There is a moment most of us have lived through without naming it.
An online meeting, six people from different remote locations, none of them native English speakers. Someone is explaining a process, walking through how a feature works. On screen, a diagram tells a slightly different story. The flow described in words and the flow drawn in boxes do not quite match. For half a second, something registers. A contradiction sits in plain sight.
Then the speaker says something like “we leverage the existing framework.” The room nods. The phrase is vague enough to cover the gap. Nobody asks which framework, or why the diagram suggests a different path, because the shared language of the room does not reward that kind of precision. It rewards movement. Next slide. Next topic. The meeting ends on time, perhaps even earlier. The contradiction, and whatever it might have revealed, is not even a memory.
Most of us have lived that moment. We have watched a catch-all term close a door that was briefly, quietly, open.
What was left behind that door is a good deal of uncertainty, failed communication, and misunderstanding. Not the dramatic kind. The quiet kind that compounds. These small failures create patterns, and patterns, once visible, can be fixed. But seeing them requires noticing what we were not looking for, connecting signals across conversations that were never meant to be read together.
There is a word for that process: serendipity.
And the tools to do it at scale exist now. Almost nobody is using them this way.
The word we misunderstand
We tend to think of serendipity as luck. A happy accident. The stories are familiar: Alexander Fleming notices mold killing bacteria in a petri dish.
A 3M engineer’s failed adhesive becomes the Post-it Note.
Percy Spencer feels a chocolate bar melt in his pocket near a magnetron and invents the microwave oven.
The stories are true. The framing is wrong.
A systematic review published in the Journal of Management Studies in 2024 identified three conditions that separate serendipity from plain luck: agency, surprise, and value.¹ Serendipity is not something that happens to people. It is something people do with what happens to them. The surprise must be noticed, interpreted, and acted upon. Each step is active. Each step requires preparation.
Louis Pasteur put it in a single sentence in 1854: “In the fields of observation, chance favors only the prepared mind.“
Fleming did not just see mold. He had the bacteriological vocabulary and the conceptual framework to understand what the mold was doing. Without that preparation, the petri dish gets thrown away. The accident produces nothing.
Here is the frame that matters: serendipity is not an event. It is a capacity.
And that capacity has prerequisites, linguistic, cognitive, and organizational, that we seem to be systematically undermining.
The thinning
The opening scene is not anecdotal. The data confirms the pattern. A study tracking nearly 30,000 American adults from 1974 to 2016 found that vocabulary declined across every level of education, with college graduates dropping from 81 to 71 percent correct on a standardized test over four decades.²
In European multinationals, research on Business English as a Lingua Franca describes the shared professional code as “simplified, hybridized,” and optimized for task completion over precision.³ The functional minimum becomes the practical ceiling.
When habitual language narrows, habitual attention narrows with it.
Each of serendipity’s three steps, noticing, framing, acting, depends on distinctions that a shrinking working vocabulary quietly deletes.
We cannot see what we are losing, because nobody counts the discoveries that were never made.
How we use the tools
The arrival of large language models should, in principle, reverse this trend. And not by accident.
LLMs are architecturally suited for serendipity. They represent concepts as vectors in high-dimensional space, where ideas that are semantically adjacent but topically distant can sit close together even though a human mind would never connect them.
The model traverses conceptual neighborhoods that our categories keep separate. That structural property is precisely what serendipity needs: unexpected but meaningful connections across boundaries we did not know we had drawn.
In practice, most people use LLMs the way they use a search engine with better manners. Ask a question. Get an answer. Move on.
The interaction is efficient. It is also the cognitive equivalent of always reaching for the same twenty words.
The tool confirms what the person already expected. Nothing unexpected survives the exchange.
The contrast is visible to anyone who has tried both modes. Research presented at ICTIR 2024 found that indirect prompting, decomposing a question into sub-questions the user would not have thought to ask, significantly outperformed direct prompting when the goal was serendipitous discovery.⁴ The finding is intuitive once stated. A question already contains its own assumptions. Breaking it apart exposes them and opens space for the unexpected.
In practice, this looks like asking for ten competing hypotheses instead of one answer. Forcing cross-domain connections. Working in two passes, first divergence, then critique. But the real shift is systemic, not per-session.
I have built a deliberate workflow around this.
Every one of my LLM sessions ends with a structured handoff: a summary of what was explored, what was decided, and what remains open.
Those handoffs accumulate and periodically, I feed them back into a session with a different instruction set, what amounts to a dedicated system prompt designed not to continue the work but to read it sideways.
The prompt asks the model to surface contradictions across sessions, flag patterns I did not intend, and connect threads from unrelated projects.
The model is not generating new ideas. It is re-reading my own thinking through a lens I did not hold while doing the original work.
This is the “second reader” principle. The serendipity does not come from the machine inventing something. It comes from the machine noticing what was already there, across sessions that were never meant to connect.
The accumulated handoffs become a personal knowledge base richer than any single session could produce, and the system prompt becomes the cognitive lens that determines what the model pays attention to.
The results are consistent enough to make the pattern clear: the tool amplifies whatever the person brings to it. A narrow prompt produces a narrow answer. A flattened vocabulary gets a flattened response. The machine does not compensate for a shrinking range of questions. It mirrors it. And when it is given range, structure, and a deliberate angle of re-reading, it surfaces connections that no direct search would have produced.
How organizations choose
The same fork exists at the enterprise level, and it is being decided right now in most organizations, usually without anyone realizing it is a choice at all.
The default deployment of AI in companies is optimization. Recommendation engines trained on accuracy converge on the familiar. Internal knowledge tools surface what was already expected. Reporting systems confirm existing patterns faster. Every deployment is measured by how well it predicts what was already known. Efficient, functional, and blind to everything it leaves out.
The alternative uses the same technology with a different intent. Recommendation systems that deliberately surface the unexpected alongside the relevant, what the research literature calls the accuracy-serendipity tradeoff.⁵
Internal AI tools that retrieve content related in meaning but coming from completely different topics, not just keyword matches. Knowledge systems that connect projects across departments that did not know they were working on related problems, or that flag contradictions between what the data shows and what the quarterly narrative claims. Not chaos. Structured divergence.
Most companies are choosing the first path. The cost of that choice never appears on any report. No KPI measures the product idea that was never surfaced, the market signal that was never connected, the innovation that never crossed a departmental boundary. And, in fairness, the optimization metrics are, by every available measure, performing exceptionally well…
What we are not counting
Serendipity is not a luxury. It is not a personality trait. It is the mechanism by which individuals and organizations encounter what they did not know they needed.
The research is clear on this. The prerequisites are identifiable: linguistic range to notice and name the unexpected, cognitive habits that accept ambiguity instead of ignoring it, and organizational structures that leave room for surprise in systems built for efficiency.
We are eroding those prerequisites on almost every front. In the words we practice, in the questions we ask, in how we deploy the most powerful tools for connection and discovery ever built.
The uncomfortable truth is not that we lack the technology to make unexpected discoveries. It is that we are using that technology, systematically and at scale, to make fewer of them. The losses are invisible by design. We will never see the insights that were not had. We will never see the discoveries we failed to make. We will never know the connections we missed. We will only notice, slowly, that the world we are building feels less surprising than it should.
So, the question is not whether we can afford to invest in serendipity.
It is whether we can see what we are already losing by ignoring it.
References
- Busch, C. (2024). “Towards a Theory of Serendipity: A Systematic Review and Conceptualization.” Journal of Management Studies, 61(3), 1110–1151. https://onlinelibrary.wiley.com/doi/abs/10.1111/joms.12890
- Twenge, J. M., Campbell, W. K., & Sherman, R. A. (2019). “Declines in Vocabulary Among American Adults Within Levels of Educational Attainment, 1974–2016.” Intelligence, 76, 101377. https://www.sciencedirect.com/science/article/abs/pii/S0160289618302198
- Kankaanranta, A. & Planken, B. (2010). “BELF Competence as Business Knowledge of Internationally Operating Business Professionals.” Journal of Business Communication, 47(4), 380–407.
- Fu, Z. & Niu, X. (2024). “The Art of Asking: Prompting Large Language Models for Serendipity Recommendations.” In Proceedings of the 2024 ACM SIGIR International Conference on Theory of Information Retrieval (ICTIR ’24). https://dl.acm.org/doi/10.1145/3664190.3672521
- See, e.g., Kotkov, D., Wang, S., & Veijalainen, J. (2016). “A Survey of Serendipity in Recommender Systems.” Knowledge-Based Systems, 111, 180–192.
