A useful tool helps us while we use it. But what remains after we stop?
A file. Some time saved. A slightly more impressive presentation. Apart from these, might a way of solving a problem that we would never have imagined become ordinary practice? We may forget who taught us, while the method stays in our hands.
One study looks beyond productivity while AI is in use. Levin Brinkmann and colleagues asked whether a machine's discovery could pass to a human, and then from that person to someone who had never encountered the machine. [1]
A route that begins with a loss
Participants played a game in which they moved through a network of points and links, collecting rewards. They wanted to earn as much as possible in ten moves. But choosing only routes with immediate gains would never lead to the highest reward. A small loss at the beginning opened the way to a much larger gain.
Lose once in order to win. Easy enough to write. Yet the move that accepts a loss right in front of us can be surprisingly hard to make.
The experiment involved 1,155 participants in the United States and United Kingdom, arranged into 30 populations spanning five generations. Each generation had eight places. In half the populations, three places in the first generation were occupied by machine players trained through deep Q-learning. That was the machines' only appearance. Later generations selected demonstrators by looking at the previous generation's performance. They observed and reproduced their moves, then solved the same network themselves. After this learning phase, they left demonstrations for the next generation.
These were not generations of parents and children. They were a relay of knowledge, with new participants taking over. AI did not sit beside every participant offering advice, nor could it be called back when someone got stuck.
You can inherit what you cannot discover
Among the 600 people in the human-only condition, just one discovered the optimal strategy independently. In contrast, nine of the 15 populations that began with AI preserved the strategy through the final generation. In the other six, it reached humans but was subsequently lost. [1]
It would be a pity to file those nine and six under success and failure and leave it there. The nine show that a machine's discovery can continue as a human practice after the machine has gone. The six show that demonstrating a superior solution once does not automatically update a culture.
The first discovery is not the only difficult part. Passing it on is a different difficulty.
If people cannot reproduce a method even after seeing it demonstrated, AI may remain indispensable. Even if they can reproduce it, they may not choose it as the next example if they cannot recognize its value. The authors distinguish difficulty of discovery, ease of transmission, and visibility of advantage, and use agent-based simulations to examine what happens when these conditions break down.
The human task, however, was designed to make all three conditions relatively favorable. In real work, the value of a method is not always apparent within minutes. The payoff may arrive six months later; the only person who knows why may have left the organization. That is a more demanding handover than this game.
Saving an answer, preserving a skill
From here, let us use the study as a starting point for reflection.
When AI helps produce a good result, we save the deliverable. We may go a step further and save the prompt. Yet when a similar problem returns, knowing what to notice, where to be skeptical, and which order to try things in may still belong to no one.
Storing an answer is not the same as increasing the number of people who can solve a problem.
This is not an argument for turning everything into a manual. If every good solution produced a hundred-page procedure, we would need to pass on the time to read it before we could pass on the knowledge. What matters may be whether the discovery has been broken into something the next person can actually try.
Superior machine capabilities that cannot be learned by humans still have considerable value. People do not need to learn to calculate as quickly as a calculator. Insisting that every use of AI improve a human skill creates a different kind of confinement.
So it helps to separate the possibilities. Is this a job for borrowing a capability? Or an opportunity to change the way we work? It can be either, but what remains will not be the same.
This was a short artificial task, not a measure of organizational culture over many years. The machines had vastly more opportunities to explore than the humans. Nor does it show that the strategy was impossible for humans to discover in principle. Still, there is a quiet surprise in seeing the ability to discover separated from the ability to preserve.
Someone who has never encountered the AI is using a method it taught us. At that point, what exactly have we introduced?