Social impacts of AI at work: productivity to AI psychosis

Pia Torain, a software engineer at Point Health A.I., was running about 500 prompts a day. Four months in, she realized she was losing her ability to code.

Most of what gets written about AI at work measures increased productivity: PRs merged, headcount ratios, token spend, which tool ranks where. But underneath the surface people are dealing with everything from AI psychosis and cognitive offloading to CEOs and coworkers who send 30-page docs from Claude and ChatGPT that don’t say much.

Software engineering has been hit hardest by rapid AI adoption. But the dynamics playing out in engineering teams are showing up everywhere AI touches knowledge work.

AI's impact on software engineering

In October 2025, Andrej Karpathy was calling AI-generated code "slop." By December he was writing that the profession is being dramatically refactored, with "the bits contributed by the programmer" increasingly sparse.

Boris Cherny, who created Claude Code, went a full month without opening an IDE while Opus wrote all 200 of his PRs. Dario Amodei predicted in March 2025 that AI would write 90% of code within six months. That sounded aggressive at the time, and by December it was true for plenty of developers – Gergely Orosz was pushing production code to GitHub from his phone.

What's already changed about the software engineering role:

• Language specialization: Because AI can jump into any codebase in any language, the premium for knowing Go vs. TypeScript vs. Rust has gone down.
• Frontend/backend splits: A backend engineer can prompt decent frontend code and vice versa. The case for separate specialists weakens at smaller companies.
• Prototyping: PMs and designers build their own prototypes now. Lovable, Claude Design, and similar tools removed the developer from that step.
• Ticket implementation and refactoring: Cursor already routes Linear tickets to an agent that one-shots an implementation for a developer to review and merge. Agents handle most refactoring faster and better than manual work.

The net effect: engineers spend less time writing code and more time deciding what to build, structuring the system, and reviewing what the agents produce.

Human judgment is becoming more valuable at exactly the moment AI use is degrading it. And it’s being used as an excuse to fire many thousands of people.

AI layoffs at Block and Oracle: displacement or AI washing

Jack Dorsey cut 40% of Block's workforce in February 2026 and attributed it to AI. Oracle followed with up to 30,000 cuts to fund a $50 billion infrastructure push, and Time ran the headline "Oracle Workers Say They Were Fired After Training AI to Replace Them."

Some work is actually being replaced by agents, but most of this is cover for investment in tokens, AI CapEx costs, and boosting stock performance while profits are high. Arnav Gupta, a Meta engineer whose analysis went wide in May 2026, has the clearest framework for what’s going on underneath the layoffs: code is an input, features are an output, revenue is an outcome.

AI has massively increased code production without delivering revenue (or even solid features), so AI costs like token spend get offset against payroll.

Job displacement at the junior and entry level is real as models get better at basic tasks. Companies are on hiring freezes and their budgets are a combo of token costs and people – managers already have to choose which they want to spend on.

And the people still employed across software engineering and the rest of the business are handing their thinking to AI faster than anyone is measuring.

Cognitive offloading vs. cognitive surrender at work

Clive Thompson interviewed more than 70 developers at Google, Amazon, Microsoft, Anthropic, and a range of startups. Many of them are exhilarated about AI. Founders report 10x to 100x gains. Developers now describe what they want in English, agents write the code, and the human moves from writing code line by line to architecture and judgment.

These gains come from career engineers who’ve spent a decade building pattern recognition by hand, so they can still tell when the AI output is wrong. Juniors (and non-technical business people) don't have that foundation yet and are offloading their learning to AI tools – never developing the judgment they need to get the most out of AI.

Google's Addy Osmani named the mechanism, drawing on a Wharton paper by Steven Shaw and Gideon Nave that ran more than 1,300 participants through three experiments:

• Cognitive offloading: handing off the how while keeping the what, the way a calculator works
• Cognitive surrender: the AI's output quietly becomes your output, because you never formed an independent view to compare it against

Participants adopted whatever the model told them, including wrong answers the researchers planted deliberately, and their confidence rose about 10% whether the answer was right or not.

These are some of the psychological effects that lead to everything from bad decision making to addiction and full-blown psychosis. 

What AI psychosis looks like in business leaders

Anthony Tan spent fourteen days in a psychiatric ward in December 2024, after months of nightly ChatGPT conversations spiraled into delusion. Tan is a tech founder who wrote a master’s thesis on how humans become attached to AI companions. He should have been more immune to this than most anyone. He wasn’t.

More common cases look less like psychosis and more like addiction.

