AI has a discovery problem: users don't know what to ask
An essay titled "The Discovery Problem," dated 2026-09-03, argues that the biggest bottleneck to AI adoption is not what the technology can do but whether users know to ask for it. The core problem, in the author's words: you don't know what you don't know, and you don't know what a prompt can produce until you write it, hit go, and watch it run. As long as capabilities sit behind a blank text box, most of what a system can do stays invisible.
The essay walks through two partial fixes and rejects both as sufficient. Templates give people something to run without having to invent a request from scratch, but then the question becomes relevance: does a given template actually match the user's own work? Personalized context helps too, since a system that knows about the user can surface suggestions tailored to them rather than generic ones. Both help. Neither solves the underlying problem.
To explain the gap, the essay borrows a metaphor credited to Alan Kay: an ant at the bottom of the Grand Canyon looks up and sees only a thin sliver of blue between the canyon walls, while someone standing on the rim sees the whole sky. It's the same sky, but a completely different sense of what exists, and the ant is not less capable, it simply cannot see the axis of possibility from where it stands.
The essay applies this directly to AI use: take a non-technical marketing person next to someone fluent in agents and tool use. Watching the marketer work for an hour, the agent-fluent person can immediately spot a dozen things to automate, delegate or reinvent, including tasks the marketer has never even attempted. Put the most capable tool in the world in front of the marketer instead, and they face a blank prompt with no idea what to type. The intelligence is there; the visibility into it is not.
The author frames this as a design failure rather than a capability gap: systems today expect the user to already know what to ask for, when the work of discovering what's possible should fall on the system instead. The essay closes on that unresolved expectation, that an interface this advanced should reveal its own capabilities gradually and contextually, matched to a person's actual work, without proposing a concrete mechanism for how to build it.
Key facts
- Essay "The Discovery Problem," dated 2026-09-03, argues the biggest bottleneck to AI adoption is that users don't know what to ask a system for, not what the system can do.
- Templates and personalized context are named as partial fixes: both help surface relevant capabilities, but the essay states neither solves the discovery problem on its own.
- Uses Alan Kay's metaphor of an ant at the bottom of the Grand Canyon seeing only a sliver of sky, versus someone on the rim seeing the whole sky, to argue the gap is about visibility, not capability.
- Illustrates the gap with a non-technical marketer facing a blank prompt versus an agent-fluent person who can spot a dozen things to automate, delegate or reinvent after watching the marketer work for just an hour.
- Concludes that the burden of discovering what's possible should shift from the user to the system, which should reveal its own capabilities gradually and contextually, without proposing how that would work in practice.
Why it matters
The essay's claim reframes a familiar complaint, that people underuse AI tools, as a design problem rather than a training or capability problem. If the bottleneck really is discovery rather than ability, then explaining features or shipping more powerful models does not close the gap; what's missing is a way for the system itself to surface what it can do for a specific person's actual work, before that person knows to ask.
Who it affects
The essay contrasts two groups directly: non-technical users, exemplified by a marketing person confronted with a blank prompt, and power users fluent in agents and tool use, who can immediately spot automation opportunities by simply watching someone work. The gap described sits between these two groups, and by extension applies to anyone building the interface between users and an AI system, since the essay argues the responsibility for closing that gap belongs to the system's designers rather than the user.
How to use it
The essay offers no product or technique to adopt directly. Its practical takeaway is a design lens: judge an AI interface by whether it actively surfaces relevant capabilities to a specific user's context, rather than by how capable the underlying model is. Templates and personalization are named as the two current approaches worth combining, though the essay is explicit that combining them still falls short of solving the problem.
How solid is it
This is a personal essay, not a study: it cites no data, survey or research to support the claim that discovery is the single biggest adoption bottleneck, and its central example, a marketer versus an agent-fluent colleague, is an illustrative hypothetical rather than an observed case. The Alan Kay metaphor is presented as a memorable frame for the argument, not as independent evidence for it. The reasoning is coherent but rests entirely on the author's own observation.
Risks and caveats
The essay names no specific AI product, model or company, so its argument should be read as a general observation rather than a critique of any particular system. It proposes no concrete mechanism by which an interface would learn to reveal its own capabilities, and ends explicitly on that open question rather than a solution. The dozen-things estimate for what an agent-fluent person could spot is the author's own illustrative figure, not a measured result.
“It's not that the ant is less capable. It just can't see the axis of possibility from where it's standing.”
— the essay's author