In the strange, capricious and often messy human world that LLMs are emerging into, it’s not unusual for people to use a culinary metaphor to describe something fundamental, a principle, or a process, or a way of thinking. Such is the case, apparently, with AI, as we wrestle with how to integrate the powerful technologies at our fingertips into our lives.

Some people call it a “human sandwich.” Others call it an AI sandwich. But since the humans are the bread (human first, human last) in this concept, I’ll go with the chosen semantics of my favorite podcaster, Nathaniel Whittemore, who touched on the idea in a recent episode of AI Daily Brief. The human sandwich is essentially a way to keep the human “in the loop” and using his or her faculties, even as AI does more and more of the heavy lifting in any given project.

The idea, in a nutshell, is that AI’s input should be bounded by human initiative and direction. Humans should take the lead, at the beginning of a project, sketching out the rough drafts, and orienting the thing, before calling in AI to refine, or add, or analyze. Then, at the end, the human should be deeply involved again, in evaluating the final result, in asking for changes – in short, taking control again, getting a handoff from the technology, and bringing the wagon home.

Describing the Human Sandwich

“The human parts are where the substance really lives: your context, clarity, and creativity on one side, and your editing, judgment, and tone on the other,” writes Andrea Wightwick on Substack at her Business of Adulting column , also revealing that she likes to call GPT “Chatty.” “The middle (AI) can be good, but it’s hard to eat by itself. And one soggy bun, and the whole thing falls apart. It’s the same principle we learned long before ChatGPT showed up: trash in, trash out. When the tool launched, early adopters quickly realized that the better your inputs, the better your outputs. Clear, detailed, well-informed prompts produce strong, thorough responses. Human intelligence and expertise fact-check and finesse the AI generation.”

There’s a difference, in other words, between doing the work, and phoning it in, just putting in a basic prompt and letting AI come up with, well, whatever.

“The expectations for what AI tools can achieve on their own are pretty high,” writes Harpreet Khurana for Russel Reynolds, designating the ‘top slice’ as human curation or insight, the ‘filling’ as AI contributions, and the ‘bottom slice’ as human decision. “So much so, that some people predict it will replace the human hand altogether in plenty of strategic areas. But, in fact, the opposite is true. If anything, it makes the human touch more important.”

Not a Completely New Thing

Since we’ve only been hearing about AI for a few years, really, one might be forgiven for thinking that the temptation to outsource cognitive work to computers only dates back to, say, 2021. Not so, according to Dr. Lane Freeman, State Director of Online Learning, North Carolina Community College System, published on the Online Learning Consortium , who writes:

“Long before AI entered the picture, we were already outsourcing pieces of our thinking to technology. Many of us can remember a time when we knew dozens of phone numbers by heart. Today, most people can recite only a handful—if any—because smartphones remember them for us. It is not that we lost the ability to memorize numbers; we simply stopped investing mental effort there because a tool took over the job.”

The new reality, Freeman concedes, is different:

“That trade-off is acceptable for phone numbers,” Freeman writes. “It is far more consequential when the thing we are outsourcing is thinking itself—our creativity, our critical judgment, our ability to wrestle with complex ideas. When learners turn to AI too early, especially at the start of a task, they risk letting the tool define the boundaries of their thinking. Generative AI often produces a polished, “cookie-cutter” response: plausible, organized, and confident.”

At this juncture, Freeman made an observation that I found valuable, arguing that it’s that first ex nihilo creation that sets the stage and directs the human to a result.

“Once a human sees that answer, it becomes cognitively difficult to imagine alternatives,” Freeman writes. “The first answer becomes the anchor. Human-first work interrupts that pattern. When individuals or teams begin by brainstorming, sketching ideas, or attempting solutions before consulting AI, they keep those neural pathways for creativity and critical thinking active.”

I think that’s a good way to describe the dynamic: if you, as a user, come to the plate empty, and let AI fill up everything, you get an “AI-only” result. And people may still be able to tell. But even if they can’t, think about what happens when you stop utilizing those parts of your brain.

Maybe, at the end of the day, the sandwich is a good metaphor. It certainly delineates parts of a multi-step process. I’m fond of talking about the late Marvin Minsky’s book, “Society of the Mind,” where he mentions that the human brain is not one computer, but hundreds of smaller ones joined together. The human sandwich model shows how collaborative AI use is, or should be. Think about it.