AI systems are not neutral technical tools; they are cultural artifacts shaped by the data, design choices, and values of their creators. This fundamental insight is reshaping how researchers and technologists approach building multimodal AI systems that interact with humans across text, audio, images, and video. As artificial intelligence becomes increasingly capable of understanding and generating content across multiple modalities, the question of whose values these systems embed has become urgent.

Why Does It Matter That AI Tools Reflect Cultural Values?

When large language models (LLMs) and multimodal AI systems are trained on vast amounts of digital content, they absorb not just patterns and language, but the cultural assumptions, biases, and values embedded in that content. This process, called "distributed authorship" by researchers, means that decisions made by engineers about which data to use, how to train the model, and what counts as "safe" behavior all get baked into the final system.

"AI tools are not the neutral technical tools of the previous century, that we could calibrate within known operating conditions. They interact with us, we shape their behavior with our prompts. By design, they please us in order to increase engagement. That dynamic is entirely new," explained Andrea Cavallaro, professor at EPFL and head of the Laboratory of Multimodal Intelligent Systems.

Andrea Cavallaro, Professor and Head of Laboratory of Multimodal Intelligent Systems, EPFL

The challenge becomes even more complex when these systems operate across multiple modalities. Hateful content, for example, can be concealed across different channels: in video frames, on-screen text, audio, or spoken words. Sometimes the meaning only becomes clear when you combine these modalities together. Researchers have developed systems that cross-reference all of these simultaneously, but hate speech itself evolves constantly, using coded language, sarcasm, and implicit references that shift over time.

How Are Researchers Embedding Human Values Into AI Systems?

The AlignAI project, an ambitious European Union-funded doctoral network, is training seventeen PhD candidates across six universities to tackle this problem directly. Rather than staffing the project primarily with engineers, AlignAI deliberately recruited social scientists, cognitive psychologists, and philosophers to help characterize and transfer human values into learning systems.

The project tests its approach across three high-stakes domains where the impact of LLMs is already significant:

• Education: Ensuring AI tutoring systems reflect pedagogical values and cultural contexts appropriate for diverse student populations.
• Mental Health: Building systems that understand and respect psychological well-being across different cultural frameworks and therapeutic approaches.
• Online News Consumption: Creating systems that help users navigate information while respecting diverse values around what constitutes trustworthy reporting.

The researchers started with Europe, which is already a very diverse territory, using existing legislation as a starting point. Legislation, they reasoned, embodies the values that a society considered important enough to codify. They even collaborated with a judge to ensure they were capturing the right angle on how values translate into practice.

Steps to Becoming an Active Auditor of AI Systems

Rather than passively accepting AI outputs, users can take concrete steps to engage critically with these tools:

• Question Edge Cases: Test AI systems with unusual or boundary-pushing prompts to see where their values become visible and where biases emerge.
• Probe Value Biases: Ask yourself whose perspective the AI seems to favor. Does it reflect a particular cultural viewpoint, political leaning, or demographic assumption?
• Understand Authorship: Remember that someone made deliberate choices about what data to use and how to train the model. These choices are not inevitable or neutral.

This shift from passive consumer to active auditor is essential because automation bias, a well-known psychological phenomenon, leads people to trust software by default. We assume that because something is code, it deserves our confidence. But that confidence is often misplaced when the code embeds cultural choices we may not agree with or even recognize.

What Does This Mean for Multimodal AI Development?

The expansion of multimodal AI systems, which integrate audio, visual, and textual understanding, is accelerating the need for this kind of value-alignment work. The global AI training dataset market is projected to expand from $8.74 billion in 2025 to $49.82 billion by 2031, with video and multimodal data emerging as rapidly expanding segments. As these systems become more capable and more widely deployed, the cultural values they embed become more consequential.

Real-world applications already demonstrate this challenge. VisionAId, an Android application that turns a smartphone into a visual assistant for people with visual impairment, integrates six on-device deep-learning models running entirely through ONNX Runtime, with optional cloud services for narrative scene description. The system includes a custom detector for Romanian banknotes trained from scratch, reflecting the specific cultural and economic context of its users. This kind of localization and customization is increasingly necessary as AI systems move beyond one-size-fits-all approaches.

The research community is beginning to recognize that designing AI systems that support human flourishing, rather than just maximizing engagement, requires fundamentally different approaches. Many of the PhD students trained through initiatives like AlignAI will go on to build tools that interface directly with humans. Understanding what it means to co-design with the people who will actually use the technology, rather than imposing a techno-solutionist view from above, is becoming a core competency.

The stakes are high. Reliable information reinforces trust and cooperation in societies, while corrupted information erodes these bonds. As AI tools become increasingly capable of generating and circulating text, images, and video that are indistinguishable from human-made content, the question of whose values these systems embed is no longer an academic concern. It is a question that will shape how citizens vote, how patients are treated, and how communities respond to crises.