Contents
The Role of AI in Our Modern Lives
Today, hundreds of millions of people utilise tools like ChatGPT for brainstorming and Midjourney for creating visuals. These artificial intelligence (AI) tools have seamlessly integrated into our daily routines, heralding a new digital era. Consequently, we now work more efficiently and tackle professional and creative challenges more adeptly, leading to accelerated innovation.
However, AI’s significance extends beyond merely supporting daily tasks. It powers critical services and maintains societal infrastructure. Whether facilitating loan agreements, granting access to higher education, or supporting health care, AI has become indispensable. Identity verification, once just a gateway to credit checks and bank account openings, now supports services from healthcare to travel and eCommerce.
Addressing Bias in AI Systems
AI systems, whilst powerful, can exhibit biased behaviours towards end-users. Companies such as Uber Eats and Google have unearthed how AI can jeopardise the legitimacy of online services. Yet, it’s worth noting humans themselves are subject to biases, such as the well-documented Own Group Bias (OGB) in facial recognition.
Pillar 1: Identifying and Measuring Bias
The initial step in combating bias is establishing robust measuring processes. Biases are often subtle, hidden within vast datasets, and only identifiable after decoupling correlated variables. Companies need good practices like using confidence intervals, appropriate dataset sizes, and suitable statistical tools. Transparency is key, as exemplified by Onfido’s 2022 "Bias Whitepaper". Public tools like NIST’s Face Recognition Vendor Test also provide valuable bias analyses.
| Practices for Measuring Bias | Benefit |
|---|---|
| Measurement by confidence interval | Increases accuracy |
| Use of varied and large datasets | Ensures robustness |
| Employment of skilled statisticians | Enhances credibility |
| Transparency through public reports | Builds trust |
Confounding Variables and Hasty Conclusions
Bias in AI often lurks among multiple correlated variables. Take facial recognition – a crucial step in identity verification. Initial analysis might show poorer recognition for dark-skinned individuals. One might hastily blame the system. However, deeper examination reveals that many African countries, with predominantly dark-skinned populations, use lower-quality identity documents. This document quality confounds results, as demonstrated by a performance improvement when isolating high-quality European documents.
Pillar 3: Rigorous Training Methods
The training phase of an AI model is pivotal for bias reduction. The datasets used significantly influence model behaviour. Correcting imbalances in these datasets can mitigate biases. For instance, an online service used mainly by one gender might show model robustness for that gender, disadvantaging the other. Balanced data sampling can remedy this, especially for critical services like higher education applications. Collaboration with regulators like the ICO has helped companies like Onfido reduce performance disparities in demographic groups.
Pillar 4: Tailoring Solutions to Use Cases
There’s no one-size-fits-all solution for bias measurement. Google’s Glossary on Model Fairness identifies various fairness definitions, each valid but resulting in different model outcomes. For identity verification, Onfido uses the "normalised rejection rate" to assess bias. While achieving perfect parity is challenging due to data limitations and external factors, striving for it is vital.
The Path Forward: A Balanced Approach
Biases, while inevitable, shouldn’t deter the adoption of AI. It’s crucial to measure, reduce, and communicate openly about these biases. Many companies, including Meta, contribute significantly to bias research, offering datasets and publications. Despite the persistence of biases, the potential benefits of AI in improving digital services are substantial.
If companies take steps to mitigate biases, customers’ digital experiences will undoubtedly improve. They will access appropriate services, adapt seamlessly to new technologies, and receive the necessary support from their chosen companies.
Olivier Koch is VP of Applied AI, Onfido.
