You may have heard the news that a college student used GPT-3 to write fake blog posts and ended up at the top of Hacker News. Last week, college student Liam Poor heard of GPT-3. By the end of the week, he posted a fake blog under a fake name. “It was super easy actually,” he says, “which was the scary part.”
Just in case you’re not acquainted with GPT-3: It’s the most recent version of a series of AI autocomplete tools designed by San Francisco-based OpenAI and has been in development for several years. GPT-3 is that the largest natural language processing (NLP) transformer released so far, eclipsing the previous record, Microsoft Research’s Turing-NLG at 17B parameters, by about 10 times. Unsurprisingly there has been much excitement surrounding the model, and, given the plethora of GPT-3 demonstrations on Twitter et al. , OpenAI has been pretty accommodating in providing beta access to the new API. This has resulted in an explosion of demos: some good, some bad, all interesting. Some of these demos are now being touted as soon-to-be-released products, and in some cases may be useful.
Machine learning models are only nearly as good, or as bad because the data fed into them during training. In the case of GPT-3, that data is huge. GPT-3 has increased the number of parameters more than 100-fold over GPT-2, from 1.5 billion to 175 billion parameters. It was trained on the Common Crawl dataset, a broad scrape of the 60 million domains on the web alongside an oversized subset of the sites to which they link. This means that GPT-3 ingested many of the internet’s more reputable outlets along with the less reputable ones. Yet, Common Crawl makes up just 60% of GPT-3’s training data.
Like all deep learning systems, GPT-3 looks for patterns in data. To simplify things, the program has been trained on an enormous corpus of text that it’s mined for statistical regularities. These regularities are unknown to humans, but they’re stored as billions of weighted connections between the various nodes in GPT-3’s neural network. Importantly, there’s no human input involved during this process: the program looks and finds patterns with none guidance, which it then uses to finish text prompts. Tucked away within the GPT-3 paper’s supplemental material, the researchers give us some insight into a little fraction of the problematic bias that lurks within. Just as you’d expect from any model trained on a largely unfiltered snapshot of the net, the findings may be fairly toxic.
How might this play move into a real-world use case of GPT-3? Let’s say you run a media company, processing huge amounts of information from sources everywhere the planet. You might want to use a language model like GPT-3 to summarize this information, which many news organizations already do today. If the model carries a negative sentiment on your site will also receive that negative slant.
The creators of GPT-3 decide to continue their research into the model’s biases, but for now, they simply surface these concerns, passing along the risk to any company or individual who’s willing to take the chance. All models are biased, as we all know, and this could not be a reason to outlaw all AI, because its benefits can surely outweigh the risks within the long run. Even if this program, this ultimate, all-knowing autocomplete, All we’d have to do would be to ask the right things.
