Less cloud, more control: a plea for local AI

Artificial intelligence is more than large language models with enormous computing power, states Andries Lohmeijer, founder of KITT Engineering and speaker at WoTS 2026. According to him, the future of AI lies not in bigger, faster, and more, but rather in small, local applications.

“The public debate focuses primarily on generative AI and language models like ChatGPT and Claude. But at its core, artificial intelligence consists of algorithms and pattern recognition,” explains Lohmeijer. “If you understand the technology behind AI, you also see that not every problem requires the brute computing power of a language model. On the contrary: simple, targeted solutions are often more efficient and more reliable.”

The downside of brute force

Lohmeijer knows what he is talking about, as he has been working with artificial intelligence since the 1990s. During that period, the general public first became aware of the possibilities of AI, including when IBM's chess computer Deep Blue defeated world champion Garry Kasparov in 1997. Even then, the human-versus-machine debate led to discussions about the opportunities of AI, but also about the risks.

According to the engineer, those risks are still very much present. “In terms of the environment, for example, because the growing demand for computing power and data places a heavy burden on our data centers. But also in terms of security and privacy, because users are entrusting their data en masse to large technology companies without knowing what happens to it.”

Lohmeijer points to a broader concern in society regarding data processing, ownership, and digital autonomy. “What do companies like OpenAI, Google, and Microsoft do with our data, and who controls that? And what happens if a party suddenly claims access to that data or gains access for itself?”

Bias and reliability

Another problem is bias. AI models make decisions based on input, but that input is rarely neutral. Incomplete or biased training data or human assumptions in the design have a skewed influence on the output. As a result, a system may draw incorrect conclusions based on coincidental correlations rather than the actual cause.

“A computational model has no knowledge,” says Lohmeijer. “It only predicts what is statistically most likely. That is precisely why I advocate for explainable AI, where it is clear how a decision is reached. This is not only important for confidence in the output, but also for control and error analysis. Controllability is essential, especially in technical applications. After all, safety, reliability, and the quality of the input determine whether a system is usable or not.”

Power of local AI

According to Lohmeijer, explainable AI combined with local AI is an important way to mitigate the risks of artificial intelligence. “If you run AI on your own PC or server, you have a better understanding of what happens to the data, who works with it, and how the processes run. Instead of solving a problem with massive datasets and brute force, you make smart use of your domain knowledge.”

He illustrates this with an example from bird identification. “Whereas a large AI model requires hundreds of thousands of images, an expert can arrive at a reliable classification based on a few simple characteristics, such as beak shape, tail, or color. By incorporating that knowledge into a model, a smaller, faster, and more accurate system is created at the local level. Your privacy is protected, you use less energy, and you remain independent of third parties.”

From cloud to edge

In practice, the engineer also sees many applications where local processing works better than a cloud solution. “A client wanted a system that measures congestion in cities. Monitoring via security cameras or smartphone tracking was not an option due to privacy legislation. Moreover, you generate enormous amounts of data that you don't need at all.”

That is why Lohmeijer and his colleagues developed a solution using radar sensors in paving tiles. These so-called 'smart tiles' detect movement locally and provide anonymized information regarding numbers and speeds. ‘Simple and effective. You shouldn't make a project more complex than necessary. Ultimately, the municipal government just wants to know how busy it is.’

Strategic choice

He also sees opportunities for local AI in internal business operations. “A lot of knowledge is locked away in internal documents and designs, sometimes belonging to employees who left the company years ago. That is actually a waste. By running open-source models locally, companies can unlock that knowledge without sharing data with external parties.”

According to Lohmeijer, it is even possible to have various embedded AI systems collaborate locally. “Then you have self-organizing systems where devices exchange information with each other and make decisions together. Think, for example, of smart rain barrels in a neighborhood that keep track of whether a rain shower is coming and whether it is time to drain the rainwater.”

“AI is not an end in itself, but a tool,” emphasizes Lohmeijer. “And like any tool, we must deploy it consciously and ethically responsibly. So take a critical look at where machine learning is truly needed, actively utilize your domain knowledge, and opt for local and explainable AI wherever possible. This way, you benefit from the advantages of AI without losing your autonomy.”

Attend presentation?

Andries Lohmeijer will give a lecture on Thursday, September 24, during the seminar 'Embedded AI for machines and devices'. Register for the presentation and the exhibition free of charge via the WoTS 2026 website.