I'll bet that Alexander Fleming wasn't thrilled to find mold growing in one of his laboratory dishes in 1928. Contamination generally isn't something scientists want or even celebrate.
But Fleming did notice something unusual. And I think that word is important. The bacteria surrounding the mold weren't growing. What he noticed helped give the world penicillin and transformed medicine. The zone of inhibition became a zone of discovery.
Fleming wasn't looking for penicillin; he was noticing something—something that wasn't supposed to be there at all. I wonder how many important scientific discoveries over time have emerged from that ability to see something new in what first appears to be a mistake or an irrelevant artifact.
About a hundred years earlier, Louis Pasteur offered an observation that nicely aligns with Fleming's observation: "In the fields of observation, chance favors only the prepared mind." Clearly, Pasteur wasn't just celebrating luck; he was describing the dynamic between an observation and the ability to recognize the significance.
Today, artificial intelligence is adding an interesting twist to what is observed and interpreted. AI is becoming very good at finding answers, but as we move more "frictionlessly" from question to answer, I wonder what happens to the unexpected bumps we encounter (and might miss) along the way.
And there's another "evolving possibility" that is worth putting under the microscope: Could AI actually manufacture the conditions for serendipity?
The Things We Weren't Looking For
Let's jump back another one hundred years. The word serendipity was coined by Horace Walpole in 1754, based on a Persian tale called "The Three Princes of Serendip." In the story, the princes made discoveries through a combination of both accident and insight.
Serendipity isn't simply about finding something surprising—it's about finding something valuable that wasn't part of the plan.
I've written about "productive" cognitive friction and how the struggle (even confusion) can align with scientific discovery. The journey toward an answer can produce questions that are more interesting than the original problem. AI's efficiency may allow us to bypass some of those detours, along with the unexpected observations they occasionally reveal.
Now, back to Fleming. His mold wasn't an unexpected connection buried in a pile of data. It was an unexpected event living in our physical world. No one had asked for it, and it wasn't part of an academically approved study design. AI can uncover surprising relationships across wide and varied data sets, including connections a researcher might not recognize. But there's something different about an observation that arrives as an unexpected knock on the lab door, particularly when it challenges what we thought we knew.
The Prepared Machine?
After looking back a few hundred years, let's look at today.
A 2024 study in Science Advances found that generative AI helped individuals write stories that were judged to be more creative. So far so good. But creativity seemed to come at the expense of originality as the AI-assisted stories also became more similar. It's important to note that the study looked at creative writing rather than science, yet it might raise a related question. Could AI help each of us produce better ideas while potentially narrowing the diversity of ideas we explore?
A 2026 preprint from Princeton and the University of California, Berkeley, offers more recent perspective. The study explored if deliberately introducing ideas from unrelated topics could stimulate creativity in both humans and LLMs.
The results showed that humans consistently benefited from unexpected associations. But AI, which generally produced more original ideas in the study, showed no overall improvement from the added topics. It's my contention that AI's vast corpus of knowledge may not "incrementally" advance the "connectivity" as much of the information may already be contained in the training data.
So, here's my question: Can we deliberately engineer serendipity? And if we build AI to identify unexpected relationships, are there discoveries really serendipitous accidents? Or have we developed a more sophisticated method of searching?
I've written about a version of this idea in medicine with what I called an AI Zebra Scan. Rather than allowing AI to settle on the "most statistically probable diagnosis," why not design it to deliberately explore less likely possibilities? This isn't serendipity itself, but maybe it's a way to expand the range of possibilities that AI considers. And perhaps that same principle can be applied more broadly to scientific discovery.
- Expand the hypothesis space. Explore possibilities beyond what's most probable.
- Preserve anomalies. Investigate outlying observations rather than dismissing them as statistical noise.
- Connect distant disciplines. Identify relationships outside established boundaries.
- Challenge the original question. Consider if the assumptions themselves are limiting discovery.
Of course, these aren’t definitive rules for producing scientific breakthroughs. And they’re probably easier said than done. But there might be a few thought starters here for designing AI that could be receptive to the unexpected rather than simply optimizing the expected.
The Observer and the Observation
In the final analysis, there's a difference between discovering an unexpected relationship among things we know and the discovery of something that makes us to reconsider what we thought was even possible. Both can lead to discoveries, but they're not the same experience.
AI can do the first at scale. It can also be designed to notice anomalies and judge their significance. Once we deliberately create the conditions where the unexpected will be found and recognized, the accident is no longer accidental. In that sense, programming serendipity is a bit like Schrödinger’s cat—the act of looking for it changes what we find.
Pasteur’s "prepared mind" may be entering unfamiliar territory where AI increasingly participates in both the fortunate encounter and the judgment that turns it into a discovery. What remains may still be valuable. But is it really serendipity?