What if GPT-3 is used as a content filtering tool, aka censorship?
I’m kind of late to the GPT-3 hype, which is probably due to my slight disdain for AI related news. For the most part, I don’t find the AI researcher’s desire to “forge the gods” all that interesting, nor do I like to catch the hype at its peak. Nonetheless, I’ve heard GPT-3 mentioned multiple times in the Lex Fridman podcast. What really made me pay more attention to it, and honestly scared me as bad as the next guy, is this video [1] where a real person does an interview with a GPT-3 conversational API. I don’t know about you, but, if judging solely based on its command of the written word, and not the deepfake-looking avatar used for anthropomorphic effects, this AI passed the Turing test for me. After digging slightly deeper, I found out that GPT-3 can also accomplish impressive tasks other than being a conversational bot, for example, translate human language description into a react.js code snippet, read an article such as the bitcoin white paper and make a summary of it, take the name of an object and output a list of affordances (door -> open, close, walk through, etc), and many more (even make memes!). If you are a web developer or a writer, this should sound pretty scary already. What made me extra worried about the is, well, the title of this post — what if GPT-3 is used for censorship?
Censorship is a broad term, nowadays it describes something like a vastly imposed suppression on ideas, speech and the systematic infringement on citizens’ privacy. Sounds familiar? If the tech giants in the west are debuting their own censorship career, they are still way behind their counterparts in China. I like to say that the only aspect where Chinese technological companies reign supreme is in the area of large scale censorship. Taking China and its Great Firewall (GFW) infrastructure as an example, I’d like to explore if GPT-3 would make a censorship apparatus more deadly, and how much. First, let’s start with a brief description of how the GFW works. Note that since there’s no leaked official documentation for the GFW that I know of, this is based mostly on experimental data, security research, and sometimes just speculation, all of which could be wrong or outdated. I’m also not an expert in either security or AI. So, take everything with a grain of salt. For more details, I suggest the book by James Griffiths, which is a very well-written non-technical account of the GFW’s history. Citizen Lab also has many papers on this topic, which are more technical.
The most common and obvious feature of the GFW is censoring western social platforms like Facebook, Twitter, YouTube. This is done in multiple ways, from simple to complex, they are: IP blocking, DNS tampering, deep packet inspection, and of course, manual inspection [2], which you could argue is the most complex since it can’t be automated by a machine. China has governmental control over all of their ISPs (Internet service providers), therefore, it can enforce a blacklist of IP addresses within gateway routers that route all of China’s internet traffic to the outside world. Outbound traffic with a blacklisted destination IP address are injected with null routes. DNS queries that contain forbidden domain name like (google.com) have their DNS records tampered with, returning a wrong IP address instead of Google’s correct IP. With a tampered DNS record, the website is effectively unaccessible. This approach is rudimentary, since it can be bypassed with proxies, which are intermediate destinations with unblocked IP addresses that re-route traffic to their intended destination. Other tools such as VPN, Onion Routing (Tor), encryption (HTTPS) also help bypassing the filter to various degrees of effectiveness. However, the GFW has evolved at a overall faster rate than tools that serve to bypass it. Machine learning and deep packet inspection are used to shutdown VPN or proxies [3]. Active probing is a technique used to shutdown Tor bridges, which led the Tor developers to develop new techniques to counter it, leading to an arms race between China and Tor.
Since I’m not a AI researcher or even follow the news in the field that closely, the best I can do is guess how GPT-3 could enhance a system like the GFW. Looking at the brief overview of its structure, it’s possible that the censors use GPT-3 to identify traffic for VPN, proxies, or Tor, since they already use machine learning for that task dating as early as 2008 [4]. However, I think the biggest use for a language processing model like the GPT-3 would be to replace manual inspection. The GFW, for all its technological achievements, is still not a completely automated system. The main purpose of the GFW is to crush voices of dissent to a degree where people who feel the same way are left in the dark. Therefore, it has to be smarter than the dissenters, which is hard because human creativity is something to behold. The Chinese languages has incredible capacity for wordplay, shifting characters for similar sounding ones, and hard to understand euphemisms that all make automated content filtering systems have a hard time. However, with something like GPT-3 who seems to understand very clever manipulation of language like humor, memes, essays, it’s not hard to imagine a way more accurate and efficient content filtering system using this new tool.
Whether GPT-3 will become a deadly weapon in the hands of censors, or whether something makes this more difficult than it seems, I don’t know enough to bet my money on it. Just like anything else in recent times, perhaps the only thing left to end the discussion with is: I guess we’ll see how it goes. In the mean time, I certainly want to keep up with news related to this topic.
[1] https://www.youtube.com/watch?v=PqbB07n_uQ4&t=5s
[2] https://anonymster.com/great-firewall-china-bypass/
[3] https://www.cs.tufts.edu/comp/116/archive/fall2016/ctang.pdf
https://www.cs.tufts.edu/comp/116/archive/fall2016/ctang.pdf
[4] https://link.springer.com/article/10.1007/s10796-008-9131-2
