Is it human or AI? New tools help you spot the bots

Researchers and other programmers have taken it upon themselves to build tools to help people figure out what has sprung from the mind of a human, and what was cobbled together by a bot. But in a period of rapid advancement such as this, any tool can have a hard time keeping up.

Daniel Morgan, a 39-year-old father of two, has experimented with OpenAI’s ChatGPT text bot a few times, writing thank-you notes, crafting bedtime stories and developing marketing material.

Mr. Morgan, who has dyslexia, says he has used ChatGPT to help him write blog posts for his real-estate investment and brokerage company, Marterra Group.

“Now I can get those content ideas mostly fleshed out, give them to someone on our team and then have them customize it and not have to worry about feeling down because I’m misspelling things or my grammar’s off," he says.

Bots from OpenAI—including its ChatGPT chatbot that can create written content, and its Dall-E 2 art engine—are part of a growing wave of tools that can generate realistic work that’s difficult to discern from that made by humans. Other AI tools create “deepfake" videos that can produce footage of words and actions that were never actually filmed.

While Mr. Morgan has told some of colleagues that he’s getting help from ChatGPT, others might not be so clear. Here are ways to try to identify AI-generated content, so you don’t get fooled by a robot.

Digital sleuths

ChatGPT has spurred fears that students may use it to write essays and other writing assignments—though teachers are still apt to spot errors and other mistakes. New York City’s Department of Education, for example, recently banned access to the product on its networks and devices. People in other industries are also worried workers might use it as a shortcut for work assignments.

Edward Tian, a 22-year-old student at Princeton University, built GPTZero earlier this month to address the growing concern that people may not know when something has been written by machines. It’s simple to use: Copy and paste any text you suspect was generated with the help of AI.

GPTZero shows you the likelihood of how fake or real the text is. The software evaluates text based on a handful of factors. One key metric it uses is the tendency of people to use a higher variation in word choice and sentence length, while text from AI is more consistent.

Mr. Tian says he doesn’t oppose people using AI to support or enhance their work. “But there are qualities of human writing, really beautiful and raw elements of written prose, that computers can’t and should never co-opt," he says of why he built GPTZero.

Hugging Face, a company that develops AI machine-learning tools, has a similar website it started in 2019. Drop in about 50 words of text, and it will serve up a percentage result of how real or fake it is.

Both tools have limitations and will require updates to keep up with AI advances. Hugging Face’s tool is trained on GPT-2, an older version of OpenAI’s text engine, and will label writing from OpenAI’s current GPT-3 text engine as real. GPTZero serves up results based on GPT-2 calculations. GPTZero might not be as good at spotting GPT-3 content, but it still offers a better assessment of whether the writing is real or generated.

Brain training

Skepticism is equally as important as a detection method, says Irene Solaiman, policy director at Hugging Face. People can look for signs such as repetition or inaccuracy to indicate that what they’re reading or seeing was AI generated, she says.

“Sometimes you can tell with a language model that it’s misunderstanding modern data, misunderstanding time frames," Ms. Solaiman says.

You can train your own eyes and brain for bot-detection, especially for images and videos that use AI generated content, including deepfakes.

Start with a research project called Detect Fakes, co-created by Matt Groh, a 34-year-old Ph.D. candidate at Massachusetts Institute of Technology’s Media Lab. The exercise will ask you to determine which of the 32 text, image and video examples are real and which are deepfaked using AI.

The team behind Detect Fakes recommends other ways to detect a deepfake as well, such as paying attention to whether someone’s lip movements look real or a bit off—a sign that something is amiss. It can also be helpful for people to take a step back and think about why such an image or video exists.

“They can pay attention to the context, how likely this is given what they know about the world," Mr. Groh says. “And they can pay attention to incentives and what someone is saying and why someone might be saying this." There’s no magical way to detect all deepfakes, though, he says.

A never-ending battle

Other companies and universities are working on detection tools for AI-generated images and videos.

Intel released FakeCatcher in November. The tool looks for indications of blood flow—slight changes in coloration indicative of typical biological processes—to classify a video as fake or real. It is available now to some companies, including news organizations and social-media companies, says Ilke Demir, a senior staff research scientist at the company.

But future AI generators will most likely find ways to fool authentic markers such as blood flow in the face, Dr. Demir says.

As AI tools proliferate, the danger will be in relying on a single model or approach to spot them, she says.

A better solution is a platform that can combine a variety of different results to determine how authentic a piece of content may be, Dr. Demir says. “It will be more trustworthy because you are not saying we have one algorithm that we’re trying to conquer," she says.