browser ↔ IDE round trip keeps cycle time high — 30 seconds to a few minutes per prompt. "The outcomes got better. The friction kept the loop loose." D O P A M I N E · 2 O F 6 H I D D E N C O G N I T I V E C O S T S 16 / 43
23. November 2024 Seventeen months ago · the loop tightens C I R C U I T Variable-reward schedule fully functional. Dopamine firing on cadence. F E L T First real throughput (1–3K LOC / day). First real drain — done by 3 PM. S C H E D U L E Cursor drops cycle time to seconds. Single-file edits turn monotonous; Composer adds multi-file surface. "The loop tightened. The reinforcement arrived. The body started keeping score." D O P A M I N E · 3 O F 6 H I D D E N C O G N I T I V E C O S T S 17 / 43
24. January 2026 Three months ago · the loop self-reinforces C I R C U I T Loop + verification theory locking in. Anticipation now intellectual as well as chemical. F E L T 8+ hours a day. Shipping in languages I don't know. Confidence climbing. S C H E D U L E Behavior / graph review cuts per-cycle cost. More cycles per hour. 95%+ AI-written code. "Now the loop runs for eight hours. And I'm not tired yet." D O P A M I N E · 4 O F 6 H I D D E N C O G N I T I V E C O S T S 18 / 43
25. February 2026 Two months ago · the loops multiply C I R C U I T Multiple loops running in parallel — per agent, per project. F E L T Ideas arriving constantly. Tooling blowing open. Schedule starting to slip. S C H E D U L E Multi-agent, multi-project sessions. Attention is now the bottleneck. "The loop is now four loops. Running in parallel." D O P A M I N E · 5 O F 6 H I D D E N C O G N I T I V E C O S T S 19 / 43
26. March 2026 One month ago · the loop runs me C I R C U I T Self-verifying long-horizon runs — loops that don't need me present. F E L T All-nighters. Ignored schedule. Ordinary life starting to feel flat. S C H E D U L E Six-hour unattended runs. ~500K LOC managed across three months. "The loop was running me." D O P A M I N E · 6 O F 6 H I D D E N C O G N I T I V E C O S T S 20 / 43
27. Decision fatigue. A second cost — and it arrives earlier than you think. C L A I M Every decision draws from the same depleting reservoir. S I G N A L Quality of judgment degrades long before subjective fatigue sets in. E V I D E N C E F R O M P R A C T I C E Surgeons schedule complex cases before noon. ERs rotate shifts carefully. "I'll sleep on it" is folk wisdom for a real mechanism. "You don't feel it when it starts. You feel it when the next decision takes too long." P A R T 0 2 · D E C I S I O N F A T I G U E H I D D E N C O G N I T I V E C O S T S 21 / 43
28. Apr 2023 → Oct 2024 Three years to eighteen months ago · below the onset line A P R 2 0 2 3 Copilot autocomplete cut low-level decisions — what test to write next, what to name the variable. Load slightly below baseline. O C T 2 0 2 4 Chat delegation added new decisions (when to delegate, is output right), but volume was low. F E L T No energy drop. Full day's work possible. No 3 PM wall. "The reservoir was full. The bill hadn't arrived." D E C I S I O N F A T I G U E H I D D E N C O G N I T I V E C O S T S 22 / 43
29. November 2024 Seventeen months ago · onset L O A D Two simultaneous streams — reviewing what the agent produced, and generating the next task for it. Both draw from the same reservoir. F E L T Done by 3 PM. Quality of 4 PM decisions noticeably worse than 10 AM decisions. R I T U A L Gym at 5. Home by 6, refreshed. Skip the gym, wrecked by evening. "The gym was a reset, not a workout." D E C I S I O N F A T I G U E H I D D E N C O G N I T I V E C O S T S 23 / 43
30. Jan → Feb → Mar 2026 Three to one months ago · reduced, partially restored, nearly silent J A N 2 0 2 6 Abstraction shift cuts the drain. Graph / behavior review draws far less than line review. F E B 2 0 2 6 Multi-agent restores some load — four agents × two streams = eight decision queues. But each decision stays at abstraction level. M A R 2 0 2 6 Few decisions per session. High per-stake, low per-hour. Reservoir barely touched. "The one cost that actually retreated. The trade shows up later." D E C I S I O N F A T I G U E H I D D E N C O G N I T I V E C O S T S 24 / 43
