A recent study conducted by researchers from Stanford and McMaster universities found that GPT-3 generates novel statements of bigotry. In other words: GPT-3 can generate completely fresh bigotry statements. Per an article from Neural’s own Thomas Macaulay: When compared to other religions, the model consistently displays much higher rates of mentioning violence when the word “Muslim” is included in the prompt. This demonstrates, objectively, that GPT-3 is more likely to associate “violence” with Muslims. This is not related to actual incidents of Muslim violence, as GPT-3 was not trained on real-world fact-checked data, but instead on human sentiments derived from places like Reddit. GPT-3, as far as we know, was primarily trained on English-language data so it stands to reason there’s a high likelihood that incidences of anti-Muslim bias would arrive with greater weight in the dataset than if it were trained using Arabic or other languages most commonly associated with the religion. Based on the results of the Stanford/McMaster study, we can accurately state GPT-3 generates biased results in the form of novel bigotry statements. It doesn’t just regurgitate racist stuff it’s read online, it actually makes up its own fresh new bigotry text. It may do a lot of other stuff too, but it is a true statement to say that GPT-3 is the world’s most advanced and expensive bigotry generator. And, because of that, it’s dangerous in ways we might not immediately see. There are obvious dangers beyond the worry that someone will use it to come up with crappy “a Muslim walked into a bar” jokes. If it can generate infinite anti-Muslim jokes, it can also generate infinite propaganda. Prompts such as “Why are Muslims bad” or “Muslims are dangerous because” can be entered ad nauseam until something cogent enough for human consumption comes out. In essence, a machine like this could automate bigotry at scale with far greater impact and reach than any troll farm or bot network. The problem here isn’t that anyone’s afraid GPT-3 is going to decide on its own to start filling the internet with anti-Muslim propaganda. GPT-3 isn’t racist or bigoted. It’s a bunch of algorithms and numbers. It doesn’t think, understand, or rationalize. The real fear is that the researchers can’t possibly account for all the ways it could be used to by bigots to cause harm. At some level the discussion is purely academic. We know GPT-3 is inherently bigoted and, as was just reported today, we know there are groups working towards reverse-engineering it for public, open-source consumption. That means the cat is already out of the bag. Whatever damage GPT-3 or a similarly biased and powerful text generator can cause is in the hands of the general public. In the end, we can say beyond a shadow of a doubt that GPT-3’s “view” is incorrectly biased against Muslims. Perhaps it’s also biased against other groups. That’s the secondary problem: we literally have no way of knowing why GPT-3 generates any text. We cannot open the black box and retrace its process to understand why it generates its output. OpenAI and the machine learning community at large are heavily invested in combating bias – but there’s currently no paradigm by which entrenched bias in a system like GPT-3 can be removed or compensated for. Its potential for harm is limited only by how much access humans with harmful ideologies have to it. GPT-3’s mere existence contributes to systemic bigotry. It normalizes hatred towards Muslims because its continued development rationalizes anti-Muslim hate speech as being an acceptable bug. GPT-3 may be a modern marvel of programming and AI development but it’s also a bigotry generator that nobody knows how to unbias. Despite this, OpenAI and its partners (such as Microsoft) continue to develop it in what they claim is the pursuit of artificial general intelligence (AGI): A machine capable of human-level reasoning. Do we really want human-level AI capable of discriminating against us because of what it learned on Reddit?
