This is a bad argument; it is a fact, for example, that child pornography by being censored creates a taboo around it. It's also true that some people will seek it out because of its taboo nature - something it would not be (or only to a lesser extent) if it weren't illegal. The fact that people will download and masturbate to child porn because it is illegal is not a good argument against child porn laws.
We (justifiably, I think) do not allow speech for the purposes of terrorist recruitment. The fact that someone is mistakenly banned for terrorist recruitment and later becomes radicalized is not a good argument to allow terrorist recruitment.
The argument is fundamentally self-defeating; it presupposes that free speech is so good and basic that it overrides all other rights. However, it also supposes that people in general lack sufficient ability to introspect after being censored - people are thought of here not as rational beings to use their rights of speech to engage in democratic deliberation, but as animals who when poked by a stick become enraged.
>He did it by talking to them, engaging them in speech.
Why should it be society's burden to deradicalize people? Does this work at scale? What of all the people who heard Davis or his ideas but were not persuaded? Is the number of people he failed to convince known?
Finally, what of the people who, according to your theory, only become more 'radicalized' when they encounter the position that their views are wrong or harmful?
There's just as good of a chance that one hundred Daryl Davis' trying to deradicalize people will actually cause radicals to dig in their heels. Maybe these hypothetical Davis' don't have a welcoming tone. Maybe the radical doesn't want to listen to a hypothetical Davis because of his race. Maybe the radical actually publicises the exchange and uses it as a megaphone to gish-gallop with 'radical' ideas. Maybe the radical convinces a hypothetical Davis that actually the radical and hateful ideas are correct. Doesn't honest and open dialogue, after all, permit both sending and receiving?
I feel as though the Daryl Davis approach has a lot more risk and a lot less going for it empirically than you suppose.