↓Skip to main content

Stages of AI Grief

·5 mins

I’ve noticed that people go through predictable stages while adopting LLMs into their work. Listen to your coworkers talk about LLMs and you’ll locate them somewhere along this progression:

  1. Denial: They refuse to try it. They look down on those who have adopted it. They explain it away as a new version of existing tools… it’s just a new google, a new stack overflow, a better autocomplete. Perhaps they explain how an LLM can’t be intelligent because it just guesses the next right word.
  2. Minimizing: They try it and cling to any mistake the LLM makes. Whenever it fails to one-shot a problem, they take it as proof that it’s overhyped. The fact they still need to contribute and drive the process is taken as proof the LLM isn’t the miracle everyone thinks it is. Perhaps they suspect they’ve mastered their tools in a way others haven’t, and this means they don’t see the same leverage from LLMs.
  3. Cope: They agree the LLM works, but it could never do what they do. They can understand how it helps some people, maybe even most people, but their contributions are special and require human insight. And besides, what LLMs can do is nothing we couldn’t do fairly easily without them.
  4. Acceptance: They embrace LLMs as useful tools. It did some tedious work for them. It got them away from their desk a little earlier on a Friday. They still may use AI synchronously and serially as opposed to asynchronously and in parallel. They probably still do a lot of work by hand the LLM could do for them. They still have some tasks they’re precious about doing themselves. They might occasionally regress to earlier stages.
  5. Exuberance: They begin to use LLMs creatively, experiment freely and customize workflows. They use LLMs to build tools for using LLMs more effectively. They give tasks to the LLM by default and take them over themselves only when that fails. They achieve flow states and higher levels of productivity than ever before. They feel empowered instead of threatened for the first time. All the gains they’ve made make it easy to discount any shortcomings. It is not possible to regress to earlier stages anymore.
  6. Addiction: They chase the dopamine rushes of early successes like a gambler. They blame model updates for ruining productivity. They complain AI used to be useful but it’s been ruined. They jump from model to model, tool to tool, trying to get back to their happy place. Work starts to feel like work again.

I suspect most people live in stages 2-4. They can’t flourish for a variety of reasons:

  • Perhaps LLM adoption is being forced on them from above. There are metrics, leaderboards, lunch and learns, new AI categories on their annual review, and worst of all consultants. That’d be enough to turn anyone off.
  • A lot of LLM tooling is very poor. Shoving an LLM chat box into an existing tool rarely makes sense. This is done to market that tool to VPs who know they need to adopt AI this quarter but don’t know how. VS Code, I’m talking to you. Bad tooling sees the LLM as a capability to add to other tools. Good tooling puts the LLM at the center and adds capabilities to it.
  • They can’t shake the anxiety that the LLM will take their job. This is not unfounded. There is a constant drip of news stories about AI driven layoffs. Refusing to adopt LLMs won’t save you. For the foreseeable future, it won’t be an LLM that takes your job, it’ll be a human using an LLM.
  • LLMs have cultural baggage. People hate change and AI is being sold as something that will change everything. People hate uncertainty and no one knows exactly how this new technology will change our world. Narratives about our impending doom spread much more easily than the truth.
  • LLMs and other AI branded technologies have been used to make very obnoxious content online where so many of us spend most of our time. We have to be constantly vigilant in a way we never had to before to avoid being “tricked” by an AI post. We have to worry about how long until someone we care about gets scammed by AI.

I predict we will look back on this as a golden age of AI. Tokens are being subsidized by investor money. The technology isn’t as locked down now as it will become in the future. I used Claude Code to proofread this post (and it suggested capitalizing those two Cs). I predict that in the future LLMs will become sophisticated and specialized enough to prevent them from being used “off label” like that unless you want to pay the big bucks for general-purpose tokens.

Furthermore, I think LLMs are improving exponentially. This may be the beginning of a new Moore’s law that will reign for a generation, or it may be the exponential regime of a logistic curve that could reach an inflection point in the next couple of years. Either way, the exponential nature means that a gap of months can mean a significant difference in productivity.

And I don’t think productivity needs to feel toxic. LLMs really just represent being able to actualize your ideas while you still care about them. If you have an idea that only you believe in, you no longer need to sacrifice nights and weekends to see it through. LLMs are no cause for grief.