The Cocklebur Principle

When people show me AI output they’re unhappy with, the writing is almost always bland in the same way. Technically fine. Says nothing only that writer could have said. They blame the tool. The tool is rarely the problem.

AI is trained on actual human writing. It’s learned not just how people write, but how they write about specific things in specific contexts. And as a predictive engine, it regresses to the mean. The simpler the input, the more average the output. Feed it a plain request and it hands back the most likely, most expected version of the thing. Which is another way of saying the most average version.

As a creative professional, I understand creativity depends on two things: constraints and inputs. You have to define the sandbox, and you have to provide the toys to play with inside it. With AI, the more toys you give it, the more unique the output becomes.

Most people only bring the expected inputs to the sandbox, which is why so much writing feels a little flat. The most creative people bring what Edward de Bono called “random inputs.” Things that don’t belong. Things that act as springboards to unexpected, unique outcomes.

Most people trying to improve a fastener focus on a better tape, a better buckle, a better hook. Then one day a guy notices how securely a cocklebur is stuck to his sock, and invents velcro. The breakthrough didn’t come from the fastener. It came from the thing that didn’t belong.

The problem with AI output is almost never the AI. It’s the inputs.

AI works the same way. When you throw a few random inputs in with all your standard ones, it skews the output in directions that feel unpredictable from the outside. So rather than asking it to write a post about the impact of benefits on workers, I’d ask for a post about health insurance that workers can actually afford to use. I’d give it the benefits list. I’d give it real quotes from people describing what that impact looked like in their own lives. I’d give it constraints: length, things it couldn’t say. I’d try it in a few different tones.

The bland version and the good version came from the same tool. The difference was everything I brought to the sandbox before I asked.

If your AI output feels generic, don’t go looking for a better tool. Look at what you handed it. The average input is the whole problem, and it’s the one part of this you fully control.