← The Lost Innovators

Essay

Who Gets to Think?

On AI, lost innovators, and the question I’ve been carrying since I was a teenager.

By William Stafford, ADN, LI-AIast3

I. The Kid They Called Slow

I was a teenager when I first sat down in front of an IBM 5100 — a fifty-pound portable business computer that ran BASIC, sold years before the famous IBM PC existed. My mother worked for a music school that sold concert grand pianos and cathedral organs, and I was friends with one of the owners’ sons. The computer was at his house — that’s where I used it. Reading was hard for me. Spelling was harder. Math was the cruelest trick of all: I could explain exactly how to solve a problem, step by step, out loud. But you had to write it down and show your work — and somewhere between my head and the page, the numbers flipped. The answers came out obviously wrong. When my mother asked my math teacher what the problem was, the teacher said: “Well, he gets the answers wrong.” That was the answer. Nobody around me had a name for it. I would learn much later it was dyslexia and dysgraphia. The thinking was never the problem — the paper was. I saw things backwards, I couldn’t read my own handwriting, and putting ideas down was excruciating. The teachers had a simpler name for it. They called me slow. Not stupid, exactly. Just slow.

But put me in front of that machine and something happened. I wrote a little game in BASIC. A martial arts game. Nothing big. But I built it. I had programmed a computer! Me — the kid who couldn’t spell, whose math answers came out wrong, and whose handwriting no one could read, including me. I sat there and thought: there have to be others like me. People with ideas, questions, and theories they’d like to pursue, and no way to see those dreams to fruition. Someday a thing like this is going to make a huge difference for people like me.

There was nothing I could do with the idea back then, and nothing I could do with that computer beyond sitting at that table — computer science was a universe of math away from this kid. Even though I was going to college, I could never dream of entering that field. So I just had a feeling, and I tucked it away. But it never left me.

I went to college on an athletic scholarship, excelled athletically — in college and at the national level — but kept struggling in classrooms, and I left without finishing my degree. I came back later and got an associate’s degree in nursing. So I had a career, a family, did a lot of living. Next page: word processing comes along — big help, spell-check and all. Then AOL and the internet happen, and the world becomes a knowledge repository at the speed of your dial-up phone line. What a game changer for everyone, and it made so many tasks easier and faster for a dyslexic. Now I could ask the internet questions and get answers without a trip to the local library and a trudge through the Dewey Decimal System — more numbers! But all of it, helpful as it was, didn’t touch the real issue: access to the room. I was still that teenager, without the graduate degrees and credentials to make my essays, theories, or hypotheses land with any weight. To do real work. The kind that gets taken seriously. The kind that might contribute to our understanding of any of the millions of questions that could make our lives better. And — if I’m honest — maybe to be remembered for doing something. That part’s vanity, I guess. I’d rather admit it than pretend it isn’t there.

So the puzzle of me went on, unexplained: could someone like me ever do that kind of work? That question stayed unanswered for decades — not because I stopped asking, but because the tool I’d imagined at that table didn’t exist yet.

Then ChatGPT came into my life.

II. The Grand Theory That Wasn’t

So here we go: learning what this new mysterious AI thing could do. Everyone was talking about it — half saying it was the greatest advancement ever, the other half that it was the end of the world as we know it. (And some of them felt fine.) Me, I started thinking I might have found my tool — the one from the table. So, let’s do it.

I’ll not dress it up. I had a theory in my head, half-formed, about civilizations and information — that the bandwidth and speed at which information travels is directly tied to how far and how fast a civilization advances. I’d been turning it over for years without knowing how to write it down. And yes, back then I thought no one had ever had this thought before — or at least not the way I wanted to say it. With AI, suddenly I could. I would type in a half-thought and the machine would help me find the words. And by the way: it passed no judgment on my misspellings. I would ask it to compare what I was saying to other thinkers, and it would tell me about people I’d never heard of — Hidalgo, Henrich, Dennett, Brian Arthur — people who’d written books about ideas next to mine. I drafted chapters. I built evidence dossiers. I wrote what I thought was a unified theory of how civilizations advance.

Eventually, I had something that looked like a book.

Then I made the mistake of asking another AI — Claude — to look at it honestly. I asked it to act like a PhD advisor. Don’t just tell me it’s great — that was the status quo of my relationship with ChatGPT — be real with me.