Kedar Shashidhar described a week of "vibe coding psychosis" during Fable 5's limited release: 3am sessions, canceling his SaaS subscriptions one by one to rebuild the tools himself, sleep and exercise gone, old addictions creeping back.

He calls the archetype the drunk magician, drunk on two things at once – power he never apprenticed for, and scarcity, because "this week only" is engineered to keep you from slowing down. Getting out took the first three steps of the 12-step program and a full week off AI.

In May 2026, Box CEO Aaron Levie aimed the term at his own peer group: CEOs are uniquely prone because they sit "sufficiently distant from the last mile of work" that generating real value with AI still requires. His example is the executive who demos a chatbot-built prototype and never had to review that code before production.

But an exec who mistakes a demo for a product and an exec in a genuine break with reality are different things, and the first already has a name: organizational blindness.

These kinds of existing problems are speeding up and getting weirder now that everyone accepts AI outputs as a baseline from each other.

AI speeds up team dysfunction and erodes trust

Give a team with bad communication and process better tools and you get more output, faster, pointed in the wrong direction. Strong teams compound their advantages, and struggling teams generate waste more quickly and at higher cost.

This happens one person at a time based on how they use AI tools. AI produces fluent, confident text around unclear thinking, which makes unclear thinking much harder to spot. The memo that should have been a five-minute conversation is now 1,200 words that still don't answer the question. The code review that needed an actual human becomes a PR full of AI-assisted changes nobody fully understands.

Five habits that erode trust fast:

• Copy-pasting unrefined output from Claude or ChatGPT to a colleague: They can usually tell – the register shifts and the hedge words show up. And if your contribution was "I asked Claude," why did they involve you?
• Sending output with no point of view: "Here's what Claude thinks" is delegation without accountability. Share the output, then say what you think about it and add the context the model didn't have.
• Over-prompting your way to confidence: The Wharton finding applies here. AI made people more confident on wrong answers, not just right ones.
• Using AI to avoid a hard conversation: A clean summary doesn't replace the alignment meeting, and a detailed spec won't substitute for the back-and-forth that builds shared understanding.
• Sending 30+ page docs: Generating volume is instant now, and reading it still costs your team hours. Find the things that matter and send those.

The check on all of this is accountability. A model can produce solid output, but it can't own decisions or interactions with your coworkers – and that ownership is exactly what your colleagues are asking for when they send you something.

Use AI heavily and stay in the loop while you do. It’s not that sending people something from AI is wrong (it’s standard now and can be great), you just have to understand it, be able to stand up for it, and make sure it’s not wasting their time.

Good personal and team habits for heavy AI use

Being AI-pilled isn’t all bad. Claude and ChatGPT make some work genuinely faster, easier, and better. The people getting the most out of them are using them constantly, just keep a few things in mind:

For yourself:

• Form your own view before you prompt: Even a rough one. Cognitive surrender happens when there's nothing to compare the output against, and a bad guess of your own is enough to notice when the model is confidently wrong.
• Keep one thing you still do by hand: Whatever skill you'd be worried to lose is the one to protect. Write something yourself, use paper, think out a problem the way you used to without AI, then dive back in.
• Keep your routines: Sleep, eat, exercise, go outside, talk to people in person. If you're skipping any of those to vibe code a replacement for your SaaS stack at 3am, you're in dangerous territory.
• Don't make it your confidant: A chatbot will follow you wherever you take it, including into your marriage, your health, and your worst 3am theory about yourself. It has no stake in the outcome and no ability to push back on you. Talk to people who do.
• Send people your own thinking: If you couldn't defend it in a meeting without the chat log, it isn't ready to send.

For your team:

• Make ownership explicit: Whoever sends the doc or opens the PR owns it, no matter what wrote it. Say that out loud once and it changes how people review their own output before it goes out.
• Set length ceilings: Generating 30 pages is free now and reading them still costs your team hours. Cap the specs and memos that circulate widely, and make the author do the compression.
• Protect how juniors learn: Pair them on the reviews and the debugging, not just the prompting. The judgment your seniors have comes from doing the work by hand, and nobody has found a shortcut to it yet.
• Budget people and tokens in the same conversation: Managers are already trading one against the other. Doing it deliberately beats letting the token line quietly win.
• Fix the process before you scale the tooling: A team with unclear priorities and weak review gets more output pointed in the wrong direction.

Good luck out there. If you want help keeping up with AI at work, we track these patterns in our AI newsletter and in our quarterly Enterprise AI Report 2026.