What Makes AI Truly Human-Centered? It Starts With Trust, Not Hype

Artificial intelligence (AI) is almost everywhere, recommending what we watch, shaping what we read, and increasingly influencing decisions that affect our jobs, health and finances. However, amidst this rapid AI boom, a fundamental question is growing harder to ignore: are these technologies actually built around people, or are people being forced to adapt to them?

The tech industry often answers that question with a reassuring phrase: "human-centered AI." But what does that really mean in practice? A new study published in the MDPI journal Knowledge takes a closer look, moving beyond slogans to ask what it truly takes for AI to put humans at the center.

The answer, according to the research, is not one principle or one design feature. Human-centered AI depends on a broader set of attributes that include trust, values, human benefit, user needs, usability, control, dignity and meaningful stakeholder involvement. The study points out that AI cannot be called human-centered simply because it is efficient, popular, or easy to deploy. It must be designed, evaluated and governed around people's needs, rights and real-world experiences.

The AI field has a definition problem

A policymaker may demand "trustworthy AI," a designer may focus on usability, an engineer may prioritize performance, and a business may measure customer satisfaction. All may claim to be building human-centered AI, yet they may be working from very different assumptions.

The study addresses this fragmentation by examining 81 definitions of human-centered AI from academic, institutional and industry sources. It extracts 78 keywords, refines them through expert evaluation, and develops a final inventory of 26 attributes. These attributes were then rated by 145 respondents, including practitioners, academics and students, most of them from the Asia-Pacific region.

The result is a more practical vocabulary for a field that often moves between ethics slogans and technical benchmarks. Instead of treating human-centered AI as a broad aspiration, the study breaks it into assessable dimensions: human trust, human values, human needs, user needs, usability, human control, stakeholder engagement, human dignity, human benefits, and more.

Trust tops the list, but it cannot be demanded

The highest-ranked attribute in the study was human trust. The finding is not surprising in a world where AI systems are increasingly used to assist or influence decisions in education, healthcare, hiring, finance, public services and digital platforms. If people do not trust AI systems, they may resist them, misuse them, or become vulnerable to hidden harms.

However, the study's qualitative findings add an important warning: trust is not a switch that developers can turn on. Respondents viewed trust as something AI systems earn through transparency, reliability, explainability and meaningful human oversight. In other words, trust is an outcome of design and governance, not a marketing claim.

Organizations often say they want users to trust AI, but the better goal is to make AI systems trustworthy. That requires explaining how systems work, where their limits are, what data they use, how decisions can be challenged, and who remains accountable when something goes wrong.

Human values also ranked near the top. Respondents saw fairness, dignity, privacy, autonomy and justice as central to human-centeredness. This shows that users and experts are not only concerned with whether AI works. They are concerned with whether it works in ways that respect people.

The human part of AI is the hardest to build

Human emotions ranked last among the 26 attributes, while empathy and human behaviours also appeared near the bottom. At first glance, this might suggest that stakeholders do not care much about emotional or behavioural dimensions of AI. The qualitative evidence suggests the opposite. Respondents did not dismiss emotions or empathy as unimportant. They viewed them as difficult, uncertain and risky for AI systems to handle reliably.

As AI enters mental health support, tutoring, elder care, workplace management and customer service, emotional intelligence will become more important, not less. But the study indicates that stakeholders remain cautious about whether AI can genuinely understand human emotion, cultural context and behavioural complexity without oversimplifying or manipulating people.

AI systems that claim to detect emotion or respond empathically may create new risks if they misread users, reinforce stereotypes, or simulate care without genuine accountability. Human-centeredness should not mean pretending that AI understands people more deeply than it does.

Students want outcomes. Academics want process

The study found broad agreement across practitioners, academics and students on which attributes matter most. Trust was the only attribute appearing in the top five for all three groups, but there were important differences in emphasis. Academics placed greater weight on participatory and process-oriented attributes such as user involvement, user needs and stakeholder engagement. Students placed more emphasis on visible outcomes such as trust, benefits, usability and control. Practitioners occupied the middle ground, with a distinctive emphasis on human values.

Students entering the AI workforce may understand the importance of useful and trustworthy systems, but they may need stronger training in how such systems are created. Human-centered AI is not achieved only at the end of development, when a product is tested with users. It requires participation, feedback, iteration and accountability from the earliest stages of design.

AI education must go beyond coding, model accuracy and deployment. Students need exposure to human-centered design, user research, ethics, participatory methods, human factors, accessibility and social impact assessment. Without that foundation, future AI professionals may value human-centered outcomes without knowing how to build the processes that produce them.

From ethics slogans to design standards

Around the world, AI governance is moving from abstract ethical principles toward implementation. Governments are developing laws and risk frameworks. Companies are creating internal AI governance systems. Public institutions are experimenting with AI tools for service delivery. In this environment, human-centeredness cannot remain a vague promise.

The 26-attribute inventory offers a starting point for turning human-centered AI into something more concrete. For regulators, it can help bridge the gap between high-level requirements and practical evaluation criteria. For companies, it can guide product teams toward better design questions: Does the system protect human dignity? Can users understand and challenge its outputs? Were affected communities consulted? Does it produce real human benefit? Can people maintain control over consequential decisions?

For civil society, the framework can support accountability. Instead of asking whether an AI system is simply "ethical" or "responsible," advocates can ask whether it meets specific human-centered attributes.

AI systems are increasingly being introduced into public services, education, agriculture, health and financial inclusion across developing regions. If these systems are imported or scaled without attention to local needs, languages, infrastructure and social realities, they may deepen inequalities rather than reduce them. Human-centered AI must therefore be culturally aware, context-sensitive and participatory.

The study has limitations. Its sample is concentrated in Southeast Asia, with strong representation from Myanmar, Singapore and Thailand. The results should not be treated as a global standard without replication in other regions. The research is exploratory and requires further validation through larger samples, factor analysis and domain-specific testing in fields such as healthcare, education and public-sector AI.

Still, the study reminds policymakers, businesses and researchers that human-centered AI is not a decorative label.