A sociologist demonstrates how numerous unseen individuals create the magic.
By John P. Nelson, Georgia Institute of Technology Postdoctoral Research Fellow in Ethics and Societal Implications of Artificial Intelligence
The media hysteria surrounding ChatGPT and other large language model AI systems covers a wide range of topics, from the mundane—large language models could replace traditional web search—to the alarming—AI will eliminate many jobs—and the overwrought—AI poses an extinction-level threat to humanity. Large language models presage the advent of artificial intelligence that will surpass human intelligence, which unites all of these subjects.
However, despite their complexity, massive language models are actually quite simple. Furthermore, although having the moniker “artificial intelligence,” they are totally reliant on human wisdom and work. Of sure, they cannot consistently produce new information, but there is more to it than that.
Without humans providing new content and instructing it on how to understand it, not to mention programming the model and creating, maintaining, and powering its hardware, ChatGPT cannot learn, improve, or even remain current. You must first comprehend the operation of ChatGPT and comparable models, as well as the part that people play in their operation, in order to comprehend why.
How ChatGPT works
According to training data sets, large language models like ChatGPT generally estimate which characters, words, and sentences should come after one another in a sequence. In the instance of ChatGPT, the training data set includes vast amounts of text that has been collected from the internet in the public domain.
ChatGPT works by statistics, not by understanding words.
Imagine I used the following collection of texts to train a language model:
Bears are enormous, hairy creatures. Grizzlies have claws. In reality, bears are robots. Bears possess a nose. In reality, bears are robots. Bears occasionally eat fish. In reality, bears are robots.
Because those terms occur the most frequently in its training data set, the model would be more likely to suggest that bears are actually hidden robots than anything else. Models trained on inconsistent and imperfect data sets, which includes scholarly literature, are plainly a concern.
People write lots of different things about quantum physics, Joe Biden, healthy eating or the Jan. 6 insurrection, some more valid than others. How is the model supposed to know what to say about something, when people say lots of different things?
The need for feedback
In this situation, input is useful. You can choose whether to rate responses as good or negative if you utilize ChatGPT. If you give them a bad rating, you will be required to give an example of a good response. Through user feedback, input from the development team, and input from outside companies hired to label the output, ChatGPT and other large language models learn what responses, what projected text sequences, are good and terrible.
ChatGPT is unable to independently compare, analyze, or assess claims or data. It can only produce text sequences that have been utilized by others in comparison, analysis, or evaluation, preferring to produce sequences that resemble those that it has already been taught are appropriate responses.
As a result, when the model provides you with a good response, it is drawing on a great deal of human effort that has already been put into defining what constitutes a good response. Behind the screen, there are countless numbers of human workers, and these workers will always be required if the model is to advance or broaden its content coverage.
According to a recent investigation by journalists for Time magazine, thousands of hours were spent by hundreds of Kenyan workers reading and labeling disturbing writing from the darkest corners of the internet, including graphic descriptions of sexual violence, in order to teach ChatGPT not to copy it. They were paid no more than US$2 an hour, and many understandably reported experiencing psychological distress due to this work.
Language AIs require humans to tell them what makes a good answer — and what makes toxic content.
What ChatGPT can’t do
The value of feedback is seen in ChatGPT’s propensity to “hallucinate,” or confidently give incorrect responses. Despite the fact that there is a wealth of reliable material about a certain subject online, ChatGPT is unable to provide insightful responses without training. By posing increasingly esoteric questions to ChatGPT, you can test this out for yourself. Because the model appears to have been more carefully trained on nonfiction than fiction, I’ve found it to be especially helpful to ask ChatGPT to explain the narratives of several fictional books.
In my testing, ChatGPT accurately summarized “The Lord of the Rings,” a well-known book by J.R.R. Tolkien, with only a few errors. However, its summary of “The Pirates of Penzance” by Gilbert and Sullivan and “The Left Hand of Darkness” by Ursula K. Le Guin—both slightly more specialized but still not obscure—come very close to playing Mad Libs with the names of the characters and locations. It doesn’t matter how well-written the Wikipedia pages for each of these books are. Not simply content, but feedback is required for the model.
Large language models rely on people to understand and assess information for them because they are incapable of doing so alone. They prey on human work and knowledge. They require new training on whether and how to construct sentences based on new sources when they are added to their training data sets.
They are unable to determine whether or not news reports are accurate. They are unable to analyze arguments or compare trade-offs. They can’t even accurately recount a movie’s plot or read a page from an encyclopedia and just make remarks that are compatible with it. They rely on someone to complete all of these tasks for them.
Then they paraphrase and remix what people have stated, relying on additional people to evaluate their paraphrasing and remixing skills. They will require substantial retraining to embrace the new consensus if the conventional thinking on a given issue changes, such as whether salt is hazardous for your heart or whether early breast cancer tests are beneficial.
Many people behind the curtain
Large language models, in short, show how many AI systems are completely dependent on their creators, maintainers, and users, rather than being the forerunners of fully autonomous AI. Therefore, if ChatGPT provides you with a good or useful response to a question, keep in mind to thank the thousands or millions of unseen authors whose words it parsed and who taught it what constitutes good and bad information.
ChatGPT is not an autonomous superintelligence; it is nothing without humans, just like every other technology.
The Conversation is an independent nonprofit news source devoted to disseminating the insights of academic authorities. Learn more about us or up to receive our weekly highlights.
Did you find this article to be enjoyable? If so, congratulate them and select “follow” from the menu in the top right corner of this page. If you do that, our articles—all of which were produced by qualified researchers—will begin to appear in your feed.
John P. Nelson has revealed no relevant relationships beyond their academic appointment, and says they are not employed by, consult for, own shares of, or receive funding from any company or organization that would benefit from this publication.
