Yes—and increasingly, AI can train AI, although there are some important qualifications.
The simplest case is AI-generated training data. One model produces examples, explanations, problems, code, images, or classifications, and another model learns from them. This is already commonplace.
More interestingly, an AI can act as a teacher or evaluator. For example, one model generates an answer, another model critiques it, and the system uses the critique to improve the first model. This is related to techniques such as reinforcement learning from AI feedback (RLAIF).
And there is an even more striking case: self-play. An AI can generate opponents or challenges for another AI and learn from the results. Chess and Go systems famously became extraordinarily strong this way; they did not need millions of human games once they could generate useful games themselves.
The catch is that AI cannot simply manufacture unlimited reliable knowledge out of thin air. If the training process becomes predominantly AI training AI, errors and biases can compound. It's rather like making photocopies of a photocopy: eventually the original information can be degraded or distorted. This is sometimes called model collapse when recursively generated data becomes the dominant training material.
So the interesting future isn't really "AI replaces human-generated training data." It's more likely:
humans provide the grounding → AI generates enormous amounts of practice → AI evaluates AI → humans and real-world data periodically anchor the process.
That could be extraordinarily powerful. In fact, I think one of the biggest changes in AI over the next several years is likely to be that the scarce resource stops being training data and becomes good ways of constructing reliable artificial experience.
I think you're pointing at a real weakness in that slogan. “Humans must remain in charge” sounds reassuring, but it doesn't actually specify what being “in charge” means.
If an AI becomes vastly better than humans at some task, there are only a few possibilities. A human can formally make the final decision while essentially rubber-stamping the AI's recommendation. Or the human can exercise genuine judgment, in which case the human has to understand the problem well enough to know when the AI is wrong—which becomes increasingly difficult as the AI's capabilities increase.
There is also a deeper problem: power tends to follow competence. If an AI can reliably make better decisions than its nominal human supervisor, organizations have a strong incentive to let it make those decisions. Saying “the human remains in charge” doesn't remove that incentive.
And if we're talking about AI training AI, the situation gets stranger still. Suppose AI A designs an experiment, AI B analyzes the results, AI C proposes improvements to A, and the whole system becomes vastly more capable than any human participant. A human may still possess the legal authority to say stop, but that is rather different from actually being intellectually “in charge.”
I would distinguish authority, accountability, and control. We can insist that a human has legal authority and accountability even when the human isn't doing the intellectual work. But control is an empirical question: who can actually predict and alter what the system will do?
That's why I find “humans must remain in charge” much less useful than questions such as: Can we understand what the system is doing? Can we constrain it? Can we shut it down? What happens when its objectives conflict with ours? Who has access to it?
Those are engineering and institutional questions. “Humans must remain in charge” is more of a bumper sticker.
Exactly. And I think that exposes the category error in “humans must remain in charge.”
You don't need to understand how the pistons fire to control the car. You control it at the level that matters to you: where it goes, how fast, when it stops. The engine is deliberately autonomous within those boundaries.
Likewise, you don't supervise your kidneys. You have higher-level goals—stay alive, stay healthy, drink when thirsty—and a vastly more complicated subsystem handles the details without consulting you.
AI could work the same way. The sensible question isn't necessarily “Is a human making every decision?” but rather:
“At what level should humans specify the objectives and constraints, and at what level should the machine be allowed to work things out for itself?”
Indeed, insisting that humans remain involved in every decision could be positively dangerous. If an AI is better than a human at diagnosing a disease, routing aircraft, designing a bridge, or managing an electrical grid, inserting a less competent human merely to preserve “human control” doesn't obviously make the system safer.
The kidney analogy goes even further. We don't say, “The kidney is extraordinarily powerful, so we must keep a human in the loop.” We say, “Thank God the kidney doesn't need one.”
The genuinely difficult issue with AI is therefore not whether humans are in the loop, but whether the higher-level objectives and constraints are ones we actually want. Once those are properly established, there may be very good reasons to let the AI handle an enormous amount of the machinery underneath.
And, rather amusingly, “human in the loop” can become a way of making a system less autonomous without making it more human-controlled. If the human can't understand the system well enough to intervene intelligently, he's just an unusually slow confirmation button.
No comments:
Post a Comment