Devlin Brown at the water cooler: AI not ready for real-world use, but it’s getting there

I asked ChatGPT to design a workout regimen for me, and I am considering following it. Do you think this is a good or bad idea?

In my world, I encounter artificial intelligence (AI) every day. Not the kind of AI we see being displayed in our public offices, the actual smart kind.

If only nonwriters could know how easy it is to tell what has been generated by a machine and what has been written by a human. Sometimes it can be harmless and helpful. Generative AI has most certainly changed the game. It has given subject-matter experts, marketers and all sorts of other business people the power to do analyses, summaries and other tasks in record time.

Those who follow best practice, and who have ethical use guidelines in place, will understand the risks of relying exclusively on AI-generated content. There are built-in biases tied to the data sets the algorithms are trained on, hallucinations, factual inaccuracies or outdated information, the potential for misinterpreted nuances or weak situational awareness, and the content may inadvertently violate things such as privacy or ethical standards. An overreliance on AI puts users at risk of just accepting outputs without critically evaluating what has been generated.

Remember the 2023 Sports Illustrated scandal in which the publication was exposed for publishing AI-generated articles under fake author names? It resulted in executives, including the CEO, being fired. The publisher’s share price plummeted and within a year a third of its workforce was laid off.

I spoke to an executive in the renewables sector last week who brought up AI in the context of decisionmakers potentially using the technology, and then relying on it exclusively to conduct comparative reports on products and solutions. She ran a test for choosing the “best products” for certain contexts, in which the source material cited included blogs and forums that were certainly not robust sources of information, and it didn’t include verified and accurate data housed in formats and in locations on the web the algorithms couldn’t access.

The point was that high-stakes analyses should use AI as a guide, a tool or as assistance, but that deep analyses needed to be conducted by actual experts. Would you consider something related to your health and wellness high stakes?

You can see where I am going with this. I have also played around with AI-generated training programmes. You get the same old routines you find on pretty much every public website. Sometimes the splits and variations don’t make sense, well at least not to me. I once asked for a regimen that did not use machines and I got an answer that included leg extension, leg press and pec dec. Exactly what I didn’t ask for. The recommendations are generally sound: get enough sleep; eat enough protein; use progressive overload; and so on. The kind of generic stuff you can pick up all over the internet.

However, the truth is that we are fiddling around with what is publicly available. Far more is happening behind the scenes that could make this little flurry of AI activity and potential seem like child’s play. Make no mistake, the world in 10 years will be very different.

I read a report about a recently published study on Personal Health Large Language Model (PH-LLM), which is a new AI system that Google researchers unveiled in 2024. The company’s researchers reported that the system’s personal health coaching functionality could interpret wearable device data and then match, and in some instances even surpass, human experts in sleep and fitness advice.

According to the study, PH-LLM outperformed actual subject experts and delivered individualised insights and recommendations that evaluators said was nearly on par with specialist-written recommendations. However, it is not yet perfect. It occasionally produces errors and introduces assumptions. And the biggest weakness, which is mirrored elsewhere, is that its training data does not represent all populations and contexts. It’s not ready for real-world application, but as with everything in this space it is getting closer by the day.

Use the AI tools for recommendations, for thought starters, as they like to say in boardrooms. Perhaps this old Water Cooler hack is a romantic fool, but I cannot believe that an algorithm will ever be able to match one aspect of humanity: the wisdom of experience.