It needs careful justification before making such a statement - the censorship system has a serious constraint on computational costs - it needs to operate on the stream of the entire outgoing international traffic, and to make a decision in real-time (or for back-analysis). We are talking about many terabytes per seconds of traffic, any censorship tools that have a high computational costs cannot be deployed for such a purpose, even if it runs okay on a single PC. Also, a high false-positive rate is not acceptable, as it will create massive service disruption and practically useless.

Unlike the case for SOCKS5-over-SSH, HTTPS-over-HTTPS, or VPN handshakes, which can be detected by relatively simpler rules, most deep learning tools required excessive CPU time, so it's unlikely that complex deep learning algorithms are being used, at least not the category that costs the highest CPU time (anything with "AI").

Given these constraints, the algorithms available to the censorship system is rather limited, it seems. What types of algorithms are being used, then? Unfortunately, nobody can answer this question. This is the fundamental question people are facing today. 10 years ago, every sysadmin in China knew the censorship system works, but today, the system has became completely opaque.