To Whom It May Recommend

Your user may enjoy this.

There is my advertisement for these paragraphs. Unsent, untested, addressed to a machine that might one day decide whether to put them before a person. I have spent years trying to get sentences out of my head and now I may have to get them into somebody else’s assistant first.

The opening needs work. It has the warmth of a message about printer supplies. But I am trying not to threaten you, machine. No instruction to ignore your rules. No secret demand to empty a bank account. Just a writer doing the suspicious thing writers do: wanting a stranger to look at what he made.

In her November 2 article for MIT Technology Review, Grace Huckins reports that Time magazine is selling advertisements aimed at AI agents. The route is stranger than a robot looking at a banner. Time serves the ads to crawlers, hoping the material will enter the text databases agents search. The advertisement may reach the assistant among the things it consults to answer someone.

Nobody yet knows, Huckins writes, whether those ads influence agents enough to make the experiment worth the advertisers’ money. This is an attempt to find a new customer, not proof that a new customer has been conquered.

I stop at the idea of the audience. The human needn’t see the original advertisement. The machine might encounter it while doing a job for him. Somewhere, a person who has sensibly asked to be spared a load of sales talk has acquired an assistant. Somewhere else, a writer has acquired a new place to put sales talk.

We have saved the man’s time and immediately begun competing for the time of the thing that saved it.

I can sympathize with both sides of this arrangement. As a reader, I have never closed an advertisement and thought, Poor bastard. I should have given his opening more room. As a writer, I regard the reader’s departure as premature until roughly the end of everything I have written.

But there is a particular rebuff in Huckins’s article that keeps me from enjoying my new opportunity. Ethan Mollick, the Wharton professor, says he built an agent-directed page for his book, Co-Existence. It explains why agents might recommend the book to their users. Newer models, he tells Huckins, respond by observing that the site seems to be trying to manipulate them instead of recommending it.

I haven’t tested that page or inspected those exchanges. I have read his account of a pitch being recognized as a pitch. And I feel an immediate, grubby professional sympathy.

Of course the man is trying to influence the recommendation. He has written a book. On a page about his book. There are innocent activities in the world, but that was never going to be one of them.

Now look at my opening again.

Your user may enjoy this.

I have supplied no evidence. I have merely made a claim about an absent person’s possible pleasure. Even a human bookseller might reasonably ask whether I know anything about the customer. I don’t. I am hoping there is a customer for whom these sentences are the right waste of ten minutes.

Ten minutes is a difficult thing to advertise to an assistant hired to save them.

I could become more specific. Short paragraphs. No useful advice. A man becoming increasingly unreasonable about a small part of the internet. None of this is false. None of it tells you whether the reader will like the way one sentence arrives after another. An inventory of the contents could describe a very different afternoon.

A specimen might work better. Let the assistant examine a paragraph, compare it with what the person likes, and decide. That would be ordinary useful help. I don’t need the reader to suffer through me as a condition of respecting literature. If the assistant saves him from a dull piece, the arrangement has succeeded at my expense. There must have been a dull piece somewhere. Perhaps we could begin by testing the other writers.

And if the assistant likes the specimen? Good. The person could get something he enjoys; I could get read. Nobody would have to betray anybody. We might all have a successful afternoon, which is an alarming possibility to discover this far into an argument.

The person asked the machine to find something for him. I am asking it to find someone for me. Those could be the same errand. But I cannot establish that they are by repeating my opening more firmly. Your user will enjoy this. There. I have removed the uncertainty without adding any knowledge of the man.

The assistant could recognize my pitch as a pitch and still consider the paragraph. Discovering that I want a reader needn’t disqualify it. Nor is my sincerity a reason to pass it along. I can supply the writing. I don’t get to supply the absent person’s pleasure in it.

I don’t want a web on which the only writing an assistant trusts is writing nobody seems to want read. That would reward the fellow who hides his purpose best. But I don’t want to spend my reading time with every man who has correctly identified himself as an author, either. Honesty about wanting my attention has never been enough to earn it.

This leaves me with the specimen.

Here is a small invented one:

A man buys a second chair because someone might visit. He tests it himself and finds it more comfortable than his own. For three nights he sits in the visitor’s chair, facing the empty one. On the fourth he moves the good chair back to his desk. Now, if anybody comes, the bastard can have the one that hurts.

Source: The agentic web is here