Can we make AI reliable? Victoria Palacin connects VTT’s AI research with the demands of the real world

Victoria’s path has moved between academia, industry and applied research. She holds a doctorate in human-computer interaction from LUT University, where she studied how human values impact crowdsourcing applications, working with environmental agencies in Finland, the US and Japan. She has held research positions at the MIT Media Lab, the University of Toronto, the University of Helsinki and the United Nations. Later she moved into industry, working as a technology consultant and leading a data and AI business unit.

This combination of research and business experience is particularly relevant in applied AI. Technologies such as sensors, machine learning models and LLM applications can create new services and solve practical challenges, but only when they are trusted by their users. Trust is a fundamental requirement for AI applications to scale and create impact in industries and societies. And trust depends on the team behind the technology, truly understanding the human problem they are meant to solve, and implementing mechanisms for human and automated transparency. 

“I like applied research because I like real-world problems. I like innovative ways to solve them, and I want to create impact that matters for society,” Victoria says.

Asked what matters most in applied AI right now, Victoria points to three things: evaluating whether AI can be relied on, using synthetic data where real data falls short, and applying AI beyond the obvious.

One of the team’s key areas of expertise is AI evaluation. As AI-enabled systems become more widely used, it is not enough to deploy them quickly. Organisations also need to understand how well they work, whether they are reliable and how they perform in the tasks humans expect them to support.

This is one of VTT’s strengths. The team evaluates AI not only from a deep technical point of view, but also in relation to purpose, context and human intention. This is especially important for large language model deployments, where trust, transparency and performance need to be assessed carefully.

VTT’s own AI infrastructure and expertise in automation and connectivity also create a strong foundation for developing and evaluating AI in a European context. Especially when organisations are working with sensitive, fragmented or scarce data, solutions need to support data sovereignty, privacy and ethical use.

A major challenge for AI development in Europe is the availability and quality of data. Many organisations have data that is confidential, fragmented or difficult to share. At the same time, AI models depend on the quality of the data underneath them. Poor data can lead to poor performance, and, in the worst cases, decisions based on assumptions that cannot be trusted.

Synthetic data offers one way to create new room for applied AI in Europe. When carefully designed and governed, it can help organisations overcome data-related bottlenecks, improve algorithm performance and create domain-specific datasets for testing and development. It can also make it possible to explore rare or difficult-to-reach scenarios that would otherwise be challenging to study.

The key is that synthetic data must remain purpose-driven and human-governed. It should be based on real data or scientifically grounded knowledge, and it should be validated not only technically, but also against the human use cases it is meant to support.

“We are not creating synthetic data for the sake of it. The human is in control, and the data needs to serve a clear purpose,” Victoria says.

By this point most of us have tried some sort of generative AI assistants powered by LLM (large language models). VTT’s Reliable Intelligence team has been applying generative AI in new ways.

This is where Victoria's enthusiasm for applied AI becomes particularly visible. For her, generative AI becomes interesting when it leaves the familiar chatbot setting and starts solving problems in unexpected places, from plant research and satellite monitoring to remote machinery maintenance.

What these applications share is that the AI supports human experts instead of replacing their judgement – and this is a must in environments where humans cannot go or quickly evolving situations that challenge our cognition. AI changes power dynamics, decision-making and the role of humans in digital systems, and these questions need active attention as the technology develops. At VTT, human-in-the-loop approaches, transparent evaluation frameworks and responsible use of data are all part of building AI that can be trusted in real applications.

For Finnish industry, this opens opportunities beyond efficiency. Growth can come from new services, products and forms of value.

“Growth does not come only from the usual ways of doing things. We need to test, try and bring in new ideas, but we also need to do it in a responsible manner. Technology is not the answer to everything. It is a tool, and purpose is important,” she says.

VTT’s newly launched Physical AI Research Area, one of the largest physical AI labs in Europe, brings together nine teams working on situational awareness (sensors, machine vision and hyperspectral imaging), AI-based reasoning (LLMs, vision-language models and spatial computing) and autonomous machines and platforms (robots, autonomous vehicles and drones).

The Reliable Intelligence team that Victoria leads is part of VTT’s physical AI unit. The team members have developed numerous methods for human-AI evaluation, hallucination detection and technology assessments, as well as invented new foresight technologies. They actively advise organisations and governments how to deploy and evaluate AI in their own environments.

Asked about the future of AI, Victoria sees foundation models for robotics as the next major development in the field, with the potential to change everyday life the way generative AI changed knowledge work. In the future, machines will operate among us, from cute assistant robots to autonomous vehicles. 

As we speak, companies across the globe are racing to make general physical AI models. But the ones who can make human-machine collaboration reliable, will be the ones that will truly change the world. People will not work next to machines they do not trust.

“In the real world, a model that is usually right is not good enough. You can ignore a bad answer from a chatbot. You cannot ignore a robot that gets it wrong next to you.” Victoria says.