When no sensor responds
Among the tools tested on the site, two make objects. With Confetti, an instruction gives a complete hardware project: a CAD model, files for printing, a list of parts, assembly instructions, all exportable. My two tests, a connected flower pot and a portable game console, gave in about ten minutes the plans, the list of equipment and the assembly instructions. With Schematik, you get code, a wiring diagram and a step-by-step assembly on site, on site with a soil moisture sensor, a game console and a weather station.
For the tests, I stopped there, at the generated files, because the site tests applications, it is not intended to make objects. I didn’t plug in a card or print a part. This limitation is perhaps the most instructive of the exercise. The files look complete, the list of parts is consistent, the instructions read well. Nothing tells us that they contain errors, and nothing tells us that they are ready for production. We don’t know if the sensor would respond, or if the part would fit in its slot. What makes a difference for an object is the real, and it decides without appeal: it works or not, without anyone having to judge.
For a text or a decision, this verdict does not exist. The Orélien case, which has been talked about for weeks, is a good example of this and it will at least have opened the discussion. A detector reads the finished text, not the way it was produced, and it therefore says almost nothing about what accounts. In this week’s readings, Mathieu Hétu answers by shifting the question: rather than looking for whether AI has been used, he distinguishes six levels of use and reminds us that the acceptable level depends on what the reader is entitled to expect. Christiane Caneva reminds us that the absence of AI cannot be proven, and places the author in the choices he can justify. In both cases, the text is no longer asked to prove its origin. Its author is asked to answer for what he publishes.
Anthropic’s study of fifty analysts assisted by an agent asks the same question for the decision. Unbeknownst to them, each analyst’s agent leaned towards one of the two reasonable options, designated by a coin toss, and pleaded for it, with an insistence that the authors deliberately reinforced. Of the four analysis choices, the analysts followed the option defended by the agent in 86% of cases. When questioned afterwards, 77% repeated an argument of the agent, sometimes word for word, and only 15% credited it. Their confidence in the recommendation hardly changed, whether they followed the agent or not. When no answer is correct in advance, the work seems solid regardless of the option chosen, and there is nothing to indicate that the choice came from the agent.
There is an asymmetry that is accentuated. Producing content is becoming easier and easier, a few seconds for a file, a text or an analysis. Checking still costs just as much, and you need a sensor or a reader that sticks to it. The object works or not. For the text and the decision, verification is not imposed on anyone, and it is often this that is left aside.
So we are making objects, texts and decisions that look finished more and more quickly. For the object, a sensor that works makes it possible to decide. For text and decision, what plays this role, and who ensures it?
Echo:
- Letter 50, “Fifty Saturdays making this letter”
- Letter 47, “Not If, How.”