For Centuries, People Dreamed of a Machine That Could Produce Language. Then OpenAI Made One

OpenAI’s GPT-2 program churns out natural language that’s remarkably coherent—and that’s a problem

4 min read
Greg Brockman & Ilya Sutskever from Open AI, in front of a background comprised of a generic language model chart, GPT-2 screenshot; and Reddit screenshots.
Greg Brockman and Ilya Sutskever from OpenAI, who chose not to release their GPT-2 language model to the world.
Photo-illustration: Gluekit

This is part six of a six-part series on the history of natural language processing.

In February of this year, OpenAI, one of the foremost artificial intelligence labs in the world, announced that a team of researchers had built a powerful new text generator called the Generative Pre-Trained Transformer 2, or GPT-2 for short. The researchers used a reinforcement learning algorithm to train their system on a broad set of natural language processing (NLP) capabilities, including reading comprehension, machine translation, and the ability to generate long strings of coherent text. 

But as is often the case with NLP technology, the tool held both great promise and great peril. Researchers and policy makers at the lab were concerned that their system, if widely released, could be exploited by bad actors and misappropriated for “malicious purposes.”

The people of OpenAI, which defines its mission as “discovering and enacting the path to safe artificial general intelligence,” were concerned that GPT-2 could be used to flood the Internet with fake text, thereby degrading an already fragile information ecosystem. For this reason, OpenAI decided that it would not release the full version of GPT-2 to the public or other researchers.

GPT-2 is an example of a technique in NLP called language modeling, whereby the computational system internalizes a statistical blueprint of a text so it’s able to mimic it. Just like the predictive text on your phone—which selects words based on words you’ve used before—GPT-2 can look at a string of text and then predict what the next word is likely to be based on the probabilities inherent in that text.

GPT-2 can be seen as a descendant of the statistical language modeling that the Russian mathematician A. A. Markov developed in the early 20th century (covered in part three of this series).

GPT-2 used cutting-edge machine learning algorithms to do linguistic analysis with over 1.5 million parameters.

What’s different with GPT-2, though, is the scale of the textual data modeled by the system. Whereas Markov analyzed a string of 20,000 letters to create a rudimentary model that could predict the likelihood of the next letter of a text being a consonant or a vowel, GPT-2 used 8 million articles scraped from Reddit to predict what the next word might be within that entire dataset.

And whereas Markov manually trained his model by counting only two parameters—vowels and consonants—GPT-2 used cutting-edge machine learning algorithms to do linguistic analysis with over 1.5 million parameters, burning through huge amounts of computational power in the process.

The results were impressive. In their blog post, OpenAI reported that GPT-2 could generate synthetic text in response to prompts, mimicking whatever style of text it was shown. If you prompt the system with a line of William Blake’s poetry, it can generate a line back in the Romantic poet’s style. If you prompt the system with a cake recipe, you get a newly invented recipe in response.

Perhaps the most compelling feature of GPT-2 is that it can answer questions accurately. For example, when OpenAI researchers asked the system, “Who wrote the book The Origin of Species?”—it responded: “Charles Darwin.” While only able to respond accurately some of the time, the feature does seem to be a limited realization of Gottfried Leibniz’s dream of a language-generating machine that could answer any and all human questions (described in part two of this series).

After observing the power of the new system in practice, OpenAI elected not to release the fully trained model. In the lead up to its release in February, there had been heightened awareness about “deepfakes”—synthetic images and videos, generated via machine learning techniques, in which people do and say things they haven’t really done and said. Researchers at OpenAI worried that GPT-2 could be used to essentially create deepfake text, making it harder for people to trust textual information online.

Responses to this decision varied. On one hand, OpenAI’s caution prompted an overblown reaction in the media, with articles about the “dangerous” technology feeding into the Frankenstein narrative that often surrounds developments in AI.

Others took issue with OpenAI’s self-promotion, with some even suggesting that OpenAI purposefully exaggerated GPT-2s power in order to create hype—while contravening a norm in the AI research community, where labs routinely share data, code, and pre-trained models. As machine learning researcher Zachary Lipton tweeted, “Perhaps what's *most remarkable* about the @OpenAI controversy is how *unremarkable* the technology is. Despite their outsize attention & budget, the research itself is perfectly ordinary—right in the main branch of deep learning NLP research.”

OpenAI stood by its decision to release only a limited version of GPT-2, but has since released larger models for other researchers and the public to experiment with. As yet, there has been no reported case of a widely distributed fake news article generated by the system. But there have been a number of interesting spin-off projects, including GPT-2 poetry and a webpage where you can prompt the system with questions yourself.

Mimicking humans on Reddit, the bots have long conversations about a variety of topics, including conspiracy theories and

There’s even a Reddit group populated entirely with text produced by GPT-2-powered bots. Mimicking humans on Reddit, the bots have long conversations about a variety of topics, including conspiracy theories and Star Wars movies.

This bot-powered conversation may signify the new condition of life online, where language is increasingly created by a combination of human and non-human agents, and where maintaining the distinction between human and non-human, despite our best efforts, is increasingly difficult.

The idea of using rules, mechanisms, and algorithms to generate language has inspired people in many different cultures throughout history. But it’s in the online world that this powerful form of wordcraft may really find its natural milieu—in an environment where the identity of speakers becomes more ambiguous, and perhaps, less relevant. It remains to be seen what the consequences will be for language, communication, and our sense of human identity, which is so bound up with our ability to speak in natural language.

This is the sixth installment of a six-part series on the history of natural language processing. Last week’s post explained how an innocent Microsoft chatbot turned instantly racist on Twitter.

You can also check out our prior series on the untold history of AI.

The Conversation (0)

Will AI Steal Submarines’ Stealth?

Better detection will make the oceans transparent—and perhaps doom mutually assured destruction

11 min read
A photo of a submarine in the water under a partly cloudy sky.

The Virginia-class fast attack submarine USS Virginia cruises through the Mediterranean in 2010. Back then, it could effectively disappear just by diving.

U.S. Navy

Submarines are valued primarily for their ability to hide. The assurance that submarines would likely survive the first missile strike in a nuclear war and thus be able to respond by launching missiles in a second strike is key to the strategy of deterrence known as mutually assured destruction. Any new technology that might render the oceans effectively transparent, making it trivial to spot lurking submarines, could thus undermine the peace of the world. For nearly a century, naval engineers have striven to develop ever-faster, ever-quieter submarines. But they have worked just as hard at advancing a wide array of radar, sonar, and other technologies designed to detect, target, and eliminate enemy submarines.

The balance seemed to turn with the emergence of nuclear-powered submarines in the early 1960s. In a 2015 study for the Center for Strategic and Budgetary Assessment, Bryan Clark, a naval specialist now at the Hudson Institute, noted that the ability of these boats to remain submerged for long periods of time made them “nearly impossible to find with radar and active sonar.” But even these stealthy submarines produce subtle, very-low-frequency noises that can be picked up from far away by networks of acoustic hydrophone arrays mounted to the seafloor.

Keep Reading ↓Show less