I don't feel we have enough context to judge the full argument, but going by what you've quoted, it seems that his line of argument is indeed that composing a symbol such that it has a compound meaning but no relationship to a compound pronunciation is problematic, because you dissociate the symbol from the pronunciation. You make it arbitrary and unpredictable. The implicit conclusion is that readers of such a writing system will have a harder time with literacy and (to go a step further that I will soon explain) with cognitive processes supposedly tied to literacy.
I consider your counterexample from Chinese excellent. In reality this dissociation does not appear to be a hindrance to reading.
However, the fear Seidenberg expresses is not a new one. It rests on the opposite premise: that there are writing systems where there is an association between symbol and pronunciation, and that the pronunciation is predictable. This of course refers to alphabetic or syllabic writing systems.
Fundamentally, we can't escape the arbitrariness of the relationship. That the symbol < a > makes (in many languages) the sound we transcribe /a/ is an utterly contingent fact. However, we can combine these letters in more or less predictable ways and thereby make the pronunciation of new words predictable. In English this doesn't work so well, but if we take a spelling-reformed language like Spanish we can observe that an invented word like calabrenoda will be accurately pronounced by all speakers.
I presume you know all this. What might be surprising is the importance that has been attached to this fact in the past. Keep in mind that language theorists have historically (before linguistics became more of a science) identified random features of their own language and argued for their being the key to the cultural superiority of speakers of their language. The argument generally runs along one of two lines: (a) The mysteriousness/power of the feature shows that the language is highly expressive and poetic ("see how many meanings arise out of one word; how subtle is the English mind!"); or (b) the regularity/sense of the feature shows that its speakers have a naturally logical mental bent. Such arguments, incidentally, are used first and foremost to further political or sociological ends; for example, such linguistic observations were used by medieval Christian critics to argue (along nonsensical lines) that Hebrew was an inherently backward language, aligning with their conviction that Judaism was an inherently backward religion.
However, such arguments are not extinct in our day. One recent theorist who argues for the significance of the compositional quality is Kieran Egan, e.g. drawing from Vygotsky in The Educated Mind. For Egan, there is a succession of stages that humans and human cultures move through in terms of refinement of thought, and each succession requires that they have certain cognitive tools. He argues that having an alphabetic writing system lends itself to developing the cognitive tools to move on to the next stage. Therefore, for him, cultures whose language has a less phonologically transparent writing system are at a cognitive disadvantage. One example he cites is that of an indigenous people in Russia who were subjected to a logic test involving syllogisms: "Everywhere in the north is snowy; Siberia is in the north; is Siberia snowy?" They responded, "I don't know. I haven't been there." He claims that their inability to understand syllogisms (as he reads that exchange) can be attributed to the lack of what he considers a sophisticated, i.e. phonologically transparent, writing system.*
So that's where Seidenberg's argument may be tending towards—I'm speculating. Now, is that view valid? I would tend to side with you and your objection, partly because of that history I referred to. Linguistic features have been tied to cultural superiority questions for a long, long time and rarely with any good reason. Is China exactly a backwards society? I think not.
To address the question more head-on might require some good, hard data on literacy (% among population, age first learned, grade reading level, etc.) among Chinese students with comparable education to that of English students or students of other languages whose language is alphabetic. But intuitively I strongly suspect that there is not a significant discrepancy.
Again, this whole answer rests on reading a lot into a short quotation from Seidenberg. :)
* If I had my book with me I would cite the passage. I will try to remember to return to it when I'm home.