Ethical, Moral and Representation Foundations for AI Standards and Regulation
In-Context Learning Strategy
Summary
-AI regulations: documents and directions
-Ethical and moral frameworks for AI laws
-Pre-training AIs with the Moral Foundation Theory: in-context learning
-Representation as code
-Regulatory automation: architectures for automated compliance monitoring
Early Recommendations Responding to Alarming Predictions
The area of AI standards and regulations is a topic for extensive studies across domains. This text is part of a growing work on regulatory and compliance frameworks for ethical and responsible AI applications.
For more than a decade, the ethics of computing algorithms and data representation have become important among scientists, corporate leaders, and policymakers. The debate has intensified because we are clearly in the "summer of AI," a period in that new successful architectures and the pervasive use of AI algorithms are widely discussed.
AI regulation appears to follow many sets of guiding principles that may or may not lead to achieving consistency in the global efforts to keep AI in check. Below I will outline some concepts that could help formulate principles for a consistent and comprehensive regulatory framework.
We are all aware of the rapid growth and development of AI and ML technologies, which have immense potential for good but pose risks if not adequately managed and regulated. Moreover, we are only aware of a part of commercially available applications. Therefore, it is safe to assume that many research areas, inventions on the drawing boards, and more secretive applications will also have a tremendous impact. In this context, countries investing in high-tech R&D and their governing bodies decided to look into regulatory measures that could curb the largely unknown implications of AI. In 2019, over 40 countries supported a global governance framework for AI through the Organization for Economic Co-operation and Development (OECD). This organization created the Principles on Artificial Intelligence that initiated efforts in political and legislative circles to do something constructive.
Source: https://oecd.ai/en/ai-principles
In Europe, governments have also published regulations and guidelines for AI around the discussion presented in the General Data Protection Regulation. As a first step, the European legislative bodies developed a white paper: On Artificial Intelligence – A European Approach to Excellence and Trust, which discussed frameworks for private citizens, businesses, and regulatory principles shaping public interest legislation.
Based on those early publications, it was clear that further regulatory efforts would follow in the spirit of the main points the OECD and European Commission have put forward. For example, the UK Government Report on Artificial Intelligence in the Public Sector discusses AI in the industrial and government settings, the importance of developing and enforcing (future) standards, and accountability. Those standards were not clearly defined when the British government report was published.
General Data Protection Regulation (GDPR) is another large document that codifies general rules for protecting personal data processed algorithmically and stored in data centers. This extensive document requires a separate discussion just to summarize the main points. It appears, however, that GDPR and prior initiatives require a common foundational framework for analyzing and implementing the rapidly growing provisions. A set of guiding principles will help achieve consistency in the global efforts to keep AI in check. Those principles will have to be very well represented (disambiguated) as actionable, if not computational, directives that can be added or applied to AI systems as a verification or compliance procedure. While the focal point of AI regulation will be any architecture, the large machine-learning pipelines appear to be the main challenge. The regulatory provisions should be highly explainable and transparent. Converting law to system designs capable of surveillance of AI systems is, in itself, a fascinating topic that will surely energize the research community. Transparent regulatory systems overseeing and benchmarking all kinds of AIs will probably require cooperation between several branches of computer science to formulate architectures based on classical, fuzzy, or nonmonotonic logic, and knowledge representation, to mention only a few.
Ethical, Moral, and Scientific Frameworks for Political Decisions and Regulatory Efforts
Moral and ethical principles provide a deep set of guidelines for determining the direction of regulatory efforts. Besides the legislative processes, we need a consistent guide for creating just and fair laws in society. For us, those principles help to develop and solidify social norms and all areas where a fundamental level of harmony is essential for a community to prosper. More formally, morality and ethics play a role in ensuring that legislation supports and respects the rights of all individuals and contributes to the common good. Additionally, laws based on moral and ethical principles planted in social norms tend to have more significant public support and legitimacy.
Besides morality and ethics, there are other powerful ramifications like existing constitutional laws, legal traditions, precedents, or international treaties. However, those complicate the picture since more fundamental laws tend to be older, further removed from current norms, and may or may not be applicable to new and dynamically evolving technological environments. Fundamentally, those regulations are rooted in similar ethical and moral foundations. It is probably fair to assume that the challenge to apply those laws properly is one that teams of experts from intersecting areas of science and law can overcome.
Source: https://digitalcommons.law.seattleu.edu/sulr/vol33/iss4/19/; Colin Prince, Moral Foundation Theory and the Law, 33 SEATTLE U. L. REV. 1293 (2010).
Scientific knowledge and evidence-based research into emerging AI technologies and applications are other areas that are relevant. Addressing many rightful alarmist concerns from authors and scientists can inform and shape legislation. Particularly in areas related to emerging technologies, we need to assume that unknown or unexpected effects will happen. Those issues will almost certainly hide in complicated mathematical, statistical, or thermodynamic formalisms. Those technical discussions alone will lack explaining powers and will not give clues into the effects of convolutions and transformers on our everyday life or society. In fact, the very opposite point appears to be true—emerging, new or unintended effects surfacing in hidden layers of artificial neural networks.
