AI for everyone? 

Imagine a video meeting where everyone is framed perfectly, except for one person.
Not because the camera failed, but because the underlying machine learning model didn’t recognize them. 

That should not happen.

AI is evolving fast. What started as experimentation is now deeply embedded in products, workflows, and everyday decisions. 

But in all the excitement around models, capabilities, and speed, one thing is easy to overlook:
AI is only as good as the data behind it.

The part we don’t talk about enough 

In machine learning, we often talk about models. Which architecture to use, how to optimize performance, and how to scale. 

But the real differentiator isn’t necessarily the model. The real differentiator is the data. 

What data you use to train the model.

And how representative that data actually is. 

If your data is biased, incomplete, fabricated, or skewed, your model will most likely be too.
Simple as that. No matter how much you tweak it. 

What this looks like in practice 

At Cisco Norway, we build machine learning models that run on the edge of our video conferencing solutions. Everything happens in real time, on the device itself. 

That means our models need to be small, fast, and precise. But more importantly, they need to treat everyone in the room equally. 

We have seen models struggle with something as simple as detecting the back of someone’s head in a meeting, especially for women. Why? Because female hair is more diverse. Different hairstyles, lighting conditions, skin tones and other variations can all impact how well a model performs. 

And if your dataset doesn’t reflect that diversity, your model won’t either.
It might be a small detail. But in practice, it affects how well the technology works, and for whom.  

A human problem, not just a technical one

When we think about the future of AI, we don’t just think about performance or innovation. We think about responsibility. And an important distinction in today’s AI landscape is building AI versus using it. 

If you are building models yourself, like we do, you have access to the data. You can adjust, improve, and iterate. You can actively work to reduce bias. 

But if you are using third-party models, you often don’t know what data was used to train the model. 

That doesn’t mean you shouldn’t use third-party models, but it does mean you need to be conscious about what bias they might have.

You need to test. Question. Validate. 

As AI becomes more integrated into our everyday tools, the impact of this grows. It’s no longer just about whether a model works. It’s about who it works for and how we, the people behind training AI models, ensure it works for everyone.

Because in the end, AI reflects the world that we show it. 

Now, AI is increasingly influencing that world, by generating data, building software, and shaping the systems we rely on. That makes it even more important to be conscious of what behaviour we are creating and reinforcing 

If we want technology to work for everyone, the data behind it needs to reflect everyone too.

The article above is written by Kari Mo who works at our partner Cisco.