A Sloppery Slope
(This post has also been published as a German article on LinkedIn and has been cross-posted here for conservation and accessibility. It has been auto-translated from German and human-reviewed.)
The dominant question of the last few years was “What can AI do?” The answer has become boring: almost everything, often well enough. The question we have to ask now is shifting from “can we?” to “may we?” And that question probably no longer has a technical answer.
There are workflows where the result is the value, and workflows where the human effort itself is the value. For the first kind, AI can become a very powerful tool. For the second, it can be just as powerful at causing harm. When human effort is the actual value, the work becomes worthless the moment an AI does it.
A good example is feedback of any kind. That can be a code review, but especially the interpersonal kind. AI can produce a lot of it and fast - but the value was never in the volume. It sits in the time someone personally spent on honest reflection. Companies with a healthy error and feedback culture get a lot right. Overeager use of AI can tear down much of it again.
The moment effort can be faked invisibly, every personal message carries a question mark: from you - or from your AI? The uncomfortable part isn’t that I doubt the content. I doubt something else: Am I worth your time? Trust doesn’t erode slowly. It often collapses instantly. “I only do it sometimes” is no defense; it’s enough that I know it happened once.
This doesn’t apply to everything of course. If an AI drafts the scheduling email, I couldn’t care less - attention was never the currency there. But it applies to a whole class: feedback, thanks, recognition. Everywhere the real message was “I was thinking of you.”
As software developers, product people, and company builders, our reflex is often to solve this kind of problem technologically. Detectors, watermarks, “AI-generated” badges, and so on. Whatever half the industry is building right now.
My own, very personal takeaway, after a few years in this industry: social problems can’t be solved technically. Technically rooted ones, though, can be contained socially.
No algorithm restores what breaks here. The whole arms race of detectors and badges answers a social question with technical means. A category error.
What’s left is less comfortable, because it’s arguably more work: norms. Explicit expectations within the team. An understanding of where delegating is fine and where it isn’t.
And with that, the question shifts one last time. In the end it’s no longer about what’s allowed, but about what kind of team you want to be. “May” can be answered with a rule. “Should” only with a shared attitude.
Photo by Markus Spiske on Unsplash