OpenAI introduced a “Preparedness Framework” just before the 2024 holiday season. This framework grants the company’s board of directors the authority to revise decisions made by CEO Sam Altman and the leadership team. It is currently a work in progress, labeled as “beta”, and was developed in response to the events of November 2023 when Altman was briefly fired from his own company due to ‘AI safety’ concerns before being reinstated. The framework serves as a demonstration of OpenAI’s commitment to responsible and safe AI practices.
The core of this framework employs “scorecards” to assess the potential risks associated with AI technology. These scorecards are regularly updated with new information, enabling quick adjustments if any aspect appears too risky. OpenAI’s aim is to base decisions on real data rather than mere speculation.
In addition, specialized teams have been established to oversee various aspects. One team focuses on the technical aspects to ensure safety, while another group, known as the Safety Advisory Group, reviews their work and provides recommendations to the company’s leaders and the board. Moreover, they plan to conduct safety tests and seek independent verification to ensure that they are following the correct procedures.
Although OpenAI’s framework is significant, it is not the ONLY one in existence. In this segment of the AI series, I intend to introduce and discuss these frameworks. It is essential to begin defining the characteristics of “Good AI.”
PART 1: CHARACTERISTIC OF GOOD AI
The concept of “Good AI” encompasses the entire lifecycle of AI systems, spanning from data acquisition and model development to system development and deployment, and finally, to continuous monitoring and governance.
Let’s begin by defining the predominantly accepted characteristics of good AI, while excluding AI performance aspects such as Accuracy and Precision that are more performance related and were being discussed for years.
ACCOUNTABILITY: Accountability runs through the entire lifecycle of an AI system, and AI systems should be designed, implemented, and operated in alignment with human values.
Are there universal human values across the globe?
That is debatable, but there are several universal human values, such as honesty and integrity, respect, fairness, and justice, that are recognized and shared across cultures and societies around the world. While the specific emphasis on these values and their interpretation may vary from one culture to another, the core principles remain largely consistent.
A practical example of making AI models responsible is seen in social media like Twitter and Facebook. These platforms use AI models to identify and remove content that violates their community guidelines, such as hate speech, graphic violence, or misinformation. However, these systems are not infallible, and sometimes they make mistakes or exhibit bias.
In such cases, the platform should continuously update and refine its content moderation AI model based on feedback from human moderators, user reports, and audit findings.
TRANSPARENCY and EXPLAINABILITY: Good AI systems are transparent and provide explanations for their decisions. For instance, a loan approval AI system can explain why it approved or denied a loan, helping borrowers understand the process.
REPRODUCIBILITY for an AI model is its ability to recreate and replicate the model’s results and performance using the same or equivalent data, code, and configuration.
Let’s say if a group of researchers developed an image classification AI model, achieving a reported accuracy of 95% on a specific dataset and published their findings along with the model’s architecture, hyper-parameters, and a summary of their data preprocessing steps.
If another group of researchers, interested in the same problem, could reproduce the same results with the same model and the same method, this is the reproducibility.
FAIRNESS and BIAS MITIGATION: Good AI is designed to be fair and free from bias, treating all individuals equally. As an example, a facial recognition technology, when developed responsibly, doesn’t exhibit racial or gender bias in its identification capabilities.
Addressing fairness is one of the most challenging aspects because AI systems learn from massive volumes of data; if that data contains biases, the algorithms may unintentionally replicate them.
ADVERSARIAL ROBUSTNESS and SECURITY: Robustness refers to the ability of an algorithm or system to deal with execution errors, erroneous inputs, or unseen data. Robustness is usually an important factor affecting the performance of AI systems in empirical environments. The lack of robustness might also cause unintended or harmful behaviour by the system, thus diminishing its safety and trustworthiness.
