Then when law enforcement claims they are not being cooperative, they can say they have a tool that meets their needs if they're patient.
Courts don't take to kindly to people trying to be cute with their demands. It's not like they are going to say "well they are technically right" and give up, they are going to just up the consequences or clarify the request until you comply in the way that everyone knows they want you to.
Edit: And according to the wikipedia page [1] for Lavabit, he was successfully held in contempt of court for the printout move.
We absolutely do not need the "not my problem, I don't care, engage in mass surveillance all you want!" engineers.
And even in the US it still isn't a sure thing that Lavabit's response would work again if someone else tried it. I don't know the details, but I believe there is still some uncertainty around if the FBI just kind of "allowed" them to close down by not pursuing it any further, or if they got what they needed, or of secret laws were changed because of this instance.
In Lavabit's case, there was a lot of FBI involvement, a lot of secret court orders and gag orders, and a lot of accusations from the owner of Lavabit that he was brought to secret courts without legal representation and no chance to appeal, and even he says that there are things he still can't talk about.
Please note that we're not seriously suggesting that encryption providers should adopt this -- not as long as there are other options. But if you're legally obligated to do something, this is the "f*ck off and leave me alone" approach to compliance.
[1] C.V. Wright and M. Varia. Crypto Crumple Zones: Enabling Limited Access without Mass Surveillance. In Proceedings of IEEE European Symposium on Security & Privacy, 2018. https://www.ieee-security.org/TC/EuroSP2018/program.php#euro... http://web.cecs.pdx.edu/~cvwright/papers/crumplezones.pdf
Thinking like an economist, you want to align the incentives to make it possible but not free to access user data. A weak key (per user) that's breakable with $10k compute cost seems about right to me, but the actual optimal cost may be higher or lower.