This article is written by Firdaus Imambaccus, a social researcher involved in impactful work around data and AI ethics in the Department of Employment and Social Development Canada, and Hugo Morin, a behavioral economist working in the world of AI Ethics in the Department of Employment and Social Development Canada.

The views expressed in this paper are solely those of the authors and do not necessarily reflect the views of Employment and Social Development Canada, nor those of the Government of Canada.

  • The problem: Innovation in the field of artificial intelligence (AI) has taken the centre stage in the last couple of years with its rapid evolution. This explosive growth compelled those involved to rally and align around shared principles to regulate its use and mitigate its risks. However, translating these principles and ideals into practice has presented its own challenges.
  • Why it matters: Guiding principles and frameworks, although necessary, are hardly effective if not put into practice. The challenge stands to figure out how to effectively operationalise these ideals in a way that enables us to regulate AI systems without compromising innovation.
  • The solution: In this article, we present strategies and components critical to the operationalisation of AI ethics, exemplifying it through a responsible data science ecosystem. This ecosystem is an assembly of tools that can be leveraged as part of or to support ethical frameworks.

Let us start with a universal statement: everybody wants to build and use AI systems in a responsible, ethical, and unbiased manner. Yet, examples of the negative impacts and consequences of AI systems are well-documented. The prevalence of these risks and the need to mitigate their actualisation prompted the growing surge of AI ethical frameworks, in both the public and private sector. However, these ethical frameworks are much like apple pies: no one really dislikes the idea of pie, but very few are good at baking it from flour and water to crispy perfection.

This then raises the question: how can we operationalise these principles in the real world?

Two sides of the same coin

Like with other ethical questions, there is no silver bullet to operationalising AI ethics as its application varies across contexts. Yet, we could categorise its crucial components in two intertwined themes:

1. Raising awareness and empowerment

A classic in risk management! The absolute best idea is entirely useless if no one is aware of it.

In this context, raising awareness involves education and capacity building. People must first know how to distinguish AI from ‘The Terminator’ or some obscure extraterrestrial-looking computer screen. What are the capabilities of artificial intelligence? What are its risks? How is it leveraged in the given context?

The next step is to define what critical practices underscore responsible AI. What are relevant regulations and policies from the bedrock of responsible AI development and use? What are the best practices to avoid amplifying and reproducing social biases? Is technological intervention the best approach in the given context? As Maslow pointed out, "[It] is tempting, if the only tool you have is a hammer, to treat everything as if it were a nail."

2. Implementation of strong AI governance oversight

To ensure that AI solutions are developed and deployed responsibly, governance processes play an instrumental role in ensuring that projects, and the people involved, remain in compliance with the relevant regulations, guidelines, and overarching objectives over the course of the project in a way that does not impede workflows, project progression and innovation.

Wrap it all!

Governance tools and processes empower people to incorporate reflexivity in their practice and continually evaluate the efficacy and risks of AI solutions. However, as ethical frameworks and relevant regulations may vary between contexts, it would be virtually impossible to suggest a blanket approach to the operationalisation of AI ethics.

Rather, we propose a set of practical tools that can be leveraged and mobilised in response to the unique project (or organisational) needs, packaged as part of a responsible data science ecosystem.

For example, this ecosystem might include:

  • Automated governance processes Automating governance processes streamlines the process while ensuring privacy, transparency, reliability, accountability, and the adherence to regulations relevant to your context.

  • Fairness assessments This could be a pipeline where the model’s performance could be tested with various subgroups of the population, to ascertain fair and equitable outcomes of the model.

  • Data trust Individuals requesting the use of data could be guided through a coordinated process involving assessment of the risks, verification of the quality of the data and its relevance, and if necessary, a high-level ethical review.

  • Ethical challenge tool This could take the form of a challenge function tool to support the ethical and responsible use and development of AI tools and to address and mitigate the associated risks. This could be as simple as a set of tough-to-answer questions, in which the users confront their implicit biases, and the solution’s potential risks are identified and mitigated.

  • Capacity building This could entail the dissemination of educational resources around relevant policy, regulations, or directives. Resources that are more general, raising organisational awareness about the key issues and risk mitigation strategies, are as crucial as are more advanced technical trainings. In developing such resources, it is important to be cognisant and sensitive to the needs of the target audience. Additionally, it is also a good practice to involve stakeholders with varying expertise and knowledge of the subject matter in the development of the resources.

Of the examples provided, let us delve into the ethical challenges. Too often, and particularly with systems procured from third-party vendors, crucial considerations can often remain unaccounted for. Such considerations could be:

No one (and no field!) is an island

Now people are aware and the proper governance tools are in place, the job is not yet complete. We must now shift our focus to the mindset people should adopt.

Every field has its own biases, perspectives, and approaches. By approaching AI ethics through a multidisciplinary lens, we can build bridges between people with diverse expertise, knowledge, and experience; and in the process ensuring that a multitude of perspectives and solutions are considered.

If AI is in its adolescence, AI ethics is in its infancy.

In doing so, we are encouraged to remain open-minded and critically confront how our origin, gender, age, and field of expertise, to name a few, influence our views on the presenting issues. Collaborating with people of different perspectives helps minimise these biases and enables a more holistic approach to the problem and its solution.

This open-mindedness must be coupled with humility. It is important to avoid the urge to control (which could be hard when hatching a program from scratch). The key driver in the process of operationalising responsible AI is to be humble, receptive, and open to feedback from other specialists and stakeholders.

Newborns grow fast

If AI is in its adolescence, AI ethics is in its infancy. It is a newborn that is dragged in all sorts of directions. When a few lines of code might result in unintended harmful consequences, and knowing that coding is addictive, clear boundaries and safeguards must be built, often from a blank slate. However, our desire for productivity and the increasing need for efficiency must also be tempered. These questions are becoming more complex with each passing day. For this reason, you are likely better off starting with a stringent ethical framework. From a cost-benefit perspective, project delays might be preferred over the consequences of the unethical use of AI. Similarly, an ill-thought-out framework might instill too much confidence in our approaches and perhaps cause us to be less critical and thorough when evaluating our solution.

Ultimately, the operationalisation of AI ethics remains a delicate balancing act requiring the long-term investment of time and effort. As Matshona Dhliwayo once said: “Roses do not bloom hurriedly; for beauty, like any masterpiece, takes time to blossom.”

👋 You can write an article like this one! Share your thoughts with a community of public servants. Learn more.

(Image Credit: Deep Mind, Unsplash)