Plotting a path to responsible AI is essential

WITH financial services spending on artificial intelligence (AI) in the Asia-Pacific region forecasted to grow 22.1 per cent annually to US$4.9 billion by 2024, reliance on the technology is heading in only one direction: up. As these investments scale, AI technology will have a growing influence on the financial sector and consumers' everyday lives. From algorithmic trading to fraud detection and portfolio optimisation; organisations are increasingly leveraging AI to automate key processes that, in some cases, are making life-altering decisions for their customers. Not understanding how these decisions are made, and whether they are ethical and safe, create enormous legal vulnerabilities and business risks.

How then can businesses harness the power of AI to unlock new efficiencies and business value, while minimising these vulnerabilities? The answer lies in ensuring AI solutions are deployed responsibly - where AI and machine learning (ML) models are robust, explainable, ethical and auditable. Yet, the primary challenge most business leaders are faced with in achieving this golden standard is a lack of understanding around their AI models. A recent survey by FICO revealed that 65 per cent of business leaders are unable to explain how specific AI model decisions are made despite having increased AI investments overall in the past 15 months.

INSUFFICIENT KNOWLEDGE

The lack of explainability around AI algorithms, also known as black-box decisioning, makes AI-powered systems susceptible to delivering erroneous outcomes. While AI systems can provide immense value to businesses, from expediting manual processes to reducing costs, they are inherently vulnerable to developing biases as they explore variable combination nonlinearities while optimising performance. The risk of using AI ultimately depends on the algorithms chosen by data scientists, the AI modelling process employed, as well as the datasets that they are fed.

One of the risks currently is that the cognitive biases of data scientists or data sets that are non-representative of the larger population can be unintentionally incorporated into the models. Without a comprehensive understanding of machine learning algorithm, business leaders are often unable to properly monitor the decision-making capabilities of AI solutions and prevent undetected biases from delivering skewed outcomes.

PERILS OF UNDETECTED BIASES

With more than half of organisations in APAC having employed AI solutions that are only in early maturity stages, with no formal strategy or coordination around their AI deployment, the implications of bad decision-making could be far-reaching. Skewed AI-powered decisions would affect the very livelihoods of individuals by obstructing access to loans, healthcare or even job opportunities. Board of directors that fail to embrace their responsibility to deliver safe and unbiased AI could potentially be battered by regulation, a cornucopia of litigation and powerful AI advocacy groups.

Financial institutions could also become susceptible to a relatively new phenomenon called adversarial AI. This is where bad actors "game" or manipulate AI models for their own benefit. These attacks could take the form of data poisoning at the time of model training, altering the training data used to build AI models, and probing attacks that explore AI model outcomes to a variety of different inputs to provide attackers with vulnerabilities associated with the model.

MAINTENANCE AND MONITORING

Thankfully, businesses recognise that their approaches to AI deployment need to change. Nearly 63 per cent of business leaders expressed confidence that AI ethics and responsible AI will become a core element of their organisation's strategy within two years. In order to achieve responsible AI deployment, enterprises need to work internally to develop rigorous processes and formal model development governance that spot and remove bias across the model development lifecycle.

The implementation of codified policies that clearly define responsible AI model development tasks then provides a framework for the active monitoring of AI models - to ensure decisions they produce are accountable, fair, transparent, and responsible. Model development governance processes include regular maintenance and monitoring, so that chief analytic officers can evolve and improve over time.

Business leaders must actively strive to eliminate chances for biased decision-making processes to protect their organisation from legal liabilities and business risks. AI has the power to transform the world, but as the saying goes: with great power comes great responsibility.

• The writer is chief analytics officer at FICO, responsible for the analytic development of its product and technology solutions.