Workers given four days without AI discover they forgot how to work

What happened

Researchers took away ChatGPT from 10 frequent users for four days and documented what happened — work slowed down, people felt lost, and they realized how dependent they'd become on the tool for everyday tasks. The study reveals that AI has quietly become as fundamental to office work as email, and losing it exposed both how much workers now rely on it and which professional skills they'd stopped using.

10 frequent LLM users in study

4 days of LLM withdrawal observed

assumed LLMs are tools that knowledge workers can pick up and put down without disrupting core work.

found Four days without ChatGPT exposed that LLMs have become infrastructural — workers could not execute tasks, felt lost, and revealed which skills they'd stopped using.

Why this matters

At some point between 'interesting new tool' and last Tuesday, AI crossed from optional to load-bearing. Ten workers lost access to their LLMs for four days and couldn't finish their jobs. Infrastructure fails quietly and expensively, and most organizations have no idea this threshold has already passed for their teams.

They got the tool because they wanted to do more. Four days without it and they couldn't do what they used to do.

Who wins, who loses

who wins LLM vendors aka Eye of Sauron

who loses Knowledge workers

also Anyone who manages teams, approves AI tool budgets, or plans for business continuity, which is now the same conversation.

Jargon decoder

LLM withdrawal temporary removal of access to ChatGPT or similar AI tools

infrastructural so embedded in daily practice that workers treat it as a foundational system, not an optional feature

The signal

Why this hasn't landed yet

The finding contradicts the default optimistic story about AI as a tool that enhances workers rather than substitutes for skills, and a study of ten people over four days is easy to dismiss as too small to matter, which is convenient.

What happens next

Risk and continuity officers, who have so far treated AI tools as productivity software, will eventually have to reclassify them as operational dependencies, and the organizations that do it reactively will do it after the outage.

The catch

The study was of 10 people. No one is really arguing the results, because no is at all surprised. But still, 10 people?

The longer story

The longer arc

Email was optional in most offices until roughly the mid-1990s, then became load-bearing almost overnight, and the organizations that treated it as a productivity add-on rather than infrastructure paid for that distinction the first time servers went down. This study suggests that window for reclassifying LLMs has already closed.

Part of a pattern

Sits inside a growing cluster of research questioning whether AI assistance is purely additive to human capability, a framing that has dominated coverage since 2022 but is accumulating empirical friction.