The same way I know which real person is serving me the website. I don't I merely know that the owner doesn't change randomly.
Still doesn't explain how I'll confirm that if the website has not been intercepted by a middle man the first time I visit it.
What about TLS certificates attested by CAs who validate the real world legal entity? Would you agree that this is a solved problem there?
This only matters, when your view of the entity isn't solely determined by the domain name. For example you care when someone impersonates google.com, because you expect it to belong to Alphabet Inc., you perceive impersonating when the entity you are talking to changes. When some domain is always constantly resolved to your ISP, then that domain is owned by the ISP in your network.
I think you've fundamentally misunderstood the problem. Nobody here is talking about the domain resolving to ISP.
We're talking about an MITM attack where a middle-man (doesn't even need to be the ISP, it could be a router in the middle acting like middle-man) intercepts our request and serving its own public key instead of the actual server's public key so that it can carry out an MITM attack.
With TOFU, there is no way to detect this attack the first time I am connecting to the website. Once you have foolishly trusted the middle-man's public key, you may not notice any problem for months or years. Then one day, the middle-man may decide to misuse your credentials that it has collected during MITM attack.
But that's exactly what we're trying to avoid. When I want to visit my bank website, I don't want the ISP to become the real website.
I don't understand your comments. Your solution to how I trust the website on first use (TOFU) is to trust whatever public key the middle man (the ISP) serves? If you're okay with that, I guess you don't have a problem. But I'm not okay with that, so TOFU doesn't solve my problem.
If you were to redesign name and address resolution, to enforce connecting to the real physical world entity, this should happen out-of-band. Well, know that I think of it, I think that's what the GNU Name System is trying to address: https://www.gnunet.org/en/gns.html .
-----
The question seams to be, what do you consider to be the real website? The one that answers to your request? Then the ISP is the real website. In your example you seem to have preconceptions of who might own the website, but these are outside of the network. TLS from Let's Encrypt only enforces that the entity never changes. There are validation schemes where the physical/legal entity is validated, but this is not the case here.