LLM withdrawal kunskapsarbetare avslöjar infrastrukturell beroende i KAIST 4-dagars dagboksstudie (N=10, frequent users, LLM-D12 instrumental dependency 16-30): temporär LLM-frånvaro exponerar fem workflow gaps, tre återvunna professionella värden, och social oundviklighet. För dig som projektledare betyder detta: LLMs fungerar inte längre som optional tools utan som implicit work infrastructure – workflows redan optimized around expectation instant assistance taken för granted. Workflow gaps: (1) asking people socially costly (“finger prince/princess” burden), (2) search keyword processing excessive (synthesize, reformulate, iterate felt cumbersome), (3) reading/writing lowered standards (less motivation comprehend/refine utan AI), (4) instant assistance loss prompted delay/avoidance (“putting things off till study över”), (5) underestimated task duration utan LLM (frustration, anxiety, schedule adjustments). Reclaimed values: clarity genom construct/trace own reasoning (vs LLM misaligned logic), ownership över outcomes (felt “not their own” med AI), awareness priorities between tasks (core profession-defining delegated för efficiency). Dependency inescapable: “developer who can’t catch a bug” (P6), “never done coding utan ChatGPT, don’t think I can, don’t have will to learn either” (P3), social norm pressure (“you’re the only one losing out if you don’t” P9). Value-driven appropriation framework: HCI community move beyond use/non-use binary → design practices där workers deliberately shape LLM engagement align professional values (clarity, ownership, integrity) snarare än pre-defined (generativity, productivity, proactivity).
Experiment design: 4-day withdrawal diary study + semi-structured interviews
KAIST recruitment (South Korea, second-largest paid ChatGPT users globally):
- 10 participants (6 female, 4 male)
- Screened på frequent use + dependency (LLM-D12 instrumental items)
- Diversity occupations: graduate students (CS, design, aerospace, electrical, semiconductor), DJ, content creator, developer, English instructor
- Average use: 6-21+ queries/day
- LLM-D12 scores: 16-30 (range 6-36)
- Compensation: 70,000 KRW (~$48)
Study procedure:
- Four consecutive working days voluntary LLM withdrawal
- Chrome extensions available (BlockSite, Bye Bye Google AI) if needed
- Web-based diary interface: write entry when urge use LLM
- Minimum 5 entries/day capturing: context, emotions, stress level (1-5), resolution strategy
- Daily reflection: strongest urge moment, feelings completing work, satisfaction, advantages/drawbacks
- Semi-structured interviews post-study (1 hour, in-person/online)
- Data: 200 entries, 40 reflection logs, 8h51min transcribed audio
- Thematic analysis: interview primary dataset, diary/reflections contextual
Fem workflow gaps exponerade genom LLM-frånvaro
1. Asking people felt socially costly: Participants accustomed seeking LLM assistance perceived asking humans som burden. Assumed others find it tiring. P10 avoided asking → feared seen som “finger prince/princess” (Korean slang: burden others med easily searchable questions). Preferred switching LLM services (ChatGPT → Grok) rather than consult people. Men withdrawal pushed human help → some found MORE helpful än LLMs. P9: “If study gone longer, say a year, discussions between people would be more active.”
2. Search keyword processing excessive: Conventional search engines (Google) increased difficulty retrieving information. Synthesize, reformulate, iterate keywords felt unexpectedly difficult efter LLM habit. P10: “ChatGPT gives you what you want in one go, but I had to put my mind into how I could get results with least number of searches.” P4: inefficient translate thoughts ways machine understand. LLMs = “technology with added power to understand humans.”
3. Lowered standards reading/writing: Reduced motivation comprehend texts eller refine writing utan LLM assistance. Viewed LLMs som enable higher-quality performance. In absence: willing accept lower-quality outcomes rather than invest time/effort (perceived inefficient/wasteful). P1: “When I write DMs or emails to professor in English, I always ask ChatGPT for revision, but I decided I didn’t have to write perfect English during study.”
4. Instant assistance loss → delay/avoidance: Accustomed immediate access → some abandoned tasks eller waited till LLM available again. P1: “I’ve been putting things off. I should start with them now that study is over.” Desire immediate support lingered → discomfort/frustration. P3: “I gave up tasks like organizing information that I usually would have asked LLM to do. Not being able to know what comes into my mind straight away was bigger thing than I expected.”
5. Difficulties adapting pace: Underestimated task duration utan LLM → not fully aware how much sped up work. Struggled adjust slower pace → frustration, anxiety, routine changes. P2: “I had to look up information on some equipments, but it took so long I didn’t think it was worth the time.” P8: nervous realizing how much longer find materials → adjusted schedules eller overtime.
Participants compared discomfort till:
- Living utan dishwasher/robotic vacuum (P8)
- Lacking convenience store (P1)
- No vehicle (P9)
- Missing Google search (P2, P6)
- Documents utan MS Word (P7)
Tre professionella värden återvunna genom självständigt arbete
1. Clarity in work: Clearer understanding genom construct/trace own logic. Contrast: using LLMs → tendency accept output utan unpacking each step. Model often gave reasoning misaligned original intent. P3: “I try to follow LLM’s logic, but I get frustrated because I have to keep telling it no, that’s not what I meant.” Engaging information greater depth + wider sources → better sustained understanding.
