What alignment means in practice
On paper, alignment seems straightforward. It means ensuring a system's goal matches what people actually want.
But this is where the real challenge begins.
We give the system a target to aim for, such as a metric to optimise or a rule to follow. The system then learns to hit that target with ruthless efficiency, even when it violates our unspoken intentions.
In the real world, with all its messy edge cases and shifting contexts, this gap between the system's objective and our own only grows wider.
A practical way to see alignment is to ask three questions at each decision point:
- What objective is the system optimising in this context?
- What outcome do users and stakeholders actually want?
- What safeguard prevents harmful shortcuts?
The same three questions can reveal misalignment across sectors. For instance:
Benefit eligibility (public sector)
Automated checks can cut fraud, but if they are too strict, people with genuine claims are blocked. The numbers look good, but the service fails its users.
Credit scoring (financial services)
Credit models can predict defaults, but if the data misses parts of the population, people get scored unfairly.
Why misalignment matters
Misalignment is not just technical. It erodes trust in institutions.
In the public sector it can look like opaque rejections with no right to challenge. In finance it can look like scores that are hard to understand and harder to dispute.
These are not rare events. They are predictable when systems optimise for narrow targets without visible safeguards.
Design has leverage here. Most exposure happens where people meet the decision. That is the interface, the notification, the appeal path and the time it takes to get a human response.
Eight principles for safe AI
A widely cited review of AI guidance shows eight themes that recur across frameworks: privacy, accountability, safety and security, transparency and explainability, fairness and non-discrimination, human control, professional responsibility and the promotion of human values.1
Reframed for UCD, they become practical requirements:
Privacy
Let people see, control and correct how their data is used.
Accountability
Make challenge and correction possible, with clear routes and timeframes.
Safety and security
Check safety rules and competence before any optimisation or dispatch.
Transparency and explainability
Pair each decision with a plain reason and what to do next.
Fairness and non-discrimination
Measure statistical bias and listen for perceived unfairness.
Human control
Keep meaningful review points and a real right of appeal.
Professional responsibility
Test the unhappy path with those most at risk of harm.
Promotion of human values
Design for dignity and human flourishing, not only efficiency.
Turning principles into design work
Alignment is not something that happens inside a model. It is decided in design. The moment a person meets an automated decision, the experience either builds trust or erodes it.
Think about what it feels like to be on the receiving end. If the system gives a result with no reason, the user is left guessing. If it blocks a path to challenge, the service feels closed. If errors cannot be corrected, people learn that the system is not for them.
Each of these moments tells us whether alignment has been taken seriously.
Alignment is earned in design. It is measured in the confidence people feel that the system will listen, adapt and allow them to act when it matters most.
Designers have levers here. We can choose to make reasons visible, not hidden. We can treat recovery and appeals as part of the core journey, not side channels. We can involve those who have the most to lose, not only the easiest users to please.
These choices are not technical extras. They are what turn principles into practice.
Sources and further reading
- Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1(9), 389–399.
- Russell, S., Dewey, D., & Tegmark, M. (2015). Research priorities for robust and beneficial artificial intelligence. AI Magazine, 36(4), 105–114.
- Mittelstadt, B. D. (2019). Principles alone cannot guarantee ethical AI. Nature Machine Intelligence, 1(11), 501–507.