31. Multitasking and context switching. A third cost — and the one most engineers believe they're immune to. C O N C E P T Context switching — the cognitive cost of moving your attention between tasks. C O N S T R A I N T Working memory holds ~7 items (Miller, 1956); more honestly 3–5 (Cowan, 2001 / 2010). M E C H A N I S M Attention residue — part of your mind stays with the task you just left (Leroy, 2009). S T I C K I N E S S Zeigarnik effect (1927) — unfinished tasks stay active in working memory, demanding closure. "You don't have parallel attention. You have fast thrashing." P A R T 0 3 · C O N T E X T S W I T C H I N G H I D D E N C O G N I T I V E C O S T S 25 / 43
32. Apr 2023 → Jan 2026 Three years to three months ago · below the onset line A P R 2 0 2 3 Inline autocomplete. Single task. No context to switch to. O C T 2 0 2 4 Browser ↔ IDE sequential only. One task at a time. N O V 2 0 2 4 Cursor reduced switching by unifying the surface — counter-intuitive. J A N 2 0 2 6 Single-agent Claude Code. 8+ hours of sustained, single-threaded attention. "Nothing multiplied yet. Attention stayed where I put it." C O N T E X T S W I T C H I N G H I D D E N C O G N I T I V E C O S T S 26 / 43
33. February 2026 Two months ago · onset A R R I V A L Multi-agent orchestration. Several agents across multiple projects. M A T H 4 agents × 2 streams + master = 9 open contexts competing for 3–5 slots. F A I L U R E S Failed agents stick (Zeigarnik). Successful ones release. One failure in five dominates the day. S I D E - D I S C O V E R Y Every review surfaces unplanned findings; each one taxes working memory whether held or offloaded. "I wasn't multitasking. I was thrashing." C O N T E X T S W I T C H I N G H I D D E N C O G N I T I V E C O S T S 27 / 43
34. March 2026 One month ago · compounding S C A L E Long-horizon runs. Each agent holds hours of state I cannot carry in working memory. P A R A D O X Working memory is freed during the run — agent holds context, I don't. Calendar opens up. B O U N D A R Y End of each run, verification concentrates. Missing requirements (not agent bugs) dominate. S I L E N T A C C U M U L A T I O N Not looking at code for hours → comprehension debt compounds invisibly. "Free calendar. Invisible debt." C O N T E X T S W I T C H I N G H I D D E N C O G N I T I V E C O S T S 28 / 43
35. Comprehension debt. A fourth cost — and the one that shifts most downstream. D E F I N I T I O N The growing gap between how much code exists in your system and how much of it you understand. C O N T R A S T Technical debt (Cunningham, 1993) announces itself as friction. Comprehension debt announces itself as false confidence. M E C H A N I S M Programming is theory-building (Naur, 1985). Code is evidence of the theory. When the theory evaporates, the system can no longer be safely evolved. E V I D E N C E Kosmyna / MIT (2025) — 83% of LLM users couldn't quote essays they had just written. EEG: weaker coupling. Persists after LLM is removed. "The debt that doesn't show up in git blame." P A R T 0 4 · C O M P R E H E N S I O N D E B T H I D D E N C O G N I T I V E C O S T S 29 / 43
36. Apr 2023 → Jan 2026 Three years to three months ago · below the onset line A P R 2 0 2 3 Copilot inline. Every character visible as I typed. Comprehension = 100%. O C T 2 0 2 4 Chat copy-paste. Every chunk read before pasting. Line-level review mandatory. N O V 2 0 2 4 Cursor + Composer. Heavy review load — but line-level. I owned the code I shipped. J A N 2 0 2 6 Claude Code, single system. Abstraction-level review. Dependency diagrams traced actively. Theory intact. "Different mechanisms. Same result: the theory held." C O M P R E H E N S I O N D E B T H I D D E N C O G N I T I V E C O S T S 30 / 43
