Researchers developed a technology capable of identifying, through brain signals, when an artificial intelligence (AI) misinterpreted a person's intention. The proposal is to allow AI systems to detect a kind of "this wasn't what I meant" without the user needing to verbalize the correction.
The work was developed by researchers from the Korea Advanced Institute of Science and Technology (KAIST), in collaboration with Microsoft Research Asia. The technology, called Neural Value Alignment (NVA), uses a brain-computer interface to analyze signals obtained through electroencephalography (EEG) and make the AI adjust its behavior in real-time.
The idea stems from a common problem in the interaction between humans and machines: a single human action can represent different objectives, while a single objective can be achieved in various ways.
For example, a person picking up a cup does not necessarily mean they intend to drink from it. They might want to deliver the object to someone, wash it, or simply place it somewhere else. Similarly, if the goal is to quench thirst, a person might take a cup, look for a water bottle, or ask someone to bring a drink.
This ambiguity between goal and action makes it possible for AI systems to interpret an action correctly, but miss what the person really wanted — or understand the goal and choose a different way than what the user expected.
The brain delivers the error
- To solve the problem, researchers decided to observe a reaction that occurs in the brain when something unexpected happens;
- The human brain constantly makes predictions about what should happen. When the result does not match the expectation, certain neural signals may appear. The team focused on two of them: the reward prediction error (RPE, in English abbreviation) and the state prediction error (SPE, in English abbreviation);
- The RPE indicates a discrepancy related to the result or goal. In the system developed by the researchers, it may indicate that the AI did not understand what the person really wanted;
- Imagine, for example, that someone asks a domestic robot for something to drink. The person wanted water, but the robot delivered coffee. The result is wrong, and the brain may produce an RPE signal;
- The SPE works differently. In this case, the goal is correct, but the method used by the AI does not match the expectation;
- If the person wants water and the robot understands this, but instead of picking up a bottle next to her, it goes to the kitchen and fills a glass from the tap, the goal was achieved, but the chosen path was not expected. This difference can produce an SPE.
AI learns to differentiate goal and method
During the experiments, researchers identified different patterns in brainwave activity when an RPE, an SPE, or both occurred simultaneously. Then, they applied deep learning techniques to the signals captured by the EEG. With this, the system was able to classify the person's response to the AI's action.
In practice, the technology can distinguish between two situations: “you misunderstood my goal” and “you understood the goal, but did it the wrong way”. The decoded signals are then used as feedback for the AI's decision-making process.
Microsoft Research Institute was involved in the project – Image: PJ McDonnell/Shutterstock
When it identifies an SPE, the system interprets that the goal is correct, but the strategy needs to be modified. Already a RPE indicates that the own objective was misunderstood and that AI needs to try again to discover what the person intended.
According to KAIST, simulations showed that the method managed to adapt more quickly than existing approaches, even in situations of uncertainty, such as sudden changes in a person's objective or absence of part of the neural feedback.
The study was published in August 2026 in the journal IEEE Transactions on Cybernetics, with the title Neural Value Alignment: Human–AI Collaboration Under Goal–Action Ambiguity.
Read more:
- The challenge of building the next generation of artificial intelligence
- OpenAI is criticized by a former employee who warns about AI risks
- Trump announces “superintelligence force” to coordinate the advancement of AI in the US
Technology can reach robots
Researchers see application possibilities in systems that need to work close to humans.
One of them involves physical robots. A domestic robot, for example, could identify, through the user's brain response, that an instruction was interpreted incorrectly and modify its behavior without waiting for a new verbal command.
The same approach could be applied to industrial robots, autonomous vehicles that need to respond to the driver's judgment, medical and rehabilitation systems aimed at people with speech or movement difficulties, and educational tools capable of adapting their behavior to the student's cognitive state.
According to Sang Wan Lee, a professor at KAIST and the lead researcher of the study, the work demonstrates that AI can go beyond interpreting human intention based solely on observable behaviors. According to him, the system allows the direct use of cognitive signals produced by the brain during human-AI interactions.
Despite the potential, the technology is still far from allowing a person to simply think that a robot is doing something wrong and making the machine stop immediately. The system was demonstrated in simulations , and not in a robot operating freely in a real environment while continuously monitoring a person's EEG and modifying its behavior.
Furthermore, EEG relies on electrodes to measure extremely small electrical signals produced by the brain. These signals can be noisy and are more difficult to obtain outside of controlled environments.
Therefore, it will still be necessary to verify the performance of NVA in natural interactions, where people move and encounter situations much less predictable than those used in laboratory experiments.
Rodrigo Mozelli is a journalist graduated from the Universidade Metodista de São Paulo (UMESP) and, currently, is a writer at Olhar Digital.
