Ethical considerations in the development of the Digital Twins of the Ocean

In recent years, Digital Twins of the Ocean (DTOs) - digital replicas of ocean processes - have emerged as a tool for modelling the complex interactions that govern marine systems, exploiting the power of Artificial Intelligence (AI) and large training datasets to understand ocean processes and predict their future in a rapidly environmentally changing world. DTOs, although very complex, offer many advantages including providing a decision support tool for areas such as optimizing fisheries, emergency reactions to tsunami warnings, adapting to sea-level rise, protection of biodiversity and improving climate prediction/climate forecast. These powerful tools offer many promises; however, we need to go beyond the technical aspects and consider AIs impact on the decision-making process.

Three key aspects of the coupling model/AI are essential for consideration by the marine scientific commmunity, the Artificial Intelligence and policy-maker communities. The hope is that by considering the limitations early in development, we can optimize the use of AI. The key aspects to consider are:

• Data is of paramount importance: the source of data, its geographical origin, its nature and its quality should be carefully considered when developing a DTO. The need for seamless interoperability of these data raises the question of which data standards are selected and how they are applied. All these considerations may introduce biases into the algorithms, which need to be identified: in specific cases, the use of open-access data may introduce bias as it may not have access to sources related to endangered species or Indigenous knowledge. Furthermore, data openness for DTO models may have ethical limitations such as compliance to Access and Benefit Sharing regulations, or sharing of data from commercially valuable or endangered species, which question the conditions under which data should be made open.
  
  
• The model itself is a mathematical object, based on physical conservation principles and a set of hypotheses that guaranty the consistency of the reasoning. It also comes with certain limitations and uncertainties, especially in the biological modelling and always involves some numerical approximation for being solved within the available computational power. This process of model development and use, and the benefits and limitations that users assume the models may contain, must be transparent and accountable as highlighted in the European guidelines on Trustworthy AI.
  
  
• Finally, the result we expect from the data-driven DTOs is a powerful decision-support tool, capable of predicting and warning. These tools need to be explicit and targeted to the end-users, leveraging the complexities of the analyses and ensuring that the results and choices of data and models ensure transparency and present all biases and uncertainties, to allow the end user to draw reasonable conclusions. End-user training is an essential aspect to consider. The decision-making chain of command must be solid, well identified and structured, and accountability is key, especially in crisis management due to natural hazards.

It is urgent that marine scientists, AI developers and policy makers work together for the best for the planet.