“This sentence was written by an AI.” Or maybe not? OpenAI’s new chatbot, ChatGPT, presents us with a whole new problem: how are we supposed to know whether what we read online was written by a human or a machine?
Since its release in late November, ChatGPT has been used by over a million people. The AI community is enthusiastic – and it seems clear that the Internet will be increasingly flooded with AI-generated texts. Many people now use ChatGPT to generate jokes, create children’s stories or better emails.
ChatGPT is an easy-to-use offshoot of the large language model GPT-3 from the company OpenAI, funded by Elon Musk, Microsoft and other corporations. The system can generate remarkably human-sounding responses to questions put to it. The magic – and the risk – of these large language models is that they provide an illusion of correctness. The sentences that the algorithms produce look real at first—they use the right parts of speech in the right order. But the AI just doesn’t know what all that is generated means. Because the models work with a prediction of the most likely next word in a sentence. So you have no idea if something is right or wrong. What’s more, they confidently present their information as true, even when in reality it isn’t.
This creates a great danger: In an online world that is already extremely politically polarized, such AI tools could distort the content of what we consume even further. Once deployed in the real world in real products, the consequences could be devastating. We therefore urgently need ways to differentiate between human-written texts and AI texts to counteract the potential misuse of the technology. That’s what Irene Solaiman, policy director at the AI start-up Hugging Face, says, who used to be an AI researcher at OpenAI and had already dealt with the detection of AI content when the GPT-3 predecessor GPT-2 was released .
Bans must be verifiable
New tools are also needed to enforce bans that are intended to exclude AI-generated texts or AI-generated program code in the future. The topic has long been topical. So bans the programming website Stack Overflow, where coders can ask for help, now output from ChatGPT. Because the system can also spit out answers to software problems, but it is by no means foolproof. And incorrect code can then lead to buggy and broken software that later becomes expensive and potentially messy to clean up.
A spokesperson for Stack Overflow says the company’s moderators have now examined “thousands of responses submitted by community members.” A number of tools were used, including recognition models and special heuristics. However, the company does not want to give any details. And in reality, enforcing such a ban is incredibly difficult – if not impossible.
There is already some methods, which researchers have used to try to recognize AI-generated text. A common technique is to use software to analyze various characteristics of the text – for example, how smoothly it reads, how often certain words occur, or whether there are patterns in punctuation or sentence lengths. “If you have enough text, a simple indication is that the word ‘the’ occurs too often,” says Daphne Ippolito, senior scientist at Google Brain, the search engine company’s deep learning research unit.
Because large language models predict the next word in a sentence, they tend to use common words like “the”, “it”, or “is” instead of less common words. This is exactly the kind of text that automatic recognition systems would be good at recognizing, like Ippolito and a team of researchers from Google in one study published in 2019 have figured out. However, Ippolito’s study also revealed another piece of news: the human participants tended to think that this type of computer-generated “clean” text looked better and was more correct. Result: They thought it came from humans.
Error rather human, perfect text from the AI
But in reality, texts written by humans tend to be full of typographical errors while at the same time being highly variable. They often contain different styles and slang terms. “Language models, on the other hand, very, very rarely generate typos. They’re much better at generating perfect text,” says Ippolito. And: “A typo in the text is actually a very good indicator that it was written by people.”
Large language models can also be used to recognize AI-generated text themselves. One of the most successful methods is to train the model on some human-written text and some machine-written text so that it learns to distinguish between the two, as explained by Muhammad Abdul-Mageed, Canadian research chair in machine learning at the university of British Columbia and has studied AI text recognition.
Scott Aaronson, a computer scientist at the University of Texas who was a researcher at OpenAI himself for a year, has also developed a form of watermarking for longer text generated by models such as GPT-3. “It’s an otherwise unobtrusive, secret wording signal that can later be used to prove it’s from GPT,” he explains in his blog. A spokesman for OpenAI confirmed that the company itself is working on watermarking for computer-generated text. In addition, its terms of service stated that users should clearly label text generated by the AI: “In a way that no one can overlook or misunderstand.”
However, the previous technical solutions are by no means perfect. Most don’t stand a chance against the latest generation of AI language models as they are still built on top of GPT-2 or other earlier models. And many of the recognition tools work best when there is a lot of text available; in some concrete use cases such as chatbots or email assistants, which are designed for shorter conversations and provide less data for analysis, they are not very efficient. And the use of large language models for AI recognition also requires powerful computers – plus access to the AI model itself. And so far, according to Professor Abdul-Mageed, tech companies are not granting that.
Train the human eye
The bigger and more powerful the models get, the more difficult it is to create AI models that can recognize which text was written by a human and which wasn’t, says Hugging Face expert Solaiman. “What’s worrying is that [ChatGPT] now delivers really impressive results. The detection models just can’t keep up. They have to catch up all the time,” she says.
And there’s no silver bullet for recognizing AI-written text, Solaiman said. “A recognition model is not the solution to synthetic text recognition, any more than filters are the solution to bias problems.” (By this we mean the tendency of AI systems to also reproduce racist and otherwise problematic content.) In order to have a chance of solving the problem, we need better technical solutions and more transparency in relation to when people interact with an AI . This means: We have to learn to recognize signs of AI-written sentences ourselves.
“It would be really nice if there was a plugin for Chrome or any other web browser that could be used to check whether text on a webpage was machine-generated,” says Ippolito. Some help is already available. Researchers at Harvard University and IBM have created a tool called Giant Language Model Test Room (GLTR) that helps people find problematic passages. It highlights those that could have been created by a computer program.
But AI systems have been fooling us again for a long time. Cornell University researchers found that people believed fake news articles generated by GPT-2 — mind you, the predecessor of GPT-3 — to be credible about 66 percent of the time. Another study found that untrained people could only recognize text generated with GPT-3 to a degree that amounted to coincidence.
Training and playing help with recognition
The good news is that humans can be trained to recognize AI-generated text better, says Ippolito. She herself has developed a game to test how many sentences a computer can generate before the player realizes that it is not a human. She has found that players only get better over time.
“If you look at a lot of computer-generated text and try to figure out what doesn’t make sense in it, you can also get better at this job,” she says. One way is to detect implausible statements, like an AI claiming that it takes 60 minutes to brew a cup of coffee.
GPT-3, the foundation of ChatGPT, has only been in use since 2020. According to OpenAI, ChatGPT is a demonstrator, but it is only a matter of time before similarly powerful models are developed and integrated into products such as freely available chatbots for use in customer service or healthcare.
And that brings us to the heart of the problem: The enormous speed of development in this sector means that any methods of recognizing AI-generated text very quickly look very old again. So it’s a race – and right now it looks like we’re going to lose.
(jle)
