A former colleague sent me a screen recording last week. He typed one sentence about a coffee subscription into one of the new GPT-3 copy tools and it handed back forty headlines in about nine seconds. Most were filler. Four were usable. Two were better than anything the brief had produced in a week of trying.
He wanted to know whether he should be worried. I watched it three times at my kitchen table with the April rain doing its usual work on the window.
Then I went back and read the brief.
The brief was the problem. It said the subscription was for people who love good coffee, which describes everyone who has ever bought coffee, and no quantity of generated headlines repairs an audience defined that loosely. The tool did not know that. The tool cannot know that.
These systems are extraordinary at producing the artifact and useless at deciding whether the artifact deserved to exist, which means they automate the cheap half of the job and leave the expensive half exactly where it has always been.
The expensive half is three decisions, and none of them are writing. Who is this actually for, stated narrowly enough that somebody is excluded. What is the company genuinely able to deliver, as opposed to what the deck says. What are we willing to be disliked for. Oatly built a whole business on that third one, printing arguments on the side of a carton that a committee would have sanded flat. No generator produces that, because a generator has no stake in being disagreed with.
Here's the part that gets lost in the demos. Forty options is not abundance. Forty options is a decision problem, and the person who has to choose still needs a point of view about the market, the category, and the customer's actual Tuesday. Give a weak strategist forty headlines and you get a slower version of the same weak campaign. Give a good one four, and it ships.
I'm not precious about the tooling. I've sat in agency rooms where three days went to headline variants that a machine could now generate over lunch, and I would happily take those days back for positioning work. The risk is not that the writing gets automated. The risk is that generation gets so cheap that nobody notices the thinking never happened.
Forty options is not abundance.
I wrote him back one line: rewrite the brief, then run it again. He did. The tool returned forty new headlines, and this time eleven were usable, because the input finally said something. The machine got better because the human did. That is the whole relationship, and I suspect it stays that way longer than the demos suggest.