
OpenAI AP
Generative AI Is More Mad Lib Than HAL9000
COMMENTARY
Published Aug. 30, 2023 • Read time: 7m
A quiet revolution has been occurring in artificial intelligence. Whereas in the past, AI was limited to analyzing information that already exists, it has now begun to create on its own, generating images, documents, speech, and video. In scenes reminiscent of science fiction, AI can now reason about the world in complex ways, write new computer programs to do its bidding, and even hire humans to perform tasks on its behalf. Pundits warn that the AI apocalypse is right around the corner, while naysayers dismiss it as akin to glorified auto-complete engines. What is the reality of AI today, and what are its risks?
So-called “generative AI” advances have been building steadily over the past few years. Image generators were the first to become widely available, prompting an initial wave of concern over “deep fake” images before the idea went mainstream through services like Midjourney, Stable Diffusion, and Dall-E. Today it has become almost routine to type in a textual description like “astronaut riding a unicorn on the moon” and, within minutes, get back stunningly realistic images created from scratch by an AI algorithm.
Image-generation tools are among the most mature generative AI tools, capable today of instantly creating photorealistic imagery from a simple textual description. Trained from billions of images drawn from across the web, modern image generators can learn any artist’s style and mimic it nearly flawlessly, in some cases yielding so much “fan imagery” that they replace the original artist in search results. Images are so realistic that they can spread virally and even affect the stock market. Human creativity is still required to decide what to ask the model to draw and to select the “best” version of the generated images, but the machine now provides the previously expensive artistic talent to render it.
The equivalent for text generation is a technology called a “large language model” (LLM). The debut of OpenAI’s ChatGPT last fall thrust LLMs into the public imagination, with ChatGPT, Google’s Bard, Microsoft’s Bing AI, and Anthropic’s Claude among some of the best-known consumer versions. Similar to their image counterparts, LLMs are trained by crawling the Internet and ingesting vast archives of books, social media posts, and other content. Critically, though, these models don’t “learn” like humans do. Instead of abstracting from the words on the page to the concepts they represent, LLMs merely memorize the statistical patterns of how words are used. ChatGPT doesn’t understand what a dog is; it merely encodes that in the sentence “he walked his _,” statistically, the next word is more likely to be “dog” than “goldfish.”
Like the search engines and social media platforms that preceded them, these systems are not neutral, factual reference databases. Instead, through constant human adjustment, they encode a variety of political and societal biases that nudge them toward specific answers, but with the aura of a neutral arbitrator. The silver lining: As social media companies explore the use of these systems for content moderation, these biases could become more transparent and discoverable than they are at present.
When asked a question, LLMs don’t think and reason at an abstract level like humans do. Instead, they merely perform a glorified game of “auto complete”: it’s more Mad Lib than HAL9000. This means that they frequently “hallucinate” by connecting unrelated snippets of text together to report false information. Asked about the Chinese spy balloon incident, for example, ChatGPT summarized it as a nuclear-capable hypersonic missile test. LLM chatbots have confidently labeled people as convicted criminals and even accused them of terrorism while fabricating evidence, quotations, and citations. According to one report, ChatGPT connected a lawyer to a “sexual harassment claim that he said had never been made, which supposedly took place on a trip that he had never taken for a school where he was not employed, citing a nonexistent newspaper article as evidence.”
Closely related to hallucination is plagiarizing, where, instead of summarizing in their own words, a lot of the text generated by AI models is simply copied and pasted from the Internet. This memorization is one of the reasons that ChatGPT regurgitates entire passages from copyrighted books rather than paraphrasing them the way a human would. Even AI-powered search engines largely work by copy-pasting sentences from the web together rather than reading and summarizing them as claimed, leading to often comical results.
To make their results more natural, AI models add a bit of randomness to their output, choosing randomly among the most probable phrases. AI researchers often talk about these new models as achieving human-like intelligence and reasoning; one of the reasons they describe them as such is that for any given question, there is always a random chance of the output being something profound. For example, one model posited a richly detailed explanation of Russia’s invasion of Ukraine that featured its own original ideas. Yet running the model a second time revealed that far from producing a work of artificial genius, the AI had merely randomly strung together a set of sentences based on the title of a television show.
The legal landscape around AI is still largely unsettled. Can you sue an AI service for libel if it falsely accuses you of a crime? Do AI companies have the right to train their models on vast archives of copyrighted content that the models can often reproduce on demand? If an AI service copies from a copyrighted work, does it commit copyright infringement? If a fee-based AI service passes off someone else’s work as its own original creation, can it be sued for contractual violation? None of these questions has robust answers yet.
Most importantly to the companies using them: Can the creations of these AI tools be copyrighted? After all, an AI-powered movie generator is of far less value to Hollywood if its resulting blockbusters can be freely copied and resold without copyright protection. Early court cases suggest significant limitations.
So, where does this leave us? In many ways, the current hype around generative AI is very reminiscent of the trajectory of machine translation. Early advances in the 1950s led to wild promises of exponential advances that fizzled out, leading to an “AI winter” of reduced funding and interest. The exaggerated promises, claims of “self-awareness,” and funding free-for-all associated with LLMs today mirror that early period. Yet, as anyone who has attempted to read a user manual machine translated into gibberish from Mandarin can attest, the limitations of technology do not always slow its adoption.
Companies will race to deploy generative AI to replace front-line employees in areas from customer service to intelligent assistants, continuing a longstanding trend. Creative professionals such as artists and illustrators will also face considerable risk, especially in the normally staid world of business communications, as Photoshop automates complex image editing and Microsoft’s PowerPoint offers instant presentations. Creative writing, like advertising and idea generation, will also face great upheaval. In short, anything without a “right answer” that AI can get wrong will be subject to challenge. The idea of replacing expensive humans with cheap AI has proven a temptation that even journalism has found too tempting to resist despite high-profile failures.
In the end, the truth is that today’s generative AI is far more hype than reality. Yet, much like the driverless cars that continue to cause chaos a decade after they were supposed to replace humans, it is also here to stay. The hype will fade as AI transitions into the workplace and collides with the reality of commercial needs, but the cold reality is that this new AI revolution will bring very real job losses to the white-collar creative class.
RCP Today
Top stories on RealClearPolitics
The Takeaway
Polling intelligence and analysis delivered to your inbox
World News
Featuring the latest world news
Policy
Featuring the latest policy news
Support Our Mission
As part of RealClearPolitics, authors aim to provide balanced, non-partisan analysis by presenting top reporting and commentary from all sides of the political spectrum.
Trusted by policymakers and citizens alike since 2000






