We’re a long way off fully dependable AI systems. Rigorous failure point analysis determines the boundaries for their safe use and shows organisations where human oversight must remain. Ensure that clear divisions of responsibility are based on respective capabilities. Continuous monitoring will be necessary because the line dividing human and AI capabilities will shift and your organisation must be quick to react to this changing dynamic. This will also enable you to re-train humans and repurpose AI systems.

Continuous assessment of AI systems to understand how users are interacting with AI, to identify and mitigate potential risks of bias, misuse, and unintended consequences. It's an active research field and something Advai are involved with directly for our clients. Technical governance isn't always enough, sometimes you need to ensure the humans involved using the system appropriately. There is a knife-edge to walk that balances trustworthy systems with user over-reliance. On the one hand, organisations should strive for transparent, accountable, dependable and fair systems. On the other hand, organisations must be wary of their users becoming overly dependent on these systems and placing too much trust in the technology. 

Discussions and preparedness for AI’s future potential and how this may be accompanied by new ethical concerns. Although current AI technologies may not yet be advanced enough to integrate into certain workflows, rapid progress in the field suggests they soon will be. AI applications once thought to be decades away unfolded within years. Therefore, organisations should proactively consider and prepare for the long-term ethical risks of AI, such as laying down best data management practices. Strategic preparation will ensure you are well placed to mitigate future ethical challenges and enable you to take advantage of new AI opportunities fast.

There are varying opinions and contexts of AI ethics. Enabling clear stakeholder communication, and closely monitoring how internal or external users are using your AI systems, are necessary to incorporate diverse perspectives.

Effective cyber security and adversarial resistance is also an ethical priority, especially in the case that your organisation handles sensitive user data. Strengthening your data protection measures and identifying your AI system vulnerabilities are both vital for respecting and protecting your user privacy. Adversarial techniques exploit system weaknesses, potentially leading to data exposure and privacy breaches. This can occur through methods like model inversion attacks, which reconstruct sensitive information, or due to vulnerabilities in encryption practices and unintended data disclosures via API leaks.

AI can be an energy intensive process and in many cases it’s unclear whether the incorporation of incremental training data will improve performance. There may also be ways of reducing the carbon footprint of AI training/re-training by removing redundant processes. Understanding the tools that are available, such as Advai’s performance predictor model, can help you deploy AI more sustainably.