a 5 minute guide to Responsible AI

Artificial intelligence development is going crazy fast. Shortly after the release of ChatGPT in November 2022 there was even a petition signed by some of the biggest names in the industry to halt innovation for 6 months.
Their reasoning: artificial intelligence is evolving too quickly, posing significant ethical risks.

As AI becomes more indistinguishable from humans, it is important that we think about its uses and effects: intentional and unintentional. We should take responsibility in our own work to build responsible tooling. Algorithms touch all aspects of life, which means that biases within these algorithms impact every party of society. And as regulations like the EU Data Act come into force in 2024, responsible development practices will soon become a required part of our work and routine.

For every stage of the data process (ingestion, exploration, pre-processing, modeling, production, etc) it brings different challenges. We all share the responsibility to stay informed and engaged, on both individual and company level. Some major tech players are taking steps: Microsoft, for example, has set up a company-wide approach to responsible AI, which includes committees like AETHER and RAISE, as well as an annual Responsible AI Transparancy report. These efforts aren’t perfect but they do represent progress.

Tools that promote Responsible AI are crucial for developers and data scientists. They assist in building more ethical algorithms, providing insight into biases built into our data, but also help us recognize the broader consequences of our work.

In this article I’ll quickly talk you through one of these toolkits: FairLearn. There are many more toolkits out there so I encourage you to take a look (and reach out to me if you find interesting ones!).

What is Responsible AI, and why does it matter?

There are multiple interpretations of what Responsible AI entails, but there are some recurring themes: transparancy, accountability, human oversight and privacy & safety. Their aim is that AI technologies are used ethically, safely, and for the benefit of all.
Let’s take a quick look at some of these themes:

Transparancy

AI systems should be understandable and explainable. This means that the decision-making processes behind algorithms must be transparant. It includes making AI models interpretable and providing clear documentation about their capabilities, limitations. training data and how the model is monitored and optimized.

Accountability

There must be clear ownership and responsibility for the system’s outcomes, whether it’s the developer, organization deploying it or regulatory bodies. Accountability ensures that any negative or unintended consequences of AI use can be addressed, and corrective actions can be taken. This also incudes impact assessments to evaluate risks, especially to vulnerable populations.

Human oversight

AI systems should be designed in a way that human judgement and intervention are possible, or in my personal opinion: necessary. This ensures that it doesn’t act in ways that are harmful or unintended. It also emphasizes the importance of keeping humans ‘in the loop’ and maintaining ultimate control over decisions made within these systems.

Privacy & safety

Protecting the privacy and security of individuals is essential. This pillar focuses on securing A applications from threats and ensuring that they operate safely without exposing users to harm or risks.

Why should I care?

You’re a busy person, I get it! As AI becomes more embedded in our daily lives, the decisions they make - and biases they carry - can have real, lasting consequences. Large Language Models (or LLM’s) and Generative AI provide, in my opinion, an extra challenge: the models are more ‘black box’ and observability is very difficult to achieve due to its size and changing nature.
That being said we all carry responsibilty to ask questions and be vigilant about these systems, even if you’re not directly involved in development of these algorithms. Examples like the ‘Toeslagenaffaire’ in the Netherlands and research by Joy Buolamwini proving gender and skin-type bias in facial recognition systems are only the tip of the iceberg.

There is also more regulation coming. The EU AI Act and Data Act (set for 2024 - 2026) will require businesses to develop and deploy AI responsibly. The Dutch government has set-up a ‘algorithm register’ where high-risk algorithms need to be registered, and which needs to be up-to-date in 2025.
These laws are not a silver bullet, especially since the pace at which AI evolves is much quicker than regulation, but I am expecting more and more policies to be introduced. And not just in the EU.

Open source toolkits

There are open source python packages and other toolkits available to help you become more aware of what’s happening in your datasets, and guide you in improving your algorithms. There is a lot more tooling available, which I encourage you to take a look at.

One such example is FairLearn: a python package that helps assess ‘fairness’, meaning ‘allocation harms’ and ‘quality-of-service harms’ (checkout the FairLearn repo README for more details), in your models. It helps provide insight into your dataset so you can review whether it’s balanced and whether all groups are treated equally, hereby highlighting biases.

The quickstart is a great place to start, and you can find my quickstart including notes here.
It uses a clinical data set of hospital re-admissions over a ten-year period (1998-2008) for diabetic patients, and investigates how racial disparities impact health care allocation in the US. It will guide you through how to assess the biases in the dataset and how to fix them.