AI Ethics & The Unexpected Golden Rule

An exploration of AI Ethics, Bias, and Governance

By Mark Gilmor & Jonathan Baier

We are in an age of firsts. I never in all my machinations of my own imaginary adventure in the Star Trek universe did I believe I would be talking about the ethics of Artificial Intelligence but… here we are. How we move forward responsibly, safely, and ethically are the key answers that AI regulation and governance aim to address. To do so we must first understand what is AI ethics and governance, what is the difference between AI ethics and bias, and how can we create accountability in our AI ecosystems.

Ethics is more important than ever.

When you're rolling out AI solutions within your organization, it’s essential to weave ethics into every step of the process. This means being transparent about how AI systems work and making sure they can explain their decisions in a way that everyone can interpret, understand and trust. Openness about data sources and the algorithms at play helps build confidence and clarity.

AI Bias and Alignment

Addressing bias is crucial. AI should be fair and equitable, avoiding any discrimination based on race, gender, age, or other characteristics. We need to constantly check and correct any biases that might creep in during the AI's learning process.

Beyond avoiding biases in responses, we also need our AI system to guide us safely and responsibly. Making suggestions that but human safety at risk or assist with criminal intent are areas where these systems need to steer clear. In addition, ensuring that the responses are respectful and not offensive is key. Harassment and hate speech can all too easily find its way into these large models which train on a large corpus of text from the uncensored internet. 

Reinforcement Learning from Human Feedback is a standard approach as described in this paper by the team from Anthropic, Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback.

In addition, annotated datasets like AEGIS from Nvidia provide a path for teams looking to fine tune responses. That said each organization is likely to have its own areas of concern and preferred style of response. Therefore the well-known foundational models may not be ideal without further alignment or tuning.

Privacy, Accountability and Security

Protecting privacy is non-negotiable. We must handle personal data with the utmost care, adhering to all privacy laws and regulations. Alongside this, robust data protection measures should be in place to guard against breaches and unauthorized access.

Accountability is key. Define who’s responsible for the AI's outcomes, ensuring that there’s a clear chain of responsibility for both successes and setbacks. Establishing governance frameworks can help oversee development and ensure ethical standards are met.

And although it may not be a surprise it is still worth mentioning that security can’t be overlooked. AI systems must be safeguarded against cyberattacks, ensuring that they remain reliable and resilient. This includes making sure the systems function correctly under various conditions and threats inclusive of the potential:

• Data integrity issues

• Security Alerts

• Odd System behavior

• Performance Issues 

• Unreliable or inconsistent information

…among others.

Compliance

Legal compliance will become more and more stringent over time. Stay informed and adhere to all relevant regulations, from data protection laws to industry-specific guidelines. Respect intellectual property rights to avoid legal complications.

The EU has already released the first regulation for AI with the EU AI Act. The US, Japan, Singapore and several other countries have drafted guidance and / or may soon pass laws of their own on the safe and responsible use of AI.

Design with humans in mind

Human oversight is also vital. AI should support human decision-making, not replace it entirely. There should be mechanisms for human intervention in critical decisions, allowing us to retain control and make judgment calls when necessary. 

Key to making judgment calls is our ability to interpret how these models create their responses. While interpretability is still an evolving practice, a recent paper from Anthropic holds promise for a much easier path for understanding and guiding the responses that a model produces.

We must consider the broader impact of AI on society and the environment. We must be mindful of job displacement, economic inequality, and AI's environmental footprint. AI should be used to promote social welfare and address societal challenges, striving for the greater good.

Human centric engagement with stakeholders is crucial. Include diverse perspectives in the conversation to gather valuable input and address any concerns. Providing education and training will help everyone understand and responsibly use AI. 

Schedule an appointment here on how to make that happen.

All of us must ensure that our AI solutions are not only effective but also responsible and fair, benefiting both the organizations we work in and society as a whole. In that, the golden rule has moved realms and enables us to build great systems that will serve the future if done correctly.