By Nik Kairinos
Training AI-powered robots with human intelligence is crucial to ensure they operate effectively and safely.
Artificial intelligence has already become embedded in many of the systems people use every day, but for the most part it still exists behind screens, at a distance from the real-world consequences of its outputs.
Robotics represents a new frontier for AI, where outputs are directly linked to the physical world. As AI begins to power machines operating in workplaces, hospitals, warehouses, public spaces and, increasingly, domestic environments, it will no longer be enough for these systems to produce convincing answers. They will have to behave appropriately in the presence of people, and to operate safely in environments that are dynamic, physical, and often unpredictable.
How do the risks of AI differ in the real world?
That shift will raise the standard for what we mean by a capable AI system. In a software environment, an output can be inaccurate and remain relatively contained. In a physical setting, the consequences are more immediate; robots operating in care, logistics, or public interactions are actively participating in situations where timing, context, and conduct all matter.
We have already seen early examples of this gap in practice. Clips of robots behaving erratically in controlled or low-stakes environments are often treated as amusing, but they point to a more serious underlying issue: what is harmless in one context can quickly become dangerous in another.
The environments these systems operate in are not neatly structured. They involve incomplete information, changing conditions, and human behavior that does not follow consistent patterns. Responding effectively in those situations requires more than prediction. It requires a level of judgement that reflects how people interpret and act within real-world contexts. In short, it calls for human intelligence.
Why is human intelligence so essential?
Current AI systems have made significant advances because they are highly effective at identifying patterns across large datasets. That progress is real, but it does not necessarily translate into behavior that is appropriate outside controlled environments.
A model may generate a plausible response or select a likely course of action without any robust understanding of why one option is more suitable than another in a given situation. This becomes more pronounced in robotics because physical environments depend heavily on forms of reasoning that are rarely captured in conventional training data.
Human decision-making is complex. It is shaped by context, experience, and the ability to interpret subtle signals that do not appear in structured datasets. Tailoring a response to a specific, unique circumstance is not easily derived from examples alone. These are central aspects of behavior, and they are precisely the areas where current systems remain limited.
For that reason, improving performance through additional data and scale is unlikely to be sufficient on its own. Continuing to rely on more data and compute alone is unlikely to deliver the level of reasoning required for real-world environments.
What needs to change?
If robotics is to develop in a way that is both effective and widely trusted, AI training will need to incorporate more direct forms of human input. This does not mean replacing existing approaches, but extending them to include how people reason and apply judgement in practice. Rather than relying primarily on static datasets, systems will need to learn from the processes that underpin human decision-making.
This requires a more deliberate approach to how AI systems learn. Instead of asking models to infer behavior indirectly from large volumes of data, we need to prioritize teaching AI more explicitly. In practical terms, this points towards a model in which humans play a more active role in shaping how AI systems evolve.
Capturing reasoning in context, rather than just outcomes, becomes essential if systems are expected to operate in environments where behavior matters as much as capability. Rather than just keeping a human in the loop, we need to leverage human intelligence as a teaching tool to directly instil human reasoning, contextual judgement, and real-world behavioral insight into AI systems.
There is some understandable concern that making AI more human-like introduces new risks, but in the context of robotics, the more immediate challenge is ensuring that systems deployed in human environments are able to act with an appropriate level of judgement from the outset.
Smarter is better
Systems that interact directly with people cannot rely on pattern recognition alone. They must be able to respond in ways that are consistent with human expectations and behavior. This change cannot be achieved with only advancements in models and raw compute. Instead, we have to learn to treat human intelligence as a resource alongside data and processing power. Until we can do that, we risk deploying AI into the physical world without the level of judgement those environments demand.
Nik Kairinos