2. Ownership över outcomes: LLM-produced work felt “not their own” (P1, P3, P7) – less time/effort invested refining themselves. Completing tasks without “some other technology” (P5) required entire process ideation → decision-making → stronger ownership/pride. P1: “I had to do some ideation, this time without LLM, and it really felt like it was mine.”
3. Awareness personal priorities: Completing tasks utan LLM required more time/effort → selective prioritize. Revealed previously delegated even core profession-defining tasks för efficiency. Regained clarity which skills/activities crucial profession. P1 (researcher): “It’s a matter of priority. Since I’m a researcher, not a developer, I don’t think I’ll ever be in a situation where I have to write code entirely from my own head without any LLM support. I don’t think learning to code is something I need to invest in.”
Inescapable dependency: Individual + social dimensions
Too dependent to work without: ALL participants reported LLM use impedes skill development/acquisition. LLMs no longer simply assisting → replacing core aspects. P6: “a developer who can’t even catch a bug” (habitual copy-paste utan understanding context). Dependency shaped motivation/attitudes learning. P3 avoided tasks only ever completed med LLM: “I’ve never done it (coding) without ChatGPT. I don’t think I can do it at all without it. But I don’t think I really have the will to learn it either. ChatGPT helps me so well anyway.”
P8 compared growing reliance → gradual loss initiative: “It’s like how when you’re standing up you want to sit down, when sitting you want to lie down, when lying down you want to sleep. I could have just sat there, but LLM is trying to send me to sleep.”
Socially inevitable use: Participants regarded LLM use NOT optional tool men requirement remaining competitive. Inability use during study = disadvantage. LLM use seen basic work capacity rather than privilege. P3: “everyone is using ChatGPT anyway.” P9: “I used to think using LLMs was cheating, but now I think you’re the only one losing out if you don’t.”
Social norm pressure: P1 “no such thing as excessive when comes to using LLMs. Used to think, why would you ask LLM something so trivial, but now LLMs are everywhere. There is no line between what you can ask for help and what you can’t.” P7: “Now even my professor tells me to just get it (paperwork) over with with an LLM, and I feel it can’t be helped in this era.”
Value-driven appropriation: Beyond use/non-use binary
LLMs som infrastructure (not optional tools): Withdrawal reveals LLMs already deeply embedded functioning implicit work infrastructure. Workflows optimized around expectation instant assistance taken granted. Gaps should NOT be regarded individual shortcomings/skill degradation → traces technology adaptation där work practices reconfigured.
Conceptual shift needed: Previous dependency research focused individual-level cognitive effects + measurement. KAIST study conceptualizes från infrastructural role → normalizes use everyday work + generates social pressure adopt. Central challenge NO LONGER whether use LLMs → HOW use should be negotiated, for which tasks, to what extent.
Value-driven appropriation framework: HCI community design work practices move beyond binary use/non-use. Withdrawal made work values visible obscured by habitual delegation: clarity, ownership, professional integrity. These values DON’T align assumed LLM design (generativity, productivity, proactivity) men crucial sustain professional identity.
Building technological appropriation concept (Dourish 2003): people adopt technologies not just intended by providers men ways better fit work practices/contexts. Value-driven appropriation = workers deliberately shape LLM engagement align professional values rather than pre-defined.
Design implication exempel (education): Structured prompting strategy support student agency – restricts LLM response till students’ intent sufficiently articulated. Creates room users negotiate/sustain own professional standards.
Praktiska strategier projektledare
1. Recognize infrastructural dependency: LLM use redan embedded workflows. Team gaps during outages/restrictions NOT individual failures → systemic adaptation traces. Design contingency plans för LLM unavailability.
2. Protect professional values: Identify core skills defining team roles. Mandate LLM-free zones för critical thinking, problem formulation, ambiguity resolution. P1’s clarity: researcher vs developer priorities.
3. Counter social pressure: “Everyone using anyway” creates bandwagon. Establish team norms: när LLM appropriate vs when undermines ownership/learning. P9 shift: “cheating” → “losing out” needs governance.
4. Design value-aligned practices: Rather than ban/mandate, create structured engagement protocols. Example: LLM drafts require human refinement/validation capturing reasoning. Structured prompting forces intent articulation.
5. Monitor skill trajectories: Track whether team can perform core tasks utan LLM. P6 “can’t catch bug” warning sign. Regular LLM-free exercises maintain capability baselines.
Källa: “Oops! ChatGPT is Temporarily Unavailable!”: A Diary Study on Knowledge Workers’ Experiences of LLM Withdrawal” av Eunseo Oh, Suyoun Lee, Jae Young Choi, Soobin Park & Youn-kyung Lim, KAIST Department of Industrial Design, publicerad april 2026 (CHI EA ’26, Barcelona).