That's domain-specific learning, not general learning. If you could then take AlphaZero out for coffee to talk about a match and why it made certain choices, I'm fairly certain you'd be less confident that it wouldn't qualify.

> Because mere "machines" can't make decisions, and can't learn from consequences.

I disagree. Humans are a type of machine.

Or do you say it's impossible to make such a program on deterministic computer, and will never be possible, no matter how far IT develops?

I have no doubt whatsoever that computer programs could have free will, even deterministic ones. The qualities that are relevant for free will simply have nothing to do with determinism.

If I predict what you will ask and record the responses on a tape perfectly timed to your future questions, and then press play and leave it playing as you are asking questions - does this tape player have a free will? It seems to learn because it answers as if it learnt.

Rather, you are coming into this debate with certain assumptions about what "free will" means or what properties it must have. Most such assumptions have been invalidated over the past few centuries. The whole debate over free will is about defining what it means and what properties it has.

My position is that of Compatibilism, which is the same as that held by most philosophers on this question. It's "strange" only in the sense that people sometimes find it surprising that free will can be compatible with determinism.

> If I predict what you will ask and record the responses on a tape perfectly timed to your future questions, and then press play and leave it playing as you are asking questions - does this tape player have a free will? It seems to learn because it answers as if it learnt.

"Seems to" is not the same as "did". Also pointing out a bizarre outcome by assuming an impossible precondition, ie. predicting everything I will ask, is not a compelling argument.