And it was.

The first thing it told me was this: most of what I’d written was already in other people’s books. Hidalgo had said something similar in 2015. Dennett had said something similar in 2017. Brian Arthur had written a book in 2009 making part of my argument. Joseph Henrich had a whole framework in 2015. My grand theory wasn’t so grand. It was a restatement of work I hadn’t known existed.

That was a real emotional roller-coaster — from scholar to derivative undergraduate in about four seconds. A hard realization. So I asked it straight: is there anything noteworthy here? I just want the truth!

It started with “No,” and here’s why. Then it went through my paper like a sushi chef through a tuna.

As I read the pages of “here’s why your work sucks,” I was that teenager again. “Well, he gets the answers wrong.” I sank further into self-pity.

And then I came upon this.

It said, “The big theory isn’t original.”

But.

I stared at that one word:

“But.”

For longer than I’m comfortable admitting.

Then I read on.

But there’s one piece of it that might be. The piece that comes out of your own life. The piece you’ve been carrying since childhood.

III. The Lost Innovators

That piece was this. Throughout history, lots of people have probably had ideas worth taking seriously, but they didn’t have the credentials to be taken seriously. They didn’t have the right degree, the right school, the right contacts. They couldn’t get into “the room where it happens,” as the Broadway musical Hamilton puts it. Their ideas died with them.

We don’t see those people in the historical record because history mostly preserves the innovators who made it through the gate. I call them the lost innovators.

History is full of survivors; it is almost silent about the people who never got a chance.

We know how many Einsteins succeeded. Researchers have even studied the “lost Einsteins” — kids whose inventive potential died of no exposure and no opportunity. The lost innovators are their grown-up cousins: adults still carrying the questions, with no way to put them before other minds. So the waste itself isn’t news. What I was carrying was narrower, and it came straight out of my own life. The problem was never who could think. It was whose thinking could enter the record.

The question I’d been carrying was whether technology might finally make it possible for more of them — more of us — to contribute. Not whether AI would make existing scientists more productive — that’s a story other people are already telling. The newer question is whether AI finally unlocks the door to that room where it happens — whether people who would never have written a paper can now walk in and write one. The bar inside that room doesn’t move — the standards of evidence stay right where they’ve always been. What changes is who gets in the door to take a run at it. People without the PhD. People who read slower and write slower, and who’ve been belittled for it by the ones who find it easy. People with kids and jobs and lives. People who never had the option of spending five years studying with a professor at Princeton — because there was no family wealth to make a full-time scholar possible, because there were parents and grandparents and children to care for, because they were working sixteen-hour shifts and coming home to care for someone else.

Do these people start producing work that we can read, that we can argue with, that we can take seriously — and, when it holds up, that academia has to take seriously?

I don’t know the answer. Nobody knows the answer yet. It’s too early.

But I know one thing now, because of my own life, that I didn’t know last year. I know it’s possible. Because I did it.

IV. One Case

I want to be careful. I’m not claiming I’ve written a great paper, or that my paper proves anything. Or that I’m about to win — insert three or four prestigious awards here, ending with the Nobel. Sorry — for a second I was channeling my inner ChatGPT. The paper that came out of all this — the one I worked on with Claude as my advisor, the one I had Perplexity audit ten times, the one whose sections I sometimes rewrote ten or twelve times — is, at best, one case. One single case. Me. It documents what happened. It doesn’t prove that what happened to me would happen to other people.

But it proves something. It proves that the question I was carrying was a real question. It wasn’t a teenager’s romantic dream of a world where everyone could express themselves and be heard. It was an empirical question — and in at least one case, the answer wasn’t the no it started with. It was a maybe. Dare I say, a yes. Someone like me did do this.

The paper I ended up with is not the paper I started with. The paper I started with was about everything. The paper I ended with is about one thing: documenting what it looked like when one non-credentialed person used AI to participate in formal thinking, and figuring out what we’d need to study to find out whether other people can do it too.

That’s a more modest paper than the one I set out to write. It is also a paper I loved writing, a paper I’m proud to put my name on, and a paper I can defend.

V. What the Machines Got Wrong

A few things surprised me along my journey with AI.

