AI and Cognitive Offloading

Language models make it tantalizingly easy to offload cognitive work to a machine. Our brains are lazy and thinking hard is actual work. The challenge with this offloading is that the "work" that gets completed by an agent when we do it often has no grounding in any reality where it is useful. The work doesn't go away. It moves to deciding what you want and saying it clearly.

Offloading a meal plan#

Let's imagine we need to create a list of meals we are going to cook for the week. We're tired from all the other stuff we've been doing all week, so finally, we get to offload some tasks to a machine, like meal planning. At least that is one thing we don't need to do ourselves.

If you genuinely do not care about what you cook and all four of you are going to eat all 21 meals that week, this might work out ok for you. But the likely reality is that there is actually a lot of relevant detail that requires effort and thinking that you haven't provided, and which is needed to get a good result.

AI can't read your mind. It doesn't know that no one wants to eat leftover chicken but is fine eating pasta every day. It can accommodate this but it has to be told. You have to give it the context somehow, whether via past grocery lists, text threads discussing what you're going to buy that week, or something. But this is work, cognitive effort, thinking. And if you avoid doing it, you're going to think AI is pretty useless.

AI changes the work#

AI doesn't really make the work go away, it changes the work. You used to have to search for the recipes, then find the ingredient list of each, then collect them all together into a single grocery list. Now you can work with an agent to help decide what to cook, pick some options you like, have it find (or generate) recipes and a grocery list from that. But you still have to be involved. You have to have opinions about the outcome and you have to make them clear to the agent.

A presentation#

This change creates a strange and murky environment for work. It used to be the case that if we wanted to create a presentation, we just had to do it. Sure, we could copy an old presentation or use a template for a starting point. Then we could copy content from all over the place, that maybe we wrote, maybe other people wrote. Regardless, the presentation had to somehow get assembled.

We'd have to make decisions about which words went where, whether there were pictures, or whether we removed pictures from a previous version. We were at least, kind of, involved in the process. With AI, this no longer holds true. Now all it takes is:

That is enough to get you your presentation. But what have you really done? Hopefully you've at least provided some document about Q2 for the agent to reference. If not, your presentation is about ACME Co. with some questionable numbers.

If the presentation is about your company, you now have molded one Q2 results document into another. What are you doing with this presentation? Are you going to present it? Are you going to share it with coworkers? What is even in this thing? Are all the numbers right? Does it have all the things it needs to have in it?

If you don't care about the answers to any of these questions, fair enough. Feel free to stop reading. If you do, consider how much easier it would be to articulate these things upfront and specify what they are, rather than having to reactively go through an entire checklist of items based on the work the agent does, tediously verifying and either fixing or prompting it to fix all its little mistakes and omissions.

Say what good looks like#

This checklist, the articulation of what "done" and "good" looks like, is what the AI needs. This is the information you need to give to it for the result to have a chance of being what you want. But this requires you to think and show what done and good looks like. Can you explain what you want? Do you have examples? Are there exceptions to rules? When would an anomaly or missing data require your intervention to fix?

And if you can, then answer these rhetorical questions. Literally, write down the answers and send them to the agent along with your instructions in a paragraph. This is the context that helps the model do a good job.

There is no shortcut#

This might sound like a lot of work, and it is. If you want to get good results from AI, you must do this work. There is no alternative. If you want to do interesting, specific things, you have to go deep on your vision and ideas. And this requires work, thinking, and cognitive effort.

Working together#

The work I do with my clients focuses on tapping their deep domain expertise and ensuring they remain in the driver's seat when working with AI. We work 1:1 on their actual work, building tools or automating processes with agents. This involves:

• talking about your work, what's hard, what's frustrating, and what needs to work better
• prototyping a solution to solve your most pressing challenge
• iterating on that solution until it looks and feels right for the job
• capturing the learnings so that next time is easier and your efforts compound

The most impactful contributions I make help to prevent this cognitive offloading. That means less

or

and more like that paragraph that addressed all the questions. The real value from AI isn't "my work is so easy now," it's "I can't believe I can do my work like this now."