The Human in the Loop Manifesto

Listen to webs read this article (9 min)

Chapters

Lately I've been reminded that learning needs time to breathe. It needs loops and collisions, and it needs new people and new ideas. The same way travel makes us appreciate home, time spent in other people's heads sharpens my own views. I've met more new people in the last few months than in the three years before that combined. I joined a book club whose first meeting refreshed my soul. I've explored new activities with my dogs, started reading a fiction series ("Hi, Zev!"), and spent more time outside just being.

I've also spent more time on LinkedIn, which is probably not where you expected this to land, so stick with me. We all know we hate being here, but does it have to be awful? I wondered what would happen if I showed up as myself and engaged with people more deeply than a passing reaction (most of the time).

One of those conversations was with my friend Goodness, about creativity in education. He pointed out that something I was doing in education came straight from the Agile Manifesto (opens in a new tab): individuals and interactions over processes and tools. That got me thinking about how I use ai across corporate work, education, and software development, and what my own ai manifesto would be. Here it is, and then I'll walk through each line.

Reasoning over rubber stamps
Augmentation over novelty
Risk over rules
Learning how it fails over trusting that it works

Reasoning over rubber stamps

Human reasoning is essential before, during, and after any ai interaction. My co-founder at The Intelligent Hoodlums (opens in a new tab), Mike Lang, coined a sequence for this: human, analog, digital.

Start with the human: our experiences, perspectives, interests, and desires. Map those out with an analog tool like pencil and paper. Then shift to digital and start working with ai.

Throughout the process, each exchange is an evaluation. Is this the direction I want to go? Is this true and valid?

When the work is complete, the questions change. Does this reflect what I wanted from the end product? Does it convey what I meant it to convey? Is there extra padding? What value does it bring to me or my audience?

The same goes for approving anyone else's work. If you approved it, can you say why? What did you look at? What did you expect to see that would have told you not to approve it? If you can't answer, you didn't review it. You rubber stamped it.

When we rubber stamp instead of reasoning through the output, we create middling slop bombs. No one wants that.

Augmentation over novelty

This one comes from my product and developer life. For the last couple of years, executives have been asking, "What's our ai story?" My answer is that you shouldn't have one. You should have customer pain points you solve, and if ai is the right tool for one of them, great. Shoving ai in so you can say your product uses it is a solution in search of a problem, and it generally doesn't bring customers value. I've been in that room. When product wanted ai in the product, I looked at the state of the data first. Putting ai on top of it would have produced a mess our customers paid for, so I built a plan to shore up the data and extend the ai from there. Augmentation means building on the strengths of a system that already works.

The same applies to you as an individual, and in some ways this is an extension of the first line. You don't need to jump to ai for everything. Start from the assumption that you don't need it, and filter problems into the "Should I use ai for this?" space. Three questions help me decide:

Is it low risk if it's wrong?
Is the popular answer what I want?
Is doing it myself the point?

A new flavor of high-protein overnight oats passes easily. It's low stakes, the most popular answer is exactly what I'm after, and I'm not trying to become a recipe developer. A student's paper fails, because the whole point of the paper is practicing how to find your own view and say it.

Instead of reaching for ai for the novelty, ask how it augments your problem solving or your executive function.

Risk over rules

This line comes from my security work, and it's why "Is it low risk if it's wrong?" is the first of my three questions. When I do a threat analysis, I weigh what's most valuable against how easy it is to reach. With ai, the question becomes: what are the ways it can get it wrong, and what happens to the humans when it does?

This school year, New York City paused student-facing generative ai through eighth grade (opens in a new tab), and Los Angeles Unified paused generative ai on district devices for students in every grade (opens in a new tab). One part of New York's policy gets the risk right. Teachers can't use ai to grade, and a grade is a judgment about a child that the teacher should be able to explain. The student bans are where I disagree. A blanket ban assumes the alternative is kindergartners with unfettered access to any LLM they want, and there's a lot of room between those two. What would it look like for a teacher to design guided experiences that model responsible ai use, and to use ai as a tool for scaffolding learning?

The same thinking applies to how we design products. Before anything ships, ask what happens if it gets it wrong, and who gets hurt.

The example I think about most is ai detection. I believe we'll look back and see that we ruined a fair number of lives when institutions suspended or expelled students because a tool like GPTZero claimed they used ai when they didn't. One study across 32 university courses (opens in a new tab) found GPTZero flagged 18% of real student work as ai-generated. Another (opens in a new tab) found detectors flagged more than half of the essays written by non-native English speakers, so the harm lands hardest on second-language learners. A 1% error rate sounds small until you do the math. Run a million honest essays through a detector that's wrong 1% of the time, and 10,000 students get accused of something they didn't do. That's too many lives to put at risk.

Learning how it fails over trusting that it works

I know a lot of people who choose not to engage with large language models at all. That's getting close to impossible now that they're built into the products we use every day, from Google Docs to Outlook to search engines. I'm of the mindset that you need to understand the limitations to speak about ai effectively and take part in the discussion around it. In my experience, people who are afraid of ai quickly see its limits once they start using it, and it becomes much less scary.

The same goes for the people we teach and lead. For too long, "I'm not a technology person" has been an acceptable thing for a teacher to say. Teachers can learn how ai works faster than their students, which puts them in a position to guide how students use it. Leaders who hand ai to their teams without teaching them how it fails get slop passed from person to person.

There's still a lot of nuance, and a lot we don't yet understand about how ai could change the way we think and process information. We've also had this debate before. Many of the worries about ai echo the worries about calculators, search engines, and autocorrect.

Underneath all of it, ai is a tool. It has pros and it has cons. The best way to understand what it can and can't do is to use it and see where its limits are. Don't just mindlessly trust that what it gives you is true or accurate.