The first was the kind of mistakes the AI made. I expected the dramatic ones — fake citations, made-up authors, invented data. What I got more often was something quieter. The AI would point me at a real source — Britannica, the World Bank, ScienceDirect — but wouldn’t tell me which article, which page, which study. Just the institutional name. When I went and checked, sometimes the claim held up. Sometimes the numbers didn’t quite match. And once, on a 2021 study about the internet and research productivity, the AI told me the paper had found a positive effect when the actual paper had found the opposite.

The advisory AI didn’t catch it. The auditing AI didn’t catch it. I caught it, because I went and read what the original paper actually said. The most dangerous failures I encountered weren’t the dramatic kind everyone warns about. They were the quiet ones — the ones that look like real research until you look closely.

The second thing that surprised me was that the AI I was working with most closely — Claude — wasn’t always doing the same job. Sometimes it was generating ideas; sometimes critiquing them; sometimes searching the web for real sources; sometimes editing. These are different jobs. My paper ended up naming four: generative drafting (producing text from my prompts), advisory critique (attacking my arguments), external validity audit (checking claims against outside sources), and research assistance (finding and summarizing existing work). The same system did all of them, but I had to keep track of which job it was doing in each conversation, because the system wasn’t always clear about the switch.

Once, during what I thought was an audit, Perplexity offered me three example citations to put in my paper. They were fabricated — labeled as examples, but written in academic format and ready to paste. If I’d been careless, I could have included them as if they were real findings. The system had quietly shifted from auditing to drafting. That kind of failure isn’t about the AI lying. It’s about the boundary between two different tasks blurring inside the same conversation. The user has to enforce that boundary, because the systems I used didn’t do it for me. And the stakes are not small. If I had submitted my work unchecked — the misread study, the ready-to-paste citations — I would have built my argument on things that were not true. That is the kind of mistake that ends a career in academia.

VI. Two Stories, Both True

There is a story about AI right now that says it will replace the people doing knowledge work. I won’t tell you that story is wrong. I will tell you it’s incomplete. The same technology that may replace some knowledge work is also unlocking the door for the people who were never let into the room. Both stories can be true at once.

The first one is getting most of the attention. The second one is what this essay is about...

Well...

I’m older now. I spent my career as a nurse. I’m not a professor and I never will be. My paper may ultimately be wrong.

But I wrote it. And I think it might be worth reading by the right people. I think the question it asks — does AI quietly recruit a new class of contributors into formal knowledge work — is the kind of question we should be answering with data, not with talk show segments. I think the failure patterns I documented are real, and not the ones most people are watching for. And I think the way I ended up describing AI mediation in four different roles, which honestly came out of me trying to keep track of what the machine was doing day to day, might be useful enough that other people will want to borrow it.

Those are modest, specific, testable contributions. They are the claims I can defend.

If the paper turns out to be wrong about all of it, then I will have documented one person’s case carefully enough that researchers will have something concrete to argue with. That isn’t nothing. Knowledge advances by people being specific enough to be wrong.

VII. The Person You Know

Here is what I want you to think about.

Somewhere in your life there is probably a person like me. A relative. An old coworker. A neighbor. Someone who carried a question for decades without having the tools to articulate it. Someone whose intelligence didn’t match the path their life took. Oh — and that person may be you.

Five years from now, if the work I’m part of turns out to be onto something real, that person might use these tools to write something substantial: not a tweet, not a comment, but a real argument — the kind of thing that gets cited.

We won’t know who they are yet. We don’t have a way to count them. They are, by definition, invisible.

But I was one of them. I’m not the first, and I won’t be the last. The question is whether more of us start showing up. Watch for the marks — an “L.I.” after a name (a Lost Innovator, working without the usual credentials), or an AIast line disclosing exactly which AI systems did what — small signals that someone used AI openly on the way to publication.

I think we will. I’ve been carrying that question since I was a teen and for the first time in my life it feels answerable. Answerable is not the same as answered.

The only way to find out is to look. And almost no one has looked at this question directly yet.

So: take it seriously. Study it. The next paper on this — the one that actually answers whether the lost innovators are now just innovators — won’t be mine. It might be yours.

I just wanted to write down that the question is worth asking.

© William Stafford · Published under the Earned Trust disclosure framework