For instance, Daniel Dennett's objections. Dennett calls Searle's thought experiment a "boom crutch", something that hinders thinking, as opposed to intuitive pumps, that are considered to be helpful.
Dennett claims that Searle's argument is phrased in a specific way to limit your options, hence a "boom crutch". Searle assumes some highly advanced Symbolic rule-based AI. He wants you to believe that it could not understand meaning because it merely assembles some input symbol to an output symbol according to the rule-book. But, immediately, we are faced with objections.
Firstly, according to Dennett, the argument shows Searle's lack of understanding on how computers and how computation is achieved. It would take billions of years of writing code for a purely rule-based system* to be able to obtain a meaningful human-level conversation. Just how many rules would it have to have? One has to construct rules convincing enough so that each response is as if a real person produced an answer, leading the conversation in Chinese, on a conversational level. Try to construct all of these rules using Symbolic logic; it would be an unbelievable number.
Secondly, even if we achieved the device that utilizes a rule-based system and such system can talk about movies, weather, Greek cuisine and ask personal questions, did we not just implement the entire process of human cognition, along with understanding? The "system" that Searle presents cannot merely assemble input to output because it has to rely on billion years worth of rule sets on what to say, and what not to say. It has to draw logical inferences involving large portions of empirical data about the world. It has to have common-sense. And if it does, it has no problems with understanding anything whatsoever, semantics included. This kind of sophisticated rule system is impossible to localize in some rudimentary boxes with instructions. What one needs is computation to do that efficiently, if at all.
Building on Dan's argument, it is clear that John Searle wants you to forget about a crucial part of what makes the system; where and how the actual stuff must be going on. If only banal manual assembly of input to output was the case, there would be no possibility of meaningful human-level conversation. Literally, no one claims that a simple program that prints out "Hello World" understands semantics. In the same way, a man in the room is unable to have a convincing conversation. Moreover, a conversation that doesn't take years for him to process, if he's using boxes of billion years' worth of information. This is a charge of natural possibility. Searle wants you to imagine something trivial and draw obvious conclusions from it - that such system cannot understand, so "strong AI" altogether is implausible. However, to pass anything resembling a Turing test, it is non-trivial. The entire discussion of the rule-set system, which has to exist aside from speech assembler, is almost "deleted" from Searle's argument.
Otherwise, in what way does the AI get the rules from? Of course, the only option is this gargantuan set of rules written over billions of years. How can one squeeze that rule-set so obtaining information is feasible? Only by using a computer. If a computer is practically used, then the computer can yield understanding.
By shifting your attention to a dumb assembler, Searle makes you forget where the real work must be going on, i.e. in the complex computational unit responsible for (and capable of) human-level cognition because otherwise, it could not pass a Turing test, which Searle's experiment is a variant of. The speaker on the other end needs to be well convinced that he/she is speaking to a real person.
*- As opposed to, say, Machine-learning system.