Some exercise nuts think they’ve found a way to do that: by using the AI chatbot ChatGPT as a sort of proxy personal trainer. Created by OpenAI, it can be coaxed to churn out everything from love poems to legal documents. Now these athletes are using it to make all the relentless running more fun. Some entrepreneurs are even packaging up ChatGPT fitness plans and selling them. 

Its appeal is obvious. ChatGPT answers questions in seconds, saving the need to sift through tons of information. You can ask follow-up questions, too, to get a more detailed and personalized answer. Its chatty tone is ideal for dispensing fitness advice, and the information is presented clearly. OpenAI is tight-lipped about the details, but we know ChatGPT was trained on data drawn from crawling websites, Wikipedia entries, and archived books so it can seem to be pretty good at answering general questions (although there’s no guarantee that those answers are correct.)

So, is ChatGPT the future of how we work out? Or is it just a confident bullshitter?

Work it out

To test GPT’s ability to create fitness regimes, I asked it to write me a 16-week marathon training plan. But it was soon clear that this wasn’t going to work. If you want to train for a marathon properly, you need to gradually increase the distances you run each week. The received wisdom is that your longest run needs to be around the 20-mile mark. ChatGPT suggested a maximum of 10 miles. I shudder to imagine how I’d cope if I ran a marathon that underprepared. I’d be in a whole world of pain—and at serious risk of injuring myself. 

When I asked it the same prompt again in a separate conversation—“Write me a 16-week marathon training plan”—it suggested running 19 miles the day before the race. Again, this would be a recipe for disaster. It would have left me exhausted on the marathon start line, and again, probably with an injury.

I wasn’t sure why ChatGPT gave me two different answers to the same question, so I asked OpenAI. A spokesperson told me that large language models tend to generate a different answer to a question every time it’s posed, adding, “This is because it is not a database. It is generating a new response with each question or query.” Open AI’s website also explains that while ChatGPT can learn from the back-and-forth within a conversation, it’s unable to use past conversations to inform future responses. 

When I asked OpenAI why ChatGPT had given me potentially harmful advice, the spokesperson told me: “It’s important to remind readers that ChatGPT is a research preview— and we let people know up front that it may occasionally generate incorrect information and may also occasionally produce harmful instructions or biased content.”

One of my AI-generated plans wisely offers the caveat that it’s a good idea to check it with a coach. Another tells me to listen to my body and take rest days. Another doesn’t contain any warnings at all. The chatbot’s answers are inconsistent, and not terribly helpful.