People used to say: learn your multiplication tables, do the math yourself - in your head or on paper.
And not everything we were told “back then” becomes ridiculous simply because it sounds old-fashioned now. Knowing your multiplication tables is still useful. So is being able to work something out without reaching for a device. What changed is something else: today, very few people would feel embarrassed about using a calculator when a calculator is simply the right tool.
And, of course, I am not really talking about calculators.
AI has long since outgrown that analogy. People have been arguing for years about how its development and widespread adoption may change work, education, society, and people themselves. Psychologists and researchers studying human-AI interaction are trying to understand how sustained interaction with AI may affect human behaviour, social norms, and psychological well-being. Recent research does indeed examine these effects, from well-being to the possible spillover of interaction habits into human relationships.
And this is no longer just distant futurism. On August 27, 2026, more than a hundred companies and organizations - spanning AI, technology, finance, cybersecurity, and infrastructure - backed an open letter calling for a major collective push on cyber defence. Its warning was fairly concrete: AI-enabled cyberattacks are expected to become far more widespread and sophisticated in the coming months.
And yet, with all the threats, arguments, and ambiguity surrounding it, AI is already part of ordinary life. Homemakers and business owners, teachers, athletes, and members of clubs devoted to exotic fish... Engineers and architects were hardly going to be the exception.
One way or another, engineers use AI: sometimes with a strange degree of embarrassment. A little like quietly reaching for a calculator in a situation where, somehow, you felt you were supposed to manage on your own.
“But you just said AI has long since outgrown the calculator,” someone could quite reasonably object.
Yes. The analogy is limping, and quite visibly so. But it still has one working leg: both a calculator and AI remain tools - even if the latter increasingly participates not only in calculation, but in framing the question, choosing a solution, explaining it, and occasionally arguing with the person who opened the tool in the first place.
Sometimes that is useful. Sometimes it is a little unsettling.
And that almost inevitably leads to another question: where did this slightly strange claim that engineers can be “embarrassed” to use AI come from?
Fair point. The word is not quite right.
But that is exactly why, this time, I did not ask AI to find me a more accurate one. I did not want to throw away the feeling that brought the word here along with the naïveté of the word itself.
I am one of those people. I came into software development at a fairly mature age and, despite years of practice, apparently never quite got rid of a very school-like habit: the uneasy feeling that using someone else’s solution, even in a modified form and in a different context, is still a little bit like copying.
The Question That Wouldn’t Go Away
That is probably one reason I find the watermarking of AI-generated text so interesting. Although the engineer in me reacts to that concern somewhat inconsistently.
A healthy modern codebase already does quite a lot to smooth out an individual developer’s personal handwriting. A formatter puts the whitespace back in line. A linter argues with your favourite constructions. Conventions make naming decisions on your behalf. Code review gradually turns a personal solution into a team one. Refactoring can eventually strip the code of almost any ability to tell whose fingers originally typed it.
In a sense, we have spent years deliberately sanding down many of the surface-level signs of individuality that we suddenly find tempting to use as proof of human origin.
And yet the question “Who wrote this?” remains remarkably difficult to kill - and surprisingly many-sided.
Thankfully, we do not yet award a share of authorship to the mouse for a particularly successful drag-and-drop, or to the keyboard for an especially expressive Enter. A tool participates in producing a result, but participation alone clearly is not enough to call the tool an author.
With AI, the boundary becomes more interesting.
Suppose we detect a reliable watermark from a particular generative system in a piece of text. That is already a useful fact: we know something about the origin of at least part of the material.
But we still do not know what the human did.
Asked one question and copied the answer? Spent several hours arguing with the model? Rewrote half of it? Brought the original idea? Rejected ten proposed directions? Asked AI to fix nothing but the commas?
The watermark does not tell us any of that.
And its absence does not automatically turn the text into proof of exclusively human origin either.
That distinction is worth keeping precise: text watermarking exists in deployed systems such as Google DeepMind’s SynthID for Gemini text, but DeepMind itself explicitly describes watermarking as a useful signal rather than a universal solution for identifying AI-generated content.
Looking Past the Snapshot
I became curious whether there was a way to look deeper, and at this point I did the most natural thing one can do in an article like this: I asked AI.
It suggested looking not for a single fingerprint in a finished fragment, but for long-lived structures: recurring features of a text or codebase, characteristic decisions, the development of the same ideas over time, persistent patterns, even irregularities that do not appear once but accompany the work as it evolves.
At first glance, it was a good idea.
Instead of a snapshot, a history. Instead of one suspicious sentence, a trail stretching across many changes.
Even mistakes seemed useful here. Humans, after all, do not make mistakes perfectly. We forget things, come back to them, change direction, and sometimes carry the same small peculiarity around with us for quite a while.
That, however, is not quite how it played out.
Modern AI can also sustain a topic over a long period, preserve structures, return to previous decisions, make mistakes, lose context, recover it, and - when necessary - produce a fairly convincing version of the same “natural imperfection” we had just been tempted to treat as a human trace.
By then, the stream of thought was getting suspiciously close to becoming a puddle.
AI had suggested a good way of distinguishing AI participation from human participation - and almost immediately became a rather good counterexample to its own suggestion.
But despite this small failure, following the thread of reasoning - into the rabbit hole and back out again - was becoming increasingly interesting.
Unexpectedly, the stream reached a familiar kind of multi-way branching: the sort of thing dear to a programmer’s heart and equally capable of offending the same programmer’s aesthetic sensibilities when served as the main course three times a day.
A Larger Bottle
Instead of staying with a locally stalled attempt to establish authorship conclusively, another, much broader switch came into view.
What exactly am I trying to preserve when using AI makes me uneasy? And what am I afraid of losing when I click an imaginary Approve button on a solution I did not come up with myself?
The skill, developed over years, of solving engineering problems independently? An understanding of every step that led to a particular solution? Professional competence? My own voice? Control? The job itself?..
Fortunately, that path remained a thought experiment - although AI rather confidently suggested that this was where “genuine freedom appears for the first time.”
I do not entirely disagree.
The narrow bottleneck of trying to establish authorship had suddenly widened. Instead of one almost binary question, we now had a rather large container capable of holding skill, understanding, voice, control, responsibility, work, and quite a few other things for which we could undoubtedly invent excellent names later.
There was only one small nuance.
Do we actually want to put ourselves into a bottle at all - even if the bottle has become large enough to contain us?
The narrow neck had been inconvenient, but it had also been rather picky. At least it attempted to filter out what failed validation.
Widening it almost all the way to complete freedom gave us far more allowed states - and with them fewer obvious restrictions, weaker boundaries, and considerably more ways of letting in things that probably should not have been let in.
For an engineer, this is suspiciously familiar territory.
Constraints, permissions, validation, security - all those boring things whose absence you generally notice at precisely the moment when it is becoming a little late to keep calling them boring.
In other words, the larger bottle was freer than the smaller one, but it still did not look much like an environment in which I would want to safely create something, verify it, change it - and let it gradually evolve.
That, in fact, is where my story with RecurseWright begins, and I would like to invite anyone interested to take part in it. The story has already begun, and it will gradually unfold under the #recursewright tag.
I will also be genuinely interested to read your own stories in the comments, even if they look nothing like the calculator story - or this particular stream. :)
P.S. As the article itself already makes clear, AI took part in writing it. In future posts, I will gradually show more of what that participation actually looked like.
Top comments (0)