The difference is that if Telegram was intercepting messages, there would never be any indication to users. If Signal did it there would be evidence on the client side, in the form of tampered safety numbers.
We can worry that a user is very unlikely to notice bad safety numbers, but "very unlikely" is much different from "impossible" when it comes to mass surveillance. All it takes is a single public scandal where Signal provably tampered with safety numbers for the whole service to be thrown into question. Not to mention the dramatic surge in user interest in "safety numbers" that would result from such attacks being detected in the wild.
How would the user prove that a particular device never existed?
Logging messages under TOFU means you are either actively attacking the users (and they get a message when you stop) or you aren't getting the messages.
Most messages are not first messages. TOFU is valid, as evidenced by years of successful use by SSH.
Then, say, Telegram in non-secret message mode, is just as secure, in the sense that you have to trust a third party. Not clearly informing the user of this trust is dishonest.
>...after studies showed non-technical users (reasonably) associate the idea of fingerprints with criminal investigations
We would be interested in a reference to one of those studies. "Fingerprint" at least implies something about identity. "Safety" is more obscure. I feel that both are less than optimal here for imparting the important concept to the user. I think that "PGP number" or "Signal number" would be better. There is already preexisting culture around "Phone numbers" and identity.