Adversarial Attacks:
AI models are susceptible to adversarial attacks, in which malicious actors manipulate input data to deceive or compromise the model’s performance. Safeguarding AI models from adversarial threats, such as data poisoning, theft, tampering, and reverse engineering, is of paramount importance. Addressing this challenge requires a comprehensive understanding of adversary tactics and techniques.
One valuable resource in this regard is MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems), which offers insights derived from real-world attack observations and realistic demonstrations conducted by AI red teams and security groups.
The OWASP Machine Learning Security Top Ten analyzes the most common vulnerabilities associated with machine learning. In Part 3 of these series, I intend to deep dive in adversarial attacks.
PRIVACY: Privacy protection is all about keeping unauthorized people from using data that could identify a person or a household, whether directly or indirectly.
The data used by AI models could cover a wide range of personal information, including things like names, ages, genders, facial images, fingerprints, and more.
Recent guidelines in AI ethics highlight that privacy is a major concern. Government agencies are making more and more rules to control how data privacy works, and one of the big ones is the General Data Protection Regulation (GDPR).
PART 2: FRAMEWORKS FOR GOOD AI
- NIST AI Risk Management Framework (RMF)
In January 2023, the U.S. Department of Commerce, in collaboration with The National Institute of Standards and Technology (NIST), released the inaugural edition of the AI Risk Management Framework (AI RMF) along with a complementary playbook.
AI Risk = Enterprise Risk
The NIST AI RMF underscores the importance of integrating AI risk management into an organization’s broader enterprise risk management framework. The AI RMF highlights the significance of taking into account AI-specific factors when determining an organization’s risk thresholds, prioritizing risks, and integrating AI effectively. Additionally, it emphasizes the need to identify and address overlapping risks, such as those related to data privacy and environmental impacts associated with AI’s computational requirements. The framework encourages organizations to take proactive steps by establishing clear roles and responsibilities to enhance risk management.
However, it’s worth noting that one of the key criticisms of NIST’s AI RMF is its high-level nature. Defining what constitutes a risk can be subjective, and the guidance provided offers limited practical advice on how organizations can mitigate these risks. This limitation stems from the fact that AI is still an emerging technology, and its deployment and use cases vary widely, making it challenging to prescribe a one-size-fits-all approach to risk mitigation.
2.ENISA (The European Union Agency for Cybersecurity):
On June 7, 2023, the European Union Agency for Cybersecurity (ENISA) released a report titled ‘Multilayer Framework for Good Cybersecurity Practices for AI’ in response to the challenges posed by artificial intelligence (AI) and cybersecurity. The report aims to establish a strong framework for cybersecurity in all stages of AI, from its creation to retirement.
The report/framework consists of three layers, each addressing different aspects of cybersecurity.
- Foundations of Cybersecurity: This layer focuses on fundamental cybersecurity measures for the technology infrastructure supporting AI systems. It emphasizes compliance with EU laws, including data protection and cybersecurity regulations. Companies using AI should have a dynamic security process, including risk analysis and management.
- AI-Specific Cybersecurity: This layer deals with AI-specific security requirements applicable to all industries. It highlights the importance of AI-related legislation, such as the EU AI Act, and ongoing assessments of AI threats. Ethical considerations like transparency and fairness should guide these assessments.
- Sector-Specific Cybersecurity for AI: This layer provides sector-specific recommendations for Energy, Automotive, Telecommunications, and Health sectors, which already have relevant cybersecurity guidelines.
WHAT IS NEXT?
Following the Rules is Essential:
As governments and regulators make sure AI is safe and responsible, good AI frameworks must follow these rules to ensure organizations do the right things.
Global Ethics Guidelines:
Efforts might be made to create worldwide ethical rules for AI development. Ethical AI frameworks can actively help shape and support these global rules.
Ethics Oversight:
There will likely be more focus on making AI systems responsible and checking them carefully. More organizations may create AI ethics committees to watch over AI development and use.
Teamwork Among Different Experts:
Working together with AI experts, ethicists, policymakers, and others will likely become more important.
