Building Trust in AI Development: Character

Character in AI development is the trust element that cannot be faked for long. You can design a system that appears clear, seems competent, and performs well in a demonstration. But character reveals itself under pressure, in edge cases, in the moments no one planned for.

For AI systems, those moments arrive constantly.

This is the third in a series exploring the eight principles of trust in AI, drawn from Justin Dean's PhD research at the University of Newcastle. Each principle matters. Character is the one that cannot be retrofitted.

‍

What Character Means in AI Development

In human terms, character is the alignment between what you believe, what you say, and what you do. It is most visible not in normal conditions but in difficult ones. A leader who maintains ethical standards when it costs them something has character. One who abandons them under pressure does not.

The same test applies to AI. A system that behaves predictably and ethically under normal inputs, but drifts or fails when inputs are unusual, does not have character built in. It has character as a performance. And performances end.

‍

The Three Tests of Character in AI

Moral standards. What is the system designed to do, and what is it designed not to do? The ethical boundaries of an AI system are set in its design and training. If those boundaries are not deliberate and documented, they do not exist. An AI system cannot have moral standards by accident.

Consistency between stated purpose and actual behaviour. Does the system do what it says it will do? A system that promises accurate information but produces confident errors fails this test. A wayfinding system that promises to guide users but gives incorrect directions fails it too. The gap between stated capability and actual behaviour is a character gap.

Ethical decision-making under pressure. What happens at the edges of the system's training? When the input is ambiguous, unusual, or outside the expected range, does the system behave safely and honestly? Or does it extrapolate beyond its knowledge in ways that could mislead or harm?

These are the moments that determine whether an AI system has genuine character or just a well-designed normal case.

‍

Why Character Comes Before Adoption

Organisations do not broadly adopt AI systems they cannot trust. They pilot them, constrain them, monitor them closely, and limit their scope. That is the rational response to a system whose character is unknown or unproven.

Building character into AI from the start, through deliberate ethical design, consistent testing at edge cases, and transparent documentation of what the system will and will not do, is what earns the trust that leads to meaningful adoption.

This is not an abstract principle. It is the practical reality of deploying AI in environments where the stakes are real: healthcare, government, public infrastructure. In those contexts, a system that fails at the edges does not just underperform. It causes harm.

‍

Building It In, Not Bolting It On

Character cannot be added to an AI system after the fact. It is set in the decisions made during design and development. The values embedded in training data. The boundaries encoded in the system's behaviour. The documentation that tells users what the system does and does not know.

This is why Justin Dean's research on trust in artificial intelligence focuses on the design phase, not the deployment phase. By the time an AI system is in the hands of users, its character is already largely fixed. The question is whether that character was built deliberately.

Ask yourself, at every AI development decision point: are we doing what is right? Not what is fast, not what the deadline allows. What is right.

That question, asked consistently throughout the development process, is how character gets built in.

‍

‍

This post is part of Envent's series on building trust in AI. Read the full framework: [Trust Is Earned. Here Are the Eight Principles That Build It.]

‍