"…people can now visualize how these models can learn from exemplars. So, my hope is that it changes some people's views about in-context learning," ... "These models are not as dumb as people think. They don't just memorize these tasks. They can learn new tasks, and we have shown how that can be done."
Source: https://news.mit.edu/2023/large-language-models-in-context-learning-0207
If unexpected side-effects of AI continue to grow in severity the legislation bodies dominated by politics might overreact and overtighten the AI regulations. This might be catastrophic and lead to more unintended consequences. Therefore a solid moral and ethical framework that is in place and in the public eye is so important.
Moral Foundation Theories: a Pre-Training Data Set for Large Machine Learning Models
Some scientists in the field of machine learning believe that large machine learning models are capable of in-context learning due to their training methodology and size. The models, such as LLMs, have billions of parameters and are trained by “reading” vast amounts of text from large corpora. The size of the input text makes it difficult to control in terms of content, bias, or simply the accuracy of the included information. Nevertheless, the model can be trained on internal patterns present there. As a result, when faced with new tasks, the model can recognize similar patterns it has seen in its training data rather than generate new activations and neural connections to learn a new task "from scratch."
Source: Jonathan Haidt (2013) Moral psychology for the twenty-first century, Journal of Moral Education, 42:3, 281-297, DOI: 10.1080/03057240.2013.817327
We can hypothesize that large models can be effectively trained on synthetic data that contains patterns coming from regulatory guidelines. Once the "guiding principles" are in place, the models can be further trained in the subject matter domains and continue to learn from new examples in a responsible and ethical manner prescribed in the pre-training set.
If we look into ways to infuse AI's with a moral compass, pre-training might be a good start. The data could be based on the moral foundation theory as a starting point since there is probably significant consensus about the content. The technical side of translating this work into a training set is a separate issue. Let's just look at a data source for now.
https://pages.stern.nyu.edu/~jhaidt/articles/miller.2008.roots-of-morality.science.pdf
Moral foundation theory argues that there are fundamental sources or building blocks of morality regardless of the culture. The foundations were created over centuries and continued to evolve in order to address existential challenges faced by people in their social and physical contexts. Many of those challenges were fundamental to the survival of the society, e.g., the importance and protections of caring for offspring, forming and maintaining strong social bonds, and the concept of fairness.
According to the theory firmly documented in many sources, there are several interrelated moral foundations: care/harm, fairness/cheating, loyalty/betrayal, authority/subversion, and sanctity/degradation. These foundations are the building blocks for our moral beliefs and values and can be used to understand and predict moral judgments and behavior. In addition, the theory suggests, while the principles might be common, people surely have different levels of sensitivity to the aspect of the theory and that a combination of biological, environmental, and cultural factors influences these differences. When looking at the main foundations, it is easy to realize that those are pretty broad in scope and challenging to interpret. Although these concepts have been the focus of research in such areas as psychology, behavioral science, neuroscience, and many others, many believe there is much more room for investigating this area. For example, some argue we need to extend the models by including psychological foundations of morality based on analyzing the concepts of loyalty, respect for authority, and spiritual purity. This is for sure not the place to present even a thin slice of this research. However, scientists who focus on morality look for ways to extend the theory by adding new features, similar to machine learning engineers who look for more features in the training data. Going further, many extensions of the basic idea of morality influence various historical, philosophical, social, or religious contexts, i.e., avoiding harm, honesty, compassion, fairness, justice, loyalty, authority/respect, respect for human dignity, fairness, responsibility toward individuals or groups, and equality.
Source: https://alextheyounger.medium.com/morality-in-the-modern-world-4b37c2ecfad1
If we look at morality from the perspective of the automated practical application carried out by computer systems, perhaps even autonomously, we need to build a representation. Only well-specified sets of guidelines and rules of action can help produce layers of control, compliance, and audit for troublesome AI systems.
AI Regulations: the Question of Representation and Data Structures
Can AI be regulated through manual processes, manual audits, or experts performing system analysis without automated tools?
Probably not.
Large AI models or (more traditional) rule-based systems with explicit rules in them are very hard to analyze. We need systems, most likely based on AI algorithms, that can do this tough job. Therefore we need to answer another question: how to convert ethical and legal rules and, most likely, social norms into data structures that can be executed by a computer. There are two main types of representation: explicit rules represented in a logical system or a lot less explicit (hidden) training set for a machine learning system resulting in the creation of a model. This is surely a topic for a separate discussion.
Source: https://suffolklitlab.org/legal-tech-class/docs/representing-rules/representing-rules/
The starting point for any representation appears to be clear: natural language statements. When transforming laws written in natural language into code, we are converting patterns with uncertainty and ambiguity embedded in them into data structures that may or may not have the capacity to capture all latent information. AI and natural language processing (NLP) scientists have struggled with this problem for decades.
Turning an unstructured or semi-structured rule into an embedding or code through consultation with a subject matter expert can be difficult because we might not reach a consensus on interpretation. Legal experts typically apply the regulations to specific situations, knowing there might be a deep debate in courts where the description of the context and interpretation of every statement (or comma) might be questioned