Values as a driver of future-proof AI

From abstract values to concrete AI choices

Many organisations want to deploy AI responsibly, but find it difficult to translate values such as transparency, privacy and fairness into concrete design and implementation choices. TNO therefore develops practical methods that help organisations explicitly incorporate societal, organisational and ethical values into the development and deployment of AI systems. These methods do more than identify the values the model should serve, such as safe roads, efficient maintenance and road users’ privacy. They also help determine which data is necessary and which data can be left out.

Existing frameworks often stop after identifying abstract values or focus mainly on completing checklists. Our methods are applied in interactive workshops, making complex moral discussions accessible and turning the outcomes into concrete results. For a road maintenance model, this could mean excluding privacy-sensitive data or incorporating clear design choices about data retention periods. Through these workshops, organisations gain insight into relevant values and interests, possible tensions between them, and practical starting points for designing responsible AI applications.

To ensure that these insights remain useful after the workshops, TNO is developing an interactive dashboard that brings together the workshop results. It gives organisations a clear overview of the values identified, stakeholder perspectives, the trade-offs made and the resulting design choices.

AI that applies values, even under uncertainty

Identifying the values that should guide an AI system is only the first step. Values such as safety and privacy are inherently qualitative, open to interpretation and difficult to capture in technical or quantitative terms. This makes it challenging to incorporate values into AI design or assess whether a system’s behaviour is consistent with the intended values.

Our Moral Model architecture helps developers design AI systems that remain aligned with the intended values. The architecture addresses key questions such as: how should a value be translated into machine-readable terms? And what impact do choices within the system have on important values in a road maintenance prediction model, such as privacy, safety and costs? By clearly separating impact modelling, value assessment, decision-making within the AI system and user interaction, the architecture helps ensure that crucial value-alignment requirements are considered throughout the design process.

For existing AI systems and autonomous applications, our technical measurement tool, the ethical adherence index, offers a practical way to measure value alignment. The index continuously assesses how system behaviour affects the relevant values. It rises when the system’s behaviour supports these values and falls when it does not. This provides insight into the extent to which the AI system acts in accordance with the intended values, particularly in situations where information is incomplete or conditions are volatile, uncertain, complex and ambiguous. Our tools enable organisations to recognise such situations, compensate for them and provide users with additional context.

Monitoring values in AI

A road maintenance prediction model may begin to deviate from its original purpose. Changes in traffic can produce wear patterns that differ from those suggested by historical data. Maintenance policies and standards may also change, causing the model’s predictions to no longer reflect reality.

Making values explicit, translating them and implementing them is therefore an essential first step, but not the end point. Value alignment must continue throughout the use of an AI system. Changes in the environment, the way the system is used or the societal context can cause its outcomes to diverge over time from what was originally intended. Monitoring is also receiving increasing attention in legislation, including the EU AI Act. TNO has therefore developed a process and the necessary tools to enable the monitoring of values.

Monitoring allows organisations to identify and address deviations in good time, before problems arise. Our process and tools help them determine when a system is displaying undesirable behaviour and what is causing it. For example, a maintenance prediction model may flag repairs early, helping to maintain the quality and safety of our infrastructure. However, it may also lead to road closures and high costs when these are not always necessary. Monitoring makes this trade-off clear and manageable for organisations.

By monitoring technical measurements, such as statistics and metrics on traffic intensity, alongside qualitative data, such as road users’ experiences, our process gives organisations insight into how system behaviour and user experiences relate to the intended values, which trade-offs are required and where adjustments may be needed. Monitoring also helps identify values that emerge during use but were not considered beforehand. In the road maintenance example, it may initially seem logical to monitor traffic volume as a key factor. Yet in rural areas, a less frequently used road may be vital to local traffic or emergency services. A failure to carry out maintenance on time could then have serious consequences.

Our monitoring process and dashboard help organisations determine whether the intended values are being upheld, clarify how different values interact and make informed choices based on monitoring results.

The challenge and the solution of value-driven AI

Value-driven AI requires trade-offs during design and procurement, implementation and use. With the methods, processes and tools developed by TNO, we can facilitate discussions about values and AI, help organisations make informed trade-offs within and between values, provide tools to implement these values in the system, and enable organisations to monitor the system and its behaviour.

Visual: Hanna Barakat & Archival Images of AI + AIxDESIGN / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/