![[DSC Europe 23] Bunmi Akinremi - Ethical Considerations in Predictive Analytics](https://cdn.slidesharecdn.com/ss_thumbnails/bunmi-akinremi-ethics-in-predictive-analytics-231129100609-f96b9ee9-thumbnail.jpg?width=560&fit=bounds)
More Related Content
Similar to Professional Ethics------------------.pdf (20)
Recently uploaded (20)
Professional Ethics------------------.pdf
- 1. Professional Ethics Dr. / Dina Saif Ragab
- 2. Table of contents 01 03 02 04 The FAST Track Principles Fairness, Accountability, Sustainability, and Transparency Background Data fairness
- 4. The SUM Values are intended to guide ethical thinking in AI projects but don't directly address the processes of AI development. To make ethics more actionable, it helps to understand why AI ethics is essential. Marvin Minsky described AI as making computers perform tasks that would need human intelligence, highlighting the need for ethical frameworks as AI takes on more complex, human-like roles. The emergence of AI ethics focuses on responsible design and use as technology advances.
- 5. "Bridging the Ethical Gap: Accountability and Responsibility in AI Systems" Humans are held responsible for their judgments, decisions, and fairness when using intelligent systems. However, these systems are not morally accountable, leading to ethical breaches in applied science. To address this, frameworks for AI ethics are being developed, focusing on principles like fairness, accountability, sustainability, and transparency. These principles aim to bridge the gap between the smart agency of machines and their inherent lack of moral responsibility.
- 7. The FAST Track Principles (Fairness, Accountability, Sustainability, and Transparency) are essential pillars that guide teams in developing responsible, ethical, and socially-responsible AI systems. These foundational principles ensure a holistic approach that addresses critical ethical considerations at every phase of a project, from ideation to deployment. Here's a detailed look at each principle, highlighting their importance and application in real-world scenarios:
- 8. The FAST Track Principles Fairness AI systems should process social or demographic data equitably, without discriminatory bias. he designs should ensure equitable outcomes and avoid disproportionate impacts on any group. Accountability AI systems should be built with accountability in mind, enabling end-to-end traceability and review. This includes responsible design, clear implementation, and active monitoring protocols. Sustainability The development and deployment of AI should consider its long-term impact on society and the environment. This principle promotes the responsible use of resources, robustness, and overall system resilience. Transparency AI systems should communicate clearly with stakeholders, explaining their functioning, purpose, and potential impacts. Transparency is key for public trust and acceptance.
- 9. Getting to Know FAST Principles in AI FAST: Fairness, Accountability, Sustainability, Transparency. These four guiding principles may not always connect in a straightforward way. Accountability 牟 We all need to take responsibility for creating AI. 牟 This ensures that every step of the process is traceable and clear. Transparency 牟 We want AI decisions to be easy to understand and explain. 牟 Its important that everyone affected knows how AI impacts them.
- 10. Fairness and Sustainability in AI FAST: Fairness, Accountability, Sustainability, Transparency. These four guiding principles may not always connect in a straightforward way. Fairness 牟 AI should be friendly and treat everyone with respect. 牟 We aim to avoid harm and discrimination for all. Sustainability 牟 AI should be safe, ethical, and work for the good of future generations. 牟 Lets support positive changes for both society and our planet!
- 11. Summary: Fast Track Principles The principles of transparency and accountability provide the procedural mechanisms. and means through which Al systems can be justified and by which their producer and implementer can be held responsible, fairness and sustainability are the crucial aspects of the design, implementation, and outcomes of these systems which establish the normative criteria for such governing constraints. These four principles are all deeply interrelated, but they are not equal. There is important thing to keep in mind before we delve into the details of the FAST Track principles: 1) Transparency 2) Accountability 3) Fairness Are data protection principles and where algorithmic processing involves personal data, complying with them is not simply a matter of ethics or good practice, but a legal requirement, which is enshrined in the General Data Protection Regulation (GDPR) and the Data Protection Act of 2018 (DPA 2018).
