It can’t be stated enough: it’s the responsibility of every organization creating artificial intelligence (AI) to do so ethically. Responsible, or ethical, AI is AI that’s unbiased, equitable, and improves the quality of life of everyone it touches. In practice, it requires AI practitioners to apply an ethical framework to every AI endeavor they pursue, ensuring the people, processes, and tools leveraged serve that larger mission.
Throughout the AI value chain, there are crucial touchpoints where responsible AI can and should play a role. If AI developers overlook any touchpoint, they put the entire project at risk for not meeting standards of fairness and equity. Understanding how each link in the chain influences the next and what considerations are relevant in each phase are foundational steps for AI practitioners launching responsible AI.
The AI Development Cycle
Before starting any AI project, there are several actions your team can take to lay the groundwork for building an ethical AI product.
- Be aware of the regulations your end product must comply with. These vary by geography; the GDPR, for instance, covers AI development and application in the EU.
- Create an AI governance framework that reflects a responsible AI lens. Include guiding questions in this framework that help you address key points of consideration along the AI development lifecycle. Data governance, in particular, should be an area of focus given its significant impact on model performance.
- Recruit a diverse team of data scientists and researchers who can bring varied perspectives and experiences to AI development.
With these initial steps completed, you’re more prepared to make responsible decisions throughout the AI build process. After you’ve selected a business problem to solve, the AI value chain starts with data collection and progresses through deployment, before undergoing the cycle again and again during retraining in post-production. We’ll cover the key responsible AI considerations to think about in each stage:
When sourcing data, make it a goal to collect data that’s as complete and inclusive as possible. The more end users (and specifically their respective use cases) represented in the data, the better your AI will perform for a wide variety of groups. Ethical AI is predicated on the idea that the product works equitably for everyone, and representative data is the foundation of creating that equity.
If you’re sourcing data from a third-party, ideally it’s your joint responsibility to ensure the data is unbiased. Still, it’s best not to make any assumptions. For example, if you’re collecting image data of scientists, you can’t assume that all types of scientists are represented in the dataset or all demographics. Even if you’ve been told as much, it’s ultimately up to you to carefully quality check your data for coverage of all possible use cases.
Often when we talk about responsible AI with respect to data preparation we focus on the annotation process itself, the goal being to apply accurate, unbiased labels. Indeed, this is mission-critical in impacting bias in performance of the final model. A key component of this step is recruiting a diverse group of people (ideally people who are highly representative of your end users) to provide data labels. Greater diversity fosters differing perspectives and reduces the chance of introducing one-sided judgments.
What we often overlook in data preparation is the treatment of the people behind the data. As an AI practitioner, ensure these people receive fair treatment, as they’re a crucial but undervalued part of the AI value chain. Fair treatment may include providing fair pay, protecting their right to privacy, and offering open lines of communication for feedback. (To see how Appen addresses the wellbeing of our contributors, see our Crowd Code of Ethics).
Model Training and Testing
Responsible AI isn’t just about the data. After you’ve built your model and started training it on your prepared data, you’ll monitor its performance. The most common metric in performance measurement is, of course, the accuracy of the model’s predictions (e.g. does it always identify pedestrians in images where people are visible crossing the street?).
More nuance is needed for accuracy, however. You need to evaluate the model’s accuracy specifically for each group of end users. The groups you select will depend on the problem you’re trying to solve, but always be aware when you’re interfacing with protected classes of people (i.e., people who share a common trait, such as race or gender, that’s legally protected). Does your model perform equally well for protected classes vs. non-protected? If not, you likely need to retrain your model on additional data representing the underperforming classes.
On top of measuring the accuracy of your model, consider including a metric that directly measures bias. Incorporating a bias metric assists you with catching instances of bias quickly—although it shouldn’t replace regular quality checks done by humans. Note that there are software options available that provide this function if you need further guidance on how to add this to your dashboard.
After you’ve deployed your model, continue to evaluate its performance across user groups and to check that it works as intended. It’s important to enable your users to easily provide feedback so you can catch and fix problems as fast as possible.
It’s expected that your model performance will degrade over time if you never retrain it; most models don’t operate in static environments and are frequently faced with changing data that they haven’t encountered before. Regularly retrain your model on new data by starting again with data collection and working through the rest of the AI development cycle.
Next Steps for Responsible AI
If we look at the overall landscape of responsible AI, progress is still needed. More companies need to understand that responsible AI is essential to success and part of the job, and not just a nice-to-have. Greater acceptance of this idea would reduce the need for strict, potentially stymieing, regulations in the future.
As an AI practitioner, what else can you do now to further the progress of responsible AI? Stay up to date on ethical AI news, explore different industries to learn how they’re approaching the concept, and ask customers and annotators for regular feedback. Thoroughly document the choices you make and the tools you use in development to help address the problem of explainability in AI, aiding all of us to better understand how these innovations work. And above all, approach each project with a commitment to fairness and inclusion throughout the AI development cycle.
Learn more! Watch CallMiner’s webinar Responsible AI Across the Value Chain to hear from industry representatives, including Appen, on this important topic.