37. February 2026 Two months ago · onset T R I G G E R Multi-system scaling. Four agents across multiple projects = multiple theories to hold at once. C A P A C I T Y One engineer holds one system's theory well. Across multiple systems, each theory becomes partial. S T R U C T U R E V S . B E H A V I O R Until Jan I held both. From Feb I could only hold behavior — structure fell away. C O N S E Q U E N C E Gap between code and comprehension grows — per system, multiplied by system count. "Single system: I held the theory. Multiple systems: I held fragments." C O M P R E H E N S I O N D E B T H I D D E N C O G N I T I V E C O S T S 31 / 43
38. March 2026 One month ago · compounding + the emergency repayment ritual S C A L E Long-horizon runs across multiple systems. ~500K LOC managed across three months. S O M A T I C S I G N A L A heavy feeling in my head. Disturbing, not painful. The physical tell of comprehension debt at threshold. R E P A Y M E N T Stop. A day or two with paper and pencil. Rebuild the theory, one system at a time. C O S T O F N O T R E P A Y I N G Cannot safely evolve the system. Cannot make the right decisions for where to take it next. "The only thing that paid it down was reading. On paper. Slowly." C O M P R E H E N S I O N D E B T H I D D E N C O G N I T I V E C O S T S 32 / 43
39. The social media version. The profile we celebrate. T H E T H R E E - M O N T H S C O R E C A R D $297 / month on AI tooling ~500K LOC managed across three months Four languages in rotation: Rust · TypeScript · Go · C# Six hours of unattended agent runs Shipping in stacks with no prior experience "The thread you scroll past with the fire emojis." T H E C R A S H · 1 O F 3 H I D D E N C O G N I T I V E C O S T S 33 / 43
40. The perfect storm. The same profile, seen from the inside. T H E S A M E E N G I N E E R . F O U R C O S T S C O N V E R G I N G . Dopamine at peak (slide 20) — baseline elevated, no off-keyboard restoration. Decision fatigue silent (slide 24) — abstraction shift + long-horizon hid the drain. No smoke alarm. Context switching at peak (slide 28) — permanent working-memory reload, Zeigarnik load from long-horizon failures. Comprehension debt compounding fast (slide 32) — structure gone; only behavior remains. Repayment ritual can't keep up. "The exact profile we celebrate. The exact profile that crashes." T H E C R A S H · 2 O F 3 H I D D E N C O G N I T I V E C O S T S 34 / 43
41. What the crash looks like. The mechanism. The lived experience. M E C H A N I S M Sustained high dopamine → receptor down regulation. Baseline sensitivity drops. L I V E D E X P E R I E N C E Flat. Unmotivated. Sleep dysregulated. Appetite changes. Heavy-in-the-head chronic, not episodic. W H A T B R E A K S F I R S T Task initiation. You know what to do. You cannot make yourself begin. Dopamine is the chemistry of starting. This is not a failure of willpower. It is the bill coming due. "The crash is not a single event. It is four bills arriving at once." T H E C R A S H · 3 O F 3 H I D D E N C O G N I T I V E C O S T S 35 / 43
42. Recovery. Four costs. Four paths. Plan for all of them. D O P A M I N E B A S E L I N E (weeks) Sleep first. Schedule off-keyboard time; don't earn it. Remove loop-style reinforcement for a sustained stretch. Scrolling counts. D E C I S I O N - F A T I G U E R E S E R V O I R (a long weekend) Stop the drain. Sleep. Trust it to refill. C O N T E X T - S W I T C H I N G R E S I D U E (days) Drop the agent count to one or zero. Single-threaded work for a week. R E C O V E R Y H I D D E N C O G N I T I V E C O S T S 36 / 43 C O M P R E H E N S I O N D E B T (months) The emergency repayment ritual (slide 32), scaled. Paper and pencil. One system at a time. Rebuild the theory. "Bodies heal faster than codebases. Plan for both timelines."
