Skip to content

“AI is more of a very dumb helper.”

Pieter Marsman

Machine learning engineer

THE 3 MOST COMMON MISCONCEPTIONS ABOUT AI



If it were up to him, Pieter would spend every working minute at his computer programming. “What I do is annotate all the Orbi photos,” he explains. “I teach a machine learning model to determine what food is seen in a photo. Basically, we add information to the photo to create pure data for the Orbisk dashboard. That gives kitchen staff insights into their waste flows, such as if they throw away mainly sandwiches, kitchen waste, or products in their original packaging.” There are quite a few misconceptions circulating about machine learning models, the AI technology used by Orbisk. These are the three biggest according to Pieter:

MISCONCEPTION 1: AI IS MAGICAL

“It seems magical when ChatGPT produces a text, but AI is just a computer that follows programmed rules. Nothing else. It works very deterministically: if this, then that. But because the process is so layered and complex, the results are often mind-blowing.

Some results even I find magically good. When I’d just started working at Orbisk and the model said ‘this picture shows food’ correctly for the first time, it was incredible. Wow, this can really work, I thought. I also experienced such moments when the machine learning model recognized a food correctly for the first time, and recently when it managed to correctly estimate the weight of food waste based on a picture.”

MISCONCEPTION 2: AI CAN THINK FOR ITSELF

“It gives me the jitters when the words ‘smart’ and ‘AI’ are used in the same sentence. People think AI is a robot with a humanlike brain, but actually it’s more of a very dumb helper. Machine learning always delivers the same output; it repeats the examples we provide. The system can’t think for itself about what it sees in a picture. “By performing simple tasks hundreds, thousands, millions of times in a row, you steer the model ever so slightly in the right direction. The machine learning says: ‘I think this picture shows brown sauce’, for example. And then we say yes, that’s correct, but actually the sauce is more to the left. That’s how it gets better and better.”

MISCONCEPTION 3: AI IS A ‘FINISHED’ TOOL

“I prefer the term machine learning model, which encompasses the infinite growth that’s still possible. Our AI tool is always changing and improving. We started by automating the process of determining whether or not there’s food in a photo. Then we taught the model to recognize specific foods. Fries are an easy food for the system to recognize, because they always look the same. So are squeezed oranges.

“Foods like chicken and mixed vegetables have much more variation between them, so those are more complicated. Or take a plate with a main course: for each element the model has to determine this is where food is located, this is a bratwurst, this is mashed potatoes and these are peas. That’s a difficult task. So we’re now teaching the model to recognize foods ever more accurately by giving manual feedback on the food determinations.”

SEND AN EMAIL

GIVE US A CALL

ADDRESS

Kanaalweg 29A, 3526 KM Utrecht

Homepage