Introduction

The emergence and adoption of Artificial Intelligence (AI) is expected to greatly affect economics, humanity, and our way of life. As such it is important to take a human centric approach when implementing or developing AI solutions to ensure that AI is adopted both responsibly and ethically.

AI will undoubtedly bring both benefits and risks which we will need to consider when considering AI applications and solutions.

This document provides our Code of Conduct with regards to i) the adoption of AI as a service/product, ii) the adoption of third-party AI tools/services within our products/services and iii) the development of AI services/products by the partner institutions. It provides our guiding principles and key risks we will consider when evaluating opportunities to use AI. We will ensure to adhere to these principles to deliver responsible AI, while mitigating the associated risks and ensuring its ethical application.

Following the EU guidelines of AI Code of Conduct we will remain absent from the adoption or development of AI use cases and applications that are considered as potentially harmful to the Union and its citizens. The AI Code of Conduct will also reflect the EU AI Act as it comes into effect. In accordance with the EU AI Act, projects where the risk of AI is deemed to be unacceptable as defined by the EU AI Act will not be pursued.

The Code of Conduct will be used in tandem with any principles or codes of conduct already established by consortium members or other collaborators. The AI Code of Conduct has been established as the guiding principles for projects to be delivered only through the DiGGiN programme and does not bound consortium members beyond this scope.

1. Risks of AI

We will seek to consider and assess the risks on a case-by-case basis and where  required we will identify appropriate measures to mitigate these risks.

The following AI risk categories will be considered when assessing applications of AI, to confirm the suitability of the use-case and to define an approach that allows the mitigation of these risks, for the benefit of both the users and society in general.

1.1. Performance Risk

AI algorithms that ingest real-world data and preferences as inputs may run a risk of learning and imitating possible biases and prejudices

We will consider risks such as:

  • Risk of error
  • Risk of bias and discrimination
  • Risk of opaqueness and lack of interpretability
  • Risk of performance instability

1.2. Security Risk

For as long as automated systems have existed, humans have tried to circumvent them. This no different to AI.

We will consider risks such as:

  • Adversarial attacks
  • Cyber intrusion and privacy risks
  • Open source software risks

1.3.  Control Risk

Similar to any other technology, AI should have organisation-wide oversight with clearly-identified risks and controls

We will consider risks such as:

  • Lack of human agency
  • Detecting rogue AI and unintended consequences
  • Lack of clear accountability

1.4.  Economic Risk

The widespread adoption of automation across all areas of the economy may impact jobs and shift demand to different skills.

We will consider risks such as:

  • Risk of job displacement
  • Enhancing inequality
  • Risk of power concentration within one or a few companies

1.5. Societal risk

The widespread adoption of complex and autonomous AI systems could result in the development of “echo-chambers” between machines and can have broader impacts on human-human interaction.

We will consider risks such as:

  • Risk of misinformation and manipulation
  • Risk of an intelligence divide
  • Risk of surveillance and warfare

1.6. Enterprise Risk

AI solutions are designed with specific objectives in mind which may compete with overarching organisational and societal values within which they operate. Communities often have long informally agreed to a core set of values for society to operate against. There is a movement to identify sets of values and thereby the ethics to help drive AI systems, but there remains a disagreement about what those ethics may mean in practice and how they should be governed. Thus, the above risk categories are also inherently ethical risks as well.

We will consider risks such as:

  • Risk to reputation
  • Risk to financial performance
  • Legal and compliance risks
  • Risk of discrimination
  • Risk of values misalignment

2. Ethical AI principles

Recognising the disruptive potential of AI we have agreed on a number of key Ethical AI Principles that will guide our decision making with regards to selecting AI use-cases and their implementation.

These principles will allow us to engage with AI to benefit organisations, while considering the moral implications and impact on society and people. The principles will underpin our decision across the AI lifecycle and will drive our key practice that will enable us to have an ethical approach to AI.

These key principles are:

2.1.  Accountability

All stakeholders of AI systems are responsible for the moral implications of their use and misuse. There must also be a clearly identifiable accountable party, be it an individual or an organisational entity.

2.2.  Human Agency

The degree of human intervention required as part of AI solutions’ decision-making or operations should be dictated by the level of perceived ethical risk severity.

2.3. Data Privacy

Individuals should have the right to manage their data when it’s used to train and run AI solutions, as well as managing how that data is reused for other purposes.

2.4. Safety

Throughout their operational lifetimes, AI systems should not compromise the physical safety or mental integrity of humans.

2.5. Lawfulness and compliance

All the stakeholders in the design of an AI system must always act in accordance with the law and all relevant regulatory regimes.

2.6. Fairness

The development of AI should result in individuals within similar groups being treated in a fair manner, without favouritism or discrimination, and without causing or resulting in harm. AI should also maintain respect for the individuals behind the data and refrain from using datasets that contain discriminatory biases.

2.7. Beneficial AI

The development of AI should promote and reflect the common good, such as sustainability, cooperation and openness in accordance also with the EU guidelines and principles.

2.8. Interpretability (Explainability, transparency, provability)

An AI system should be able to explain its model decision-making overall, as well as what drives an individual prediction to different stakeholders. In cases where model explanations cannot be provided due to a black-box approach, an alternative transparency method should be adopted. Information and descriptions on training data, and outputs should be documented as part of interpretability.

2.9. Reliability, robustness, security

AI systems should be developed so that they will operate reliably and safely over long periods of time using the right models and datasets.