- 12. Fairness in AI System Design and Deployment AI models rely on historical data, which may carry inherent biases. Data may contain social and historical patterns that reinforce cultural biases. There's no single solution to completely eliminate discrimination in AI systems. AI systems may appear neutral, but are influenced by the decisions of those who design them. Designers backgrounds and biases impact AI models. Biases can enter at any stage: data collection, problem formulation, model building, or deployment. Human Influence on AI Systems Challenges with Data-Driven Technologies
- 13. Approaches to Fairness in AI - Combines non-technical self-assessment with technical controls and evaluations. - Aims to achieve fair, ethical, and equitable outcomes for stakeholders. - Ensures AI systems treat all parties fairly. Importance of Fairness-Aware Design Principle of Discriminatory Non-Harm - A minimum threshold required to achieve fairness in AI systems. - Guides developers to avoid harm from biased or discriminatory outcomes.
- 14. Principle of Discriminatory Non-Harm Fundamental Fairness Principles for AI Systems Data Fairness Design Fairness Outcome Fairness Implementation Fairness requires that the data used in training and testing is comprehensive, accurate, and represents the full diversity of the population it will affect. If the dataset is not representative ,the AI could develop biased models. To build the AI model so that it doesnt contain any biased or morally questionable features. Designers need to avoid including certain variables (like race, gender, or socioeconomic status) unless they are genuinely relevant and justifiable. For instance, a loan approval AI shouldnt include factors that unfairly disadvantage certain groups without a valid reason. Outcome fairness is about the real-world impact of the AI system. After deployment, its essential to evaluate if the AI systems decisions have a fair and positive effect on peoples lives. For instance, if a healthcare AI model favors certain groups over others in terms of treatment suggestions, this would signal an outcome disparity. Implementation fairness focuses on the responsibilities of those deploying the AI systems. Proper training is crucial for the users of AI models (such as employees or decision-makers) to understand how to use these tools impartially and ethically. For instance, in hiring, this means HR professionals should interpret AI recommendations with an understanding of any possible biases, so the tool is applied justly. Goal: Prevent AI systems from causing unfair or biased impacts on individuals or communities.
- 15. Summary: Representativeness Sampling bias can lead to the underrepresentation or overrepresentation of disadvantaged or legally protected groups, which can disadvantage vulnerable stakeholders in model outcomes. To mitigate this, domain expertise is essential to ensure that the data sample accurately reflects the target population. Technical teams should, when possible, provide solutions to address and correct any representational biases in the sampling.
- 16. Summary: Fit-for-purpose and sufficiency In data collection, it is essential to determine if the dataset is large enough to meet the projects goals, as data sufficiency impacts the accuracy and fairness of model outputs. A dataset that lacks sufficient depth may fail to represent important attributes of the population, leading to potentially biased outcomes. Technical and policy experts should work together to assess whether the data volume is adequate and suitable for the AI systems intended purpose.
- 17. Summary: Source integrity and measurement accuracy Bias mitigation starts effectively at the data extraction and collection stage, where both sources and measurement tools may introduce discrimination into the dataset. Including biased human judgments in training data can replicate this bias in system outputs. Ensuring non-discriminatory outcomes requires verifying that data sources are reliable, neutral, and that collection methods are sound to achieve accuracy and reliability in results.
- 18. Summary: Timeliness and Recency Outdated data in datasets can impact the generalizability of a model, as shifts in data distribution due to changing social dynamics may introduce bias. To avoid discriminatory outcomes, its essential to assess the timeliness and recency of all data elements in the dataset.
- 19. Data Relevance and Best Practices Data Relevance & Domain Knowledge: Select appropriate data sources for reliable, unbiased AI. Leverage domain knowledge for choosing relevant inputs. Collaborate with domain experts for optimal data selection. Dataset Factsheet for Responsible Data Management: Create a Dataset Factsheet at the alpha stage Track data quality, bias mitigation, and auditability Record key aspects: data origin, pre-processing, security, and team insights on representativeness and integrity.
- 20. Do you have any questions? Thanks!