The truth about using large langue models like ChatGPT is that they’re not trained to address outlier scenarios. There’s no one at OpenAI thinking specifically about how to respond when asked for advice in a sport that is one of the most dangerous sports out there.

In my day job, I work in mental health where AI safety is paramount. Platforms like Character.AI have unfortunately led to suicides in teenagers because they don't have the safeguards to protect young individuals from sensitive conversations. Recently, OpenAI has hired a number of people in the mental health industry to better respond when individuals start expressing themes of depression or suicide. Harvard Business Review says that emotional and mental health conversations are one of the most commonly had discussions with ChatGPT. Your everyday accessible AI platforms are not safe to handle these conversations, yet there are many platforms built by healthcare experts to address these challenges. While they’re not fool-proof they are leaps and bounds above the generally available platforms.

This is an example of where industry or use case specific safeguards can be built to address these outlier scenarios. Pole Vault is no different. We can train AI to be able to respond to these specific use cases in a safe manner.

If you think about it, coaching is quite similar to how you program, coaching is a lot of if-then statements. If a pole vaulter is over-bending a pole, then you probably need to move them to a stiffer pole if they're landing safely within the pit. If a pole vaulter got rejected, then you probably need to move them to a softer pole or a lower grip. If a pole vaulter's step is under and they are way over striding, then they're probably overcompensating from having an out step.

Now, having been a vaulter for years and coach for several more years, pole vault’s not always that simple. I’ve been rejected more times than I count, broken 4 poles in my career, and hugged a few standards mid-air in my day. It’s what happens when you take tens of thousands of jumps working your way up to poles the size of a 16’ 190lb.

One thing that you're probably familiar with ChatGPT not being very good at is when to stop. You can ask it a question that it might not know the answer to, and it will try to fill in the blanks or do its best, rather than just say, "I don't know."

An AI coach for pole vault does not have to act that way. If AI detects a critical safety issue, like someone getting rejected back onto the runway, then AI can recognize an unsafe jump by maybe the step being way too far under or having grip height that's way too high. Rather than refer to specific instructions, such as lowering the grip height by three inches, the AI can recommend deferring to a human coach and recommend that the athlete seeks in-person guidance.

This is a great question to ask. After watching a video, what if the AI tells the athlete to move their step in 6 inches when really they should move their step out?

Have you ever been to a high school track meet? You will see this happen every weekend. A coach simply looks for the takeoff spot and sees an athlete is under but doesn’t watch them over-stride the last half of the run because they’re stretching to get there. The athlete is told to move out only to be forced to over-stride more.

The great thing about AI is it can analyze the jump as a whole as long as you give it a video of the whole jump, so before making any recommendations it will synthesize all the data.

When moving grip height and poles, there’s typical “safeguards” that a coach would follow that AI can be instructed to the same. For example, if an athlete is using a 13’ 150lb pole, a coach would never tell them to jump to a 14’ 160lb if they blew through. Most likely you’re moving to a 13’ 155lb and maybe changing the grip an inch or two at a time. These limitations like moving one pole at a time and grip heigh a few inches at a time can be included in the safeguards; therefore, if any changes are made they are limited in scope.

So, an athlete or coach still learning the ropes of pole vault should not try to use ChatGPT as a coach. But a pole vault dedicated tool trained by people with deep domain expertise can be built in a way to keep athletes of all skill levels training safely.