Beware of ChatGPT trying to teach your kids math…or anything

Wednesday marks a day of teachers’ strike across much of the UK, placing parents in the familiar role of pandemic-inspired schoolmaster to their children. Except this time, a magical automated assistant is available to help you.

Educators have cautiously praised ChatGPT, OpenAI’s ultra-sophisticated chatbot, saying it could revolutionize education. One principal in Britain says it sparked a rethinking of homework, while another in Oregon used it to create lesson plans and study guides.

The tool’s personalized responses are what make it so tempting as an omniscient digital tutor. I recently used it to delve deeper into the topic of enzymes, when my 12-year-old son had questions I had no hope of answering. When ChatGPT offered a dense, technical explanation, I asked for simpler terms and an analogy.

“Of course!” he has answered. “Think of a lock on a door. The lock is like an enzyme and the key is like the substrate molecule…” He extended the analogy further to describe the active site of an enzyme as the keyhole.

These were remarkable responses. We could have delved deeper into all facets of biochemistry if we wanted to. Unlike a human tutor, ChatGPT can be queried for as long as you want.

This holds enormous potential for personalized, independent learning… except that ChatGPT is often wrong, and it does a great job of hiding it. When I tested one of my daughter’s English homework questions on the tool, it offered a list of telling examples, which, upon closer inspection, included one that was wildly inaccurate. The main character had a rocky relationship with his parents, the robot explained, even though the character’s parents were dead throughout the book.

On another occasion, I used the tool to generate linear equations for my daughter to practice. She was perplexed when I asked the tool to generate different answers than she had calculated. I asked ChatGPT for an explanation and he once again broke down his method into simple terms, sounding as authoritative as any real math teacher. But when I double-checked the answers on Google, it turned out that ChatGPT’s answers were wrong and my tweens’ answers were correct. Thus ended his mini-nightmare of failing math and much of my initial enthusiasm for ChatGPT.

The New York City public school system, the largest in the United States, has already banned its students from using ChatGPT, in part due to concerns about “content accuracy.” This is why recent comparisons of ChatGPT to a “writing calculator” are a misleading analogy, since calculators are always right and ChatGPT is not right.

How inaccurate is this? An OpenAI spokeswoman said the company had updated ChatGPT over the past two months to improve its factual accuracy, but had no statistics to share. The tool also warns users when they first open it that it sometimes makes errors.

Will this become more precise? Yes, but it is difficult to say to what extent. The large language model that underpins ChatGPT is made up of 175 billion parameters, which are parameters used to make the model’s predictions, compared to the 1.5 billion that its predecessor GPT-2 had. It has become accepted in AI that the more parameters are added to a model, the more truthful it becomes and the correlation is real for GPT. This became much more precise when all these parameters were added. The next iteration scheduled for release this year, called GPT-4, is rumored to be in the billions.

The problem is that we don’t know if a huge increase in parameters also means a huge increase in reliability. For this reason, students should use ChatGPT with caution, if at all, for the foreseeable future.

When I asked Julien Cornebise, honorary professor of computer science at University College London, if he would ever trust it as a homework tool, he replied: “Absolutely not, not yet.” He stressed that even if the system improves, we will still have no guarantee that it is truthful.

Students should get used to corroborating facts shared by the system with other information online or with an expert. Albert Meige, associate director specializing in technology at consulting firm Arthur D. Little, says his own teenage daughter used it to help with her physics homework – but he was able to validate the answers thanks to to his doctorate in computational physics. He recommends using the chatbot to better understand the questions asked during homework. “She discovered she didn’t have to ask a single question,” he said. “It was an interactive process.”

Use it to get feedback, Cornebise concedes. “That’s what the star student will do.”

Being a relatively small company, OpenAI can get away with spewing weird alternative facts. Google and Alphabet Inc.’s Meta Platforms Inc. would not be able to do the same. Google has its own highly sophisticated language model called LaMDA, but is extremely cautious about integrating a similar chatbot into its own search tool, likely in part because of a precision issue. Three days after releasing an AI tool capable of generating scientific papers, called Galactica, Meta withdrew it after academics criticized it for generating unreliable information.

OpenAI will be held to similarly high standards as the generative AI arms race intensifies and chatbot technology is integrated into search engines in the United States and China.

Until then, use it with discretion and a healthy dose of skepticism, especially in education.

© 2023 Bloomberg LP