The ring was never the bottleneck

A ring, worn by ordinary people recruited into an intergalactic police force, turns willpower into solid light. Whatever the wearer can picture hard enough, it builds.

That's the whole rule behind Green Lantern, and it holds up better than most comic-book premises: the power behind the ring is nearly unbounded, but what it can build in any given second is not. The ring turns intent into a physical construct, and it can only construct what the wearer can actually picture, right then.

Under pressure, that gap shows. A scared or untrained wearer builds a fist, or a wall, or a battering ram, because that is the most that fear or habit lets him picture in that second, not because the ring ran short on power. The ring had exactly as much to give as it ever does: the wearer just couldn't reach for anything better.

Hal Jordan, a test pilot, was the first human recruited into the Corps in 1959, the first to wear one of the Guardians' rings, though not the first person called Green Lantern. An older, unrelated lantern predates him by two decades, a different mechanic entirely, magic rather than willpower. This is worth saying plainly, so the point that follows doesn't rest on a mistake: the ring this essay is about, the willpower ring, is Jordan's, and it has passed through many hands since.

Over sixty years, the ring has passed to a test pilot, to an architect trained as a Marine, to a freelance artist, Hal Jordan, John Stewart and Kyle Rayner among them, and the range in what each built with it varied wildly. That gap was always the wearer, never the ring.

Give two people the same AI model, the same context, and ask them for the same kind of thing, and you'll get the same split: one comes back with something usable, the other with a fist, generic and forgettable, the safest and most average answer the training data can produce.

Nielsen Norman Group ran roughly this test directly, giving an AI prototyping tool the same brief twice, a profile page for a live-training product. A detailed prompt, one that specified layout, fields and behaviour, produced something close to what a trained designer would ship. A vague prompt, the kind most people actually type, produced something inconsistent, a different result each time you asked. Same tool, same brief, more or less: the gap was the ask.

This is the part that gets misread most often. A vague ask gets the AI equivalent of a fist, and the fist reads as a tool failure, blamed on the model when the failure happened upstream, in what was asked for. The model didn't run out of anything: it built exactly what it was given enough to picture.

That said, not every bad output works this way. Some failures really are the tool's, though most of what gets called a tool failure is this instead.

A ring fails on its own terms too. A battery that runs dry mid-fight leaves its wearer falling, and there's yellow: for decades, the ring simply couldn't affect anything coloured yellow, an old rule in the comics, well known, since retconned. This is worth stating plainly rather than softening it into something vaguer: it was a real, fixed limitation, and willpower didn't reach around it.

A model fails on its own terms too: it gets facts wrong, forgets an instruction from three messages back, or hits the edge of what it was trained on and produces something confident and false instead of admitting the edge exists.

In late 2022, Jake Moffatt asked Air Canada's website chatbot a plain, specific question: could he book now and claim a bereavement fare discount afterwards, following his grandmother's death? The chatbot said yes, and described a policy letting him apply for the discount within ninety days of travel, a policy that did not exist. Air Canada refused the claim, then argued before the Civil Resolution Tribunal of British Columbia that the chatbot was a separate legal entity, responsible for its own words. The tribunal disagreed. In February 2024 it ordered Air Canada to pay Moffatt CA$812.02.¹ Moffatt's question was not vague, he asked exactly what he needed to know, and the model failed anyway.

The honest disanalogy is worth naming rather than hiding: yellow is a fixed, catalogued rule, and once you know it, you plan around it forever. Most real model failure modes aren't like that, closer to unpredictable, invisible until they bite, harder to route around than a comic-book weakness.

There's a rough tell for telling the two apart in the moment: rephrase the same ask two or three different ways. If the model gets it right at least once, that was probably a fist, the first attempt didn't picture the target clearly enough, and a later one did. If it fails the same way no matter how the ask changes, that's closer to yellow, a limit inside the tool rather than the request.

Naming this here, not at the end, matters for a second reason beyond honesty. "The limit is you, not the tool" is a convenient thing to hear, and it happens to be the exact line AI companies use to deflect criticism of their own products. Saying that risk out loud is what keeps the argument in this essay from becoming apologetics with a comic-book paint job.

Set the exceptions aside. Most of the range between a fist and something worth keeping comes down to three things, and none of them are about the tool.

Specificity. Describing the actual shape wanted, not the category of thing. "A wall" is a fist wearing a longer name. "A wall curved to deflect, six feet high, holding for thirty seconds while I get behind it" is a construct.

Taste. Knowing what good looks like well enough to reject the first plausible answer. The first output a model gives you is rarely wrong exactly, it's just the average of everything like it that's ever been written. Recognising that it's average is its own skill.

Reid Hoffman got early access to GPT-4 before its public release and used it to write a book, working through hundreds of exchanges: draft a section, push back, ask for a different angle, redirect it. Impromptu², published in March 2023, is the record of doing that, not just the result of it. What made it worth reading was never the model's raw ability to produce paragraphs, it was Hoffman knowing what he wanted well enough to reject the drafts of what he didn't, over and over, until one held.

Nobody sees the discarded drafts, and that's true of the ring too. Nobody remembers the fists that got thrown away before the construct that actually worked.

There's a further move past specificity and taste, and it's where abstraction earns its keep, or costs you, depending on how carefully it's done.

A wearer who has really trained stops picturing one construct and starts picturing a kind of construct, not a wall but a shape that can be redrawn on demand, a small grammar instead of one object.

Kyle Rayner, the freelance artist who inherited the ring in the 1990s, is the clearest case of this in the comics themselves. His constructs don't just get bigger the way a soldier's might under more training, they adapt: a shield that reshapes mid-hit, a construct that keeps doing something after he stops actively picturing every detail of it. That's not more power, it's a different level of instruction, a construct built to behave like a small system rather than sit there as one object.

The same move shows up with AI: not one prompt for one output, but a reusable structure, a way of prompting that produces many good outputs instead of coaxing out a single one. A checklist instead of one answer, a process instead of one draft.

It needs care, though, and the care is the whole point of raising it here. Abstraction done badly is the fist reflex again, wearing a nicer coat: it sounds sophisticated, but it produces nothing specific. The tell is the same one from the honesty section, run in reverse: does the abstraction still cash out into something concrete when tested, or does it just float.