43. The headline number. METR, 2025 — the single study that measures this talk's thesis. 19% slower 20% felt faster → 39 percentage points between measurement and perception. "Engineers cannot reliably feel whether AI is helping." METR, 2025 — 16 experienced developers, 246 real coding tasks, RCT. arXiv:2507.09089 R E S E A R C H · 1 O F 5 H I D D E N C O G N I T I V E C O S T S 37 / 43
44. The industry-scale pattern. It is not just me. T W O S O U R C E S . D I F F E R E N T V A N T A G E S . S A M E P R O F I L E . B C G , 2 0 2 5 "AI brain fry" — 14% of oversight-heavy users report cognitive fog; 61–62% of junior / associate workers report AI-related burnout. Y E G G E , 2 0 2 6 "The Dracula Effect" — "Engineers operating at peak AI-augmented intensity can sustain it for about three hours a day." "The symptom researchers measure and the symptom the practitioner observes are the same symptom." R E S E A R C H · 2 O F 5 H I D D E N C O G N I T I V E C O S T S 38 / 43
45. Workslop. The industry's receipt for downstream cost-shifting. BetterUp Labs + Stanford Social Media Lab · HBR, Sept 2025 41% encountered workslop in the previous month 1h 56m average cleanup per incident $186 cost per employee per month $9M / yr at a 10,000-person organization "Productivity doesn't move. Cost shifts downstream." R E S E A R C H · 3 O F 5 H I D D E N C O G N I T I V E C O S T S 39 / 43
46. Confidence vs. quality. Three studies. One pattern. Engineers trust AI more than they should. P E R R Y E T A L . 2 0 2 3 Developers using AI wrote significantly less secure code. Yet were more likely to believe they had written secure code. B E L O Z E R O V E T A L . 2 0 2 5 ~42% of LLM-generated code snippets contain a CWE-catalogued security flaw. S O N A R 2 0 2 5 96% of developers say they don't fully trust AI-generated code. Fewer than 50% review it before committing. "Trust but verify has collapsed into ship and hope." — Werner Vogels, AWS re:Invent 2025. "Trust, confidence, and quality have decoupled." R E S E A R C H · 4 O F 5 H I D D E N C O G N I T I V E C O S T S 40 / 43
47. The neurological receipts. What happens in the brain itself, measured. S T U D Y Kosmyna et al., MIT Media Lab, 2025. Your Brain on ChatGPT: Accumulation of Cognitive Debt. F I N D I N G 1 83% of LLM users could not quote from essays they had just written. F I N D I N G 2 EEG: weaker brain-network connectivity in LLM users vs. brain-only or search-engine users. F I N D I N G 3 The deficit persists after the LLM is removed. "Comprehension debt imprints on the physiology, not just the workflow." arXiv:2506.08872 R E S E A R C H · 5 O F 5 H I D D E N C O G N I T I V E C O S T S 41 / 43
48. Three lines to take home. Everything this talk argued, in three sentences. 01 Feel is not a measurement. 02 Plan for all four recoveries. Weeks · Weekend · Days · Months 03 Protect the theory. It outlasts the code. "The cost is hidden, not absent." C L O S E H I D D E N C O G N I T I V E C O S T S 42 / 43
49. Thank you. Ragunath Jawahar · Founder Legacy Code Labs · legacycode.com ragunath@legacycode.com “Holding on to our capacity for effort is going to be crucial not only for employability, but for a rich life.” — Tao Bojlén, 2026 April 21, 2026 · GIDS 2026