What is Human-Centric AI?

17/04/2024 - With great regularity, you hear the term Human-Centric AI (HCAI) or Human-Centered AI (HCAI) being used. During these discussions, it becomes apparent that there is often a different interpretation of the concept of HCAI. The consensus, however, is that HCAI contributes to an ethical use of AI. Sometimes HCAI and ethical AI are even seen as synonymous. In the blog below, a number of elements are discussed that are often mentioned as (essential) parts of HCAI and thus contribute to an ethical use of AI.

AI as a complement to humans

The idea behind HCAI is that it enhances and enlarges human capabilities instead of trying to replace them. In the implementation of HCAI, human control is maintained in a way that ensures AI meets our needs. In practice, this means that an AI system does not take over the role of the doctor in the hospital, but rather supports the doctor with fast and accurate analyses.

Transparency and explainability

Transparency can have different meanings. The transparency discussed here is different from, for example, the transparency requirements under the GDPR. Here, transparency means opening up the processes, decisions, and operations of the AI system. Although transparency may seem obvious to achieve HCAI, practice proves challenging. Increasingly complex AI models make it more difficult to arrive at an outcome transparently. Additionally, large models from organizations like OpenAI are only transparent to a limited extent about the development and creation of these models. This also shows that transparency is a layered challenge when it comes to HCAI. Not only should there be transparency about the development, the choices made, and the training data used, but also about how an AI model arrives at an outcome. And you could argue that transparency in itself does not directly lead to HCAI. Because when an AI model is transparent about how it arrives at an outcome, it is not necessarily understandable or explainable to the person who has to make decisions based on that outcome. If someone only sees formulas and numbers, this may be transparent but does not directly lead to more human-centered AI. 

Fair outcomes

The next element that is part of HCAI is 'fair outcomes'. Striving for fair outcomes in HCAI is important to prevent discrimination and bias. Several steps in the (development) process are crucial to achieving fair outcomes. Firstly, it is important to ensure correct data collection and analysis. There must be a representative and balanced dataset. Additionally, it is important for organizations to test their models for potential bias. And here too, transparency is important in various ways to achieve HCAI. Organizations must be transparent about the algorithms, datasets, and decision-making processes used to enable users to understand the results and be accountable. It can help, and perhaps even be crucial, to ensure diversity and inclusion in the development process. When a diverse team is responsible for developing and implementing AI systems, there is a greater chance that different perspectives will be taken into account and potential biases will be identified and addressed.

Data protection

Finally, protection of personal data is an indispensable element when we talk about HCAI. What can an organization do to ensure data protection in the development and deployment of AI? 'Privacy by design' can be applied in many different ways. For example, as an organization, you must ensure data minimization by taking measures during the design of an AI system to ensure that only strictly necessary data is collected and processed for the intended purpose. An organization can also choose to anonymize or pseudonymize data to reduce identifiability. Additionally, the organization can ensure encryption of (sensitive) data during storage, processing, and transmission. These are just a few of the options organizations have to ensure privacy (by design).

The above elements can be part of HCAI. It is important to conduct checks to ensure that the AI system is human-centered and used ethically. In the context of privacy by design, a Data Protection Impact Assessment can be conducted to assess whether data is being processed correctly and lawfully, what risks exist, and whether additional measures are needed to prevent or reduce those risks. To ensure that an AI system is used in a human-centered or ethical manner as a whole, a Fundamental Rights Impact Assessment can be conducted. An example of this is the Fundamental Rights and Algorithms Impact Assessment (FRAIA) (Dutch: Impact Assessment Mensenrechten en Algoritmen).