The Ethics Institute defines ethics as the balance between what is good for ourselves and what is good for others. When applied to AI, this definition is extended to include the goal of leading to individual and collective flourishing, aiming to build societies that work for everyone while protecting individual rights and liberties. A crucial element is the concept of “human-centric AI,” which means developing AI that strengthens human dignity, capabilities, and decision-making, rather than weakening or replacing them. The concern is that AI could potentially prioritise what is good for machines or corporations over what is good for human beings.

The guidebook (freely available for download from The Ethics Institute here) distinguishes between machine learning and generative AI. Machine learning is described as more analytical, used for tasks like identifying patterns in data (e.g., fraud detection in medical aid applications). Generative AI, on the other hand, exemplified by tools like ChatGPT, learns patterns to generate new information rather than just analysing existing data. This distinction is crucial because they present different types of ethical risks and require different management approaches. Generative AI poses a “diffuse risk” due to its widespread, everyday use by employees, necessitating systemic culture interventions like widespread training and policies. Machine learning, being more project-based, presents a “contained risk” that requires a systemic risk management intervention with specific processes, checks, and balances.

For Large Language Models (LLMs), a primary ethical risk is accuracy, as LLMs can “hallucinate” or generate false information, such as fake legal case law. This highlights the need for users to understand that LLMs predict logical words and phrases rather than necessarily providing accurate or truthful information. Another significant risk is data security, as employees often upload copyrighted or sensitive company information to generative AI tools without proper safeguards, risking intellectual property breaches and training AI on proprietary data.

For project-based AI, key ethical risks include:

• Bias and Fairness: AI systems can exhibit bias if trained on biased data, leading to unfair outcomes, especially in sensitive areas like recruitment or parole decisions.
• Data Privacy: Ensuring that data used by AI is sourced appropriately, is accurate, and is used only for its intended purpose, aligning with privacy regulations like GDPR.
• Transparency and Explainability: The difficulty in explaining how AI makes certain decisions can be an ethical concern, especially when those decisions impact individuals.
• Recourse: Individuals affected by AI decisions need clear avenues for human recourse, as seen in cases where people were unfairly penalised by automated systems with no human oversight.
• Autonomy: AI should not be used to manipulate people into harmful actions, respecting their agency and personhood.
• Accountability: Determining who is accountable for AI’s actions, as machines do not possess moral agency.
• Job Replacement: The potential for AI to cause significant job losses, which is a major concern, particularly in developing countries.

Organisations should adopt a risk-based approach tailored to their level of AI use.

1. Governance Oversight: At any level of AI use, there needs to be some form of governance oversight, with the board being at the forefront of managing emerging technologies.
2. Commitment to Responsible AI: Before implementing any measures, organisations must decide if responsible AI use (human-centric AI) is a core value.
3. Set Standards and Codify Guidance: For diffuse risks (e.g., employees using LLMs), simple guidelines are essential. A sample code might include ensuring accuracy, avoiding bias, safeguarding data security, and reminding employees that they remain accountable for their AI-assisted output.
4. Socialise and Monitor: Policies need to be communicated and made accessible to everyone, followed by monitoring AI use and integrating feedback into reporting cycles.
5. Project-based Risk Management (for contained risks):

• Assign Responsibility: A senior human should be assigned overall ownership of each AI project.
• Consider Team Diversity: Projects should involve diverse teams in terms of demographics, skills, and backgrounds, and testing should reflect this diversity.
• Assess Against Standards: Projects must be evaluated against established ethical principles (e.g., beneficence, avoiding harm, safeguarding human autonomy, competence, transparency, explainability, oversight, and accountability).
• Keep a Human in the Process: Maintain human involvement in decision-making, even when AI provides recommendations.
• Provide Recourse Avenues: Ensure clear and accessible human-led channels for individuals to seek redress if they are negatively impacted by AI decisions.
• Regular Review: Continuously test and review AI models, for example, every two years or with every major revision, and actively solicit feedback from users.

Human-centric AI is central to the ethical framework, aiming to develop AI that strengthens human dignity and capabilities. It ensures that AI serves human well-being and progress, rather than diminishing it. This means AI should augment human decision-making, making people more reflective and capable, instead of replacing their wisdom or creative abilities. The goal is for AI to contribute to individual and collective flourishing by building societies that work for everyone, protecting individual rights and liberties, and fostering greater human consciousness, awareness, and intellectual abilities. The concern is that if AI is not human-centric, it could lead to societies where machines or corporate interests are prioritised over human needs, potentially weakening human agency and leading to undesirable societal outcomes.

It is problematic to view AI, especially large language models, as having moral agency because they do not possess the capacity for moral choice, understanding of right and wrong, or accountability in the human sense. As explained, LLMs operate by predicting the “next logical word or phrase” based on patterns in their training data. They are not conscious entities attempting to lie or deceive; rather, their “apologies” or “emotive” responses are simply generated based on these predictive patterns. Human moral agency is deeply embedded in vulnerability—the understanding of one’s own and others’ susceptibility to harm. AI, not being vulnerable in the same way, cannot experientially comprehend this fundamental aspect of human ethics. Therefore, attributing moral agency to AI is an unrealistic expectation that can lead to a misunderstanding of its capabilities and limitations, potentially leading to a “dehumanisation or unhumanisation” of governance and decision-making.

Broad societal concerns related to AI include the digital divide, societal polarisation, job loss, environmental impact (due to high energy consumption), and even existential threats to humanity. While AI offers opportunities to address these issues (e.g., better education, more meaningful work), there’s a dilemma: the opportunities are often organisational (enhanced profitability, effectiveness, efficiency), while the risks are frequently societal. This creates an ethical challenge where organisations might prioritise their immediate benefits over the broader, long-term societal detriments, such as known job losses or significant energy use. The sources emphasise that it is a risk to not utilise AI, as this could lead to falling behind and missing potential human-centric opportunities. However, this must be balanced with a clear understanding and mitigation of the societal risks involved.

International standards and frameworks like ISO/IEC 42001, the NIST AI framework, and the OECD AI principles are crucial for governing AI ethics. While laws often lag technological advancements, these standards provide a “minimum acceptable societal norm” and guidance for organisations. The EU AI Act, for instance, offers a useful framework for categorising AI projects by risk (high, intermediate, low) and calibrating the necessary governance accordingly. By aligning with these standards, organisations can ensure they are considering ethical implications, such as human values and capabilities, throughout their AI development and deployment. Leveraging these documents can help organisations establish robust governance policies, set a clear “tone from the top,” and ensure mindful and purposeful management of AI, ultimately building trust and avoiding negative impacts.