Focus of the letter 42

This week, I brought together two seemingly distant tools.

Drafted generates a house plan from a list of rooms and style, with a library of over 85,000 pre-produced, searchable, and remixable plans.

Swishy promises to go fast and succeeds. You type an instruction, you pick from a gallery of more than a hundred templates, and the motion design video comes out in a few seconds, animated typography, clean transitions and correct rendering. In this week’s tests, the result holds up.

In both cases, the promise is the same: made-to-measure in appearance, ready-to-wear in the mechanics. You choose from a catalog, you personalize a few details, and the result is out in a few seconds. But, as with fast fashion, by imagining it in hundreds of different hands, a question quickly arises: what will distinguish the house from one, the video from the other, if everyone starts from the same pattern?

This is exactly what Noémie Petit humorously denounces in a video I saw this week on Instagram about AI-generated event posters. Too much information crammed into a single image, all the bars, restaurants, associations, concert halls, etc. that end up producing the same poster, to the point that we no longer know who is proposing what, when and where. Her rant hits the nail on the head, and it hits three main points for me.

First of all, a question of knowledge, that of the implicit codes of good communication, which we do not acquire by formulating a prompt but by learning, watching, failing and starting again.

Then the necessary skills: visual communication requires skills, that’s why it’s even a job… with its rules of hierarchy of information, readability, what to keep and what to remove. A chatbot, if the instruction does not specify it, does not know that an effective poster is a poster that removes rather than adds.

And then, once again in this letter, a question of delegation: the machine is entrusted not only with the execution but with the choice, and the default choice of a model trained on millions of posters is the average, never the singularity. We also find a similar mechanism in the University of Maryland study shared by Myrtille Gardet on the “argumentative collapse”: when questioned separately on the same subject, language models converge on the same arguments, whereas human arguments are plural. What affects ideas seems to affect communication media as well…

But for me, none of this condemns the tool. Communicating with generative AI is possible, and it can even work very well, as long as it’s worked on: iterated, cropped, contradicted, retouched by hand until the result says something specific rather than everything at the same time. The difference is not between human and machine, it is between a generation accepted as it is and a generation that is led to the end.

What is lost, when you skip this step, is not a question of technical quality. The renderings are clean, the colors are right, the text is legible. What is lost is the trace of a decision. And a poster, a visual, a communication video, it has never been just a text or colors: it is first and foremost a decision, assumed and artistic choices, and that can be recognized among others.

And behind these decisions, there are often jobs that we tend to forget in this content generated in two clicks: that of illustrator, graphic designer, videographer, whose know-how is precisely that we bypass every time we let the machine choose for us.

Echoing

 →In letter 34, we focused on what looks like local information via the retrieval of content and which, by dint of industrialization, ends up replacing it.