AI: A Wider Lens
How AI turns one question into a systems view
A small business owner asks the model whether she should open a second café. The answer comes back with the cash-flow math, yes. But also with the lease timing, the staffing lead time, the risk that two locations split her customers instead of doubling them, two cafés in similar towns that closed within a year, and the question of who runs the first shop while she opens the second.
She asked about expansion. The model answered about her business.
Ask AI a narrow question and you often get a wider answer. That is not a quirk. It is how the model is built, and it is the closest thing modern technology gives us to systems thinking on demand.
Post 3 of a series on AI: concepts, systems, and how we think with the new tool.
What the model actually does
Large language models are pattern recognizers. They were trained by reading a vast amount of human writing and learning, statistically, what tends to appear near what. Ask the model about caffeine and it has a strong sense of what travels with caffeine in the texts it has seen: coffee, sleep, anxiety, withdrawal, cardiovascular research, the global trade in coffee beans, the price of arabica futures.
A human mind associates too, but only from what one person has read in one lifetime. The model has read at a scale no person can match. When you give it a question, it pulls from a much wider conceptual neighborhood than any single expert carries in their head. Same move, vastly more ground.
The associations are not random. They reflect how the world connects, because human writing already encodes those connections. Doctors write about medications and side effects in the same papers. Farmers write about rotation and crop insurance in the same forums. Engineers write about loads and permits in the same reports. The links the model surfaces are the links humans have already drawn across the systems they live in.
The model retrieves the conceptual neighborhood around your question. The neighborhood, by virtue of being human, is shaped like a system.
A question, three lives
The widening shows up across very different fields.
A patient newly diagnosed with type 2 diabetes asks about a medication her doctor mentioned. The answer arrives with the drug’s mechanism, yes. But also with the diet and exercise changes that do more than any pill to protect her eyes and kidneys over the next twenty years, the interaction with the blood-pressure tablet she is already taking, the cost difference between brand and generic, and the routine checks the diagnosis now implies. She asked about a pill. The model answered about a chronic condition.
A farmer asks whether to switch a field from corn to a cover crop. The answer covers soil chemistry, yes. But also crop insurance rules that change with the rotation, regional water-rights implications, equipment that may need to be sold or bought, market timing for the new crop, and three growers in nearby counties who tried the same switch. The farmer asked about a field. The model answered about a business.
A school principal asks how to integrate AI into next year’s curriculum. The obvious surfaces first: lesson plans, assessment, what to do about students who use AI to write their essays. But the model keeps going. Those essays were designed the old way. A 500-word piece on a favorite animal once told you something about a student’s effort; a model can produce one in seconds. The deeper question is what to assign instead: open-ended projects where students use AI to explore a hard problem, ask sharper questions, widen their view, assess what the model tells them, and decide what to ask next. The skills the new curriculum has to teach are not the old skills with AI bolted on. They are critical thinking and systems thinking. The principal asked about a tool. The model answered about how teaching has to change.
Familiar insight, field-wide insight
Before AI, the natural way to research a decision was to ask a specialist. A specialist gives you a deep answer inside a narrow frame. That is valuable, but it has a known cost: you get the answer inside the question you knew to ask.
What you did not know to ask stays invisible.
AI changes that default. The model has no single frame. It has the whole conceptual neighborhood, and it surfaces what travels with your question whether or not you thought to ask. The result is not always deeper than the specialist. It is wider. You get the familiar insight a domain expert would give, and you also get what the expert usually cannot: the connections that cross the boundaries that organize human knowledge but never organized the problems people actually face.
It is hard to overstate what this does for the quality of a decision. Most choices go wrong not because the obvious answer was wrong, but because something off to the side was never considered: the regulatory consequence of an engineering choice, the family consequence of a career move, the customer consequence of a financial one. A wider view brings those into the open before you commit. The decision you make is then better informed, more considerate of the people it affects, and more honest about the complexity and the risk it carries. Decisions made that way tend to hold up. You are seeing the road, not just the next step.
The honest counterpoint
Width is not the same as judgment. A model can list every consideration around your question without weighing any of them correctly. The associations it surfaces are statistical, not causal: it shows what tends to appear near your question, not what actually matters. Critics like Gary Marcus argue that this kind of fluency is a confident map of a landscape the system has never directly seen. In Rebooting AI and in years of essays since, he makes the case that a system trained on text statistics has no working model of the world beneath the words, so its confidence is no evidence that it is right.
The width is useful precisely when you bring the judgment to it. The model can surface what bears on a choice, but it cannot tell you which of those things matters most in your situation, and it does not carry the consequences when it gets it wrong. You do. This is the line at the center of the public argument about AI: not whether the tools are powerful, but how much of our own thinking we hand to them. Treat the model as a research assistant that widens your field of view, and it makes you sharper. Treat it as the decision-maker, and you have automated away the one part of the work that was meant to stay yours. The tool can draw the map. Reading it is still your job.
Questions for the reader
Pick a question you would normally bring to a specialist: a medical question, a legal one, a business decision, a parenting question. Ask the model. Then notice what it surfaced beyond the narrow answer. What did it bring into view that you would not have thought to ask?
A note on timing. AI is changing fast enough that some of what is described here will read differently in six or twelve months. That is the nature of the subject, not a flaw in the snapshot.
Forward pointer: Next week, hallucinations, and how to work with a fluent collaborator that sometimes makes things up.
Now this:
Further reading
• Bender & Koller, Climbing Towards NLU: On Meaning, Form, and Understanding in the Age of Data — the skeptical voice: language models do not understand meaning.
• Gary Marcus and Ernest Davis, Rebooting AI: Building Artificial Intelligence We Can Trust — the case that today’s systems have no working model of the world beneath the words they produce. See also his ongoing essays at Marcus on AI.
• Yejin Choi, “Why AI is incredibly smart and shockingly stupid” (TED, 2023).
• Stuart Russell, Human Compatible: on intelligence as breadth.
***
"Systems thinking on demand"--Great concept, nicely explained in terms of LLM fundamentals, and pointing to one of the most valuable aspects of the tools: Providing insight into related questions you did not think to ask. Then why that matters: "Most choices go wrong not because the obvious answer was wrong, but because something off to the side was never considered." Also clarifies the role of human judgment (something the narrow expert still brings to the table). Excellent flow. Thank you.
Great point about AI changing how one teaches. Many school districts and educators are asking to restrict student access at various grade levels. It occurs to me that calls for restricting AI acess to students, especially at young ages, is less about what benefits the students and more about controlling resources as well as not knowing enough about being a teacher to understand how to change.