Event Round-Up: Introducing techUK’s Inaugural Digital Ethics Unconference with Responsible AI Practitioners

This group of responsible AI practitioners differentiated existing AI governance SaaS platforms into three categories:,compliance and standards, governance and technical assessments. In each category, the group noted the advantages and disadvantages they are experiencing when engaging with each type of tool/platform.  

1. Regarding compliance and standards, there are several advantages to consider. These include the ability to integrate with existing compliance processes and ML Ops, support for defining AI in governance contexts, and help in managing the rapidly changing landscape, especially for global organizations. Standards can also support interoperability and assist in managing regulatory overlap or conflicts across multiple jurisdictions. However, there are drawbacks to this approach. It may lead to responsible AI becoming a mere tick-box exercise, oversimplifying complex ethical considerations. Value-laden terms like 'fairness' might be reduced to simple compliance metrics, potentially missing the softer impacts of emerging technologies. Additionally, these frameworks may not adequately prompt corporations to consider "what is the right thing to do" beyond compliance, and they might struggle to address highly context-specific risks. 

2. In terms of governance, platforms offer several benefits. They allow for risk management at a portfolio level, providing a view across all projects and systems. They can enable automated aspects of governance, driving operational effectiveness and promoting a pro-governance culture. Users can link training requirements to approved roles and responsibilities, making accountability clear. Unlike oversight in standardized formats, governance platforms can help prevent "lost" or "shadow" AI. However, challenges remain. Determining acceptable levels of risk is subjective, and there's a need to properly account for organizational risk appetite alongside regulatory requirements. This impacts inherent risk impact scoring. Another challenge lies in determining the appropriate rigor for system documentation proportionate to the risk involved. 

3. Technical assessment platforms bring consistency to evaluations and support responsible engineering by standardizing ways to assess fairness, drift, and robustness across an organization. However, they face limitations. The diversity of tests required means that one tool cannot fit all needs. Acceptable measures differ per jurisdiction, particularly for concepts like fairness. There's a risk that automation might lead to missing important ethical risks. Moreover, the outputs are highly dependent on the accuracy and rigor of inputting models and data, raising questions about how to best train and educate people on the right ways to evaluate AI systems. 

This group of participants spoke about the need for maturity models, assurance teams for methodology and customer joint risk management. The poster suggested there is a lack of oversight, asking should systems and processes be codified and audited? Further, pointing to the need for ‘teeth’ in corporate ethics culture for actions to be taken by individuals with domain knowledge. This group of participants posed the following questions:  

• How do we measure responsible AI? 

• What is unique to AI’s risk profile? 

• Who is marking the practice of ethics in companies, we should not grade our own homework? 

• What is the board’s role in supporting responsible AI? 

• Questions of liability: who goes to jail when/if something goes wrong, and harm is caused?  

• Is it truly too soon to regulate AI? 

This group of RAI practitioners presented their dialogue in the format of questions, stating ‘it wasn’t just questions... but we do have a lot of questions!’.  The poster included the following questions which suggest areas of dialogue for responsible AI practitioners to keep considering:  

• How do we achieve cultural change? 

• How do we change business thinking and not just dump responsible AI on technical teams? 

•  How can people be empowered to act ethically and cause a fuss when harm is caused by emerging technologies? 

• What role does accreditation play – is it more of a driver? 

• Ethics can seem conceptual – how do we make it concrete and communicate why? 

• How do we develop diverse skills and foster different perspectives? 

This group of RAI pracitioners kept their poster to the point by outlining five things to keep in mind when considering GenAI’s impact on IP: 

1. Publicly accessible doesn't equal public domain  

2. Should we protect style? 

3. How much control should we have over expression? 

4. Solutions for ‘opting out’ works for training  

5. Technical protections exist such as ‘poisoning’ models  

This group suggested a few ways that organisations can work to embed ethics in their culture, while also asking pressing questions: 

Answers 

• Red teaming and games – give people chances to practice  

• Horizon scanning allows for forward thinking 

• Information literacy  

• Tap into existing teams and be receptive to their needs – champions and representatives  

• Combine policies, initiatives, champions and leadership  

• Be sector specific  

• Enable teams to do their jobs ethically  

Questions 

• The role of HR – onboarding psychological safety – what new skills will they need? 

• What incentives do you need to lean on (safety for brand)? 

• How can we train? Is it too risky to put too much responsibility on non-experts? 

This group drew a large vehicle and suggested that AI is currently like a car with an engine that keeps getting larger but has no brakes or wheel to steer. Noting that ethics is not easy to codify as it is socially and culturally dependent. Bridging the gap between proportional mechanisms for risk and having iterative improvements to operationalised ethics through assurance techniques. Finally, operationalising ethics is often about optimising while working through trade-offs. 

This group revisited where tech has gone wrong in the past, identified the harms and lessons learned to inform and support best ways forward.  

1. Ethics must be considered at the start – have organisations thought about the potential harms before beginning the development  

2. 1st level causes: immediate oversights (“we forgot”, “we rushed”), versus 2nd level causes: systemic issues (no processes in place) 

3. Oversight, due diligence - ‘nominated manager; VP sign-off' 

4. Spillover – trying to be ethical but end up causing another problem/harm 

This group of RAI practitioners decided to translate their conversation into three risk categories that should be considered in neuroethics, namely corporate, societal and individual. This is accompanied by real world use cases and regulation for neuroethics.