The new chatbot from OpenAI, ChatGPT, gives us a problem: how can we detect whether we read online was generated by a person or an AI?
Since it came out at the end of November, more than a million people have used ChatGPT. It is fascinating to the AI community, and it is clear that AI is making more and more text on the internet. People use it to think of jokes, write stories for children, and write better emails.
Why do we need AI Text Detection?
Irene Solaiman, a former AI researcher at OpenAI who worked on AI output detection for the release of GPT-3’s predecessor GPT-2, says we need to find ways to tell the difference between text written by humans and text written by AI as soon as possible. This is to prevent people from misusing technology.
New tools are also needed to enforce bans on AI-generated text and code, like the one that Stack Overflow, a website where programmers can ask for help. ChatGPT can give you answers to software problems with confidence, but it’s not perfect. If you mess up the code, you could end up with software that doesn’t work right and has bugs. It could be expensive and hard to fix.
A Stack Overflow spokesperson said that the company’s moderators are “looking at thousands of reports from community members using a variety of tools, such as heuristics and detection models,” but would not give any more information.
In reality, it is very hard, and the ban is probably tough to implement.
Detection tools of today
Researchers have tried many different things to find AI-generated text. One common way is to use software to look at how the text flows, how often certain words appear, and if there are any punctuation or sentence length patterns.
“If you have enough text, it’s easy to tell if the word ‘the’ is used too often,” says Daphne Ippolito, a senior research scientist at Google Brain, which is the company’s deep learning research unit. ChatGPT is a branch of OpenAI’s large language model GPT-3. When asked questions, it gives answers that sound surprisingly human.
The illusion of correctness is both the magic and the danger of these big language models. The sentences they come up with look good because they use the right words in the right order. But AI has no idea what it all means. These models work by guessing what the next word in a sentence is most likely to be. They have no idea if something is true, so they confidently present false information as true.
AI tools could make the information we get even more skewed in an online world that is already polarized and politically tense. If they are used in real products, the results could be terrible.
Large language models work by figuring out what the next word in a sentence will be. Because of this, they tend to use common words like “the,” “it,” and “is” rather than weird, rare words. Ippolito and a team of Google researchers found in 2019 research that this is the kind of text that automated detector systems are good at picking up.
But Ippolito’s study also showed something interesting: people tended to think this kind of “clean” text looked better and had fewer mistakes, so they thought a person must have written it.
In reality, human-written text is full of typos and has many styles and slang, whereas “language models rarely make mistakes. Ippolito says, “They are much better at making perfect texts.”
“A typo in the text is a very good sign that a person wrote it,” she says.
Large language models can also be used to find text that AI made. Muhammad Abdul-Mageed, who studies detection, says one of the best ways to do this is to retrain the model on some texts written by humans and others written by machines. This way, the model learns to tell the difference between the two.
Scott Aaronson, a computer scientist who has been working as a researcher at OpenAI, has been working on watermarks for longer pieces of text generated by models like GPT-3.
OpenAI’s spokesperson confirmed that the company is working on watermarks and said that the company’s rules say that users should clearly mark text made by AI “in a way that no one could miss or misunderstand.”
Challenges on AI Text Detection tools.
But there are a lot of things wrong with these technical fixes. Most of them can’t compete with the latest AI language models because they are based on GPT-2 or other older models. Many of these tools work best when there is a lot of text to look at. They will be less useful in some real-world situations, such as chatbots or email assistants, where conversations are shorter, and there are less data to look at.
And Abdul-Mageed says that tech companies don’t let people use large language models for detection because they need powerful computers and access to the AI model itself, which they don’t give out.
According to Solaiman, the larger and more powerful the model, the more difficult it is to develop AI models to figure out what text is written by a person and what isn’t.
“What worries me now is that ChatGPT has really good results. Detection models just can’t keep up. “You’re always trying to catch up,” she says.
Training people to detect AI Text
Solaiman says that there is no easy way to find text written by AI. “A detection model won’t help you find fake text, just like a safety filter won’t help you get rid of biases,” she says.
To have a good chance of solving the problem, we’ll need better technical fixes and more openness about when humans are talking to an AI. People will also need to learn how to spot sentences that an AI wrote.
Ippolito says, “It would be great if Chrome or whatever web browser you use had a plug-in that would tell you if any of the text on your web page was made by a computer.”
There is already some help out there. Researchers at Harvard and IBM made a Giant Language Model Test Room (GLTR) tool. This tool helps people by pointing out parts of text that a computer program could have made.
AI is already tricking us, though. Researchers at Cornell University found that about 66% of the time, people thought fake news stories made by GPT-2 were true.
In another study, it was found that untrained people could only spot text made by GPT-3 at a level similar to random chance.
Ippolito says that the good news is that people can be taught to be better at spotting text that AI has made. She made a game to see how many sentences a computer can make before a player figures out that it’s not a real person. Over time, people got better and better at figuring out that it’s not a real person.
“You can get better at this task if you look at a lot of generative texts and try to figure out what doesn’t make sense about them,” she says. One way is to listen for statements that don’t make sense, like when the AI says it takes an hour to make a cup of coffee.
GPT-3, which came before ChatGPT, has only been around since 2020. OpenAI says that ChatGPT is just a demo, but it’s only a matter of time before more powerful models are made and used in products like chatbots for customer service or health care. And that’s the crux of the problem: because this field is changing so quickly, every way to spot text that AI made becomes old very quickly. We’re losing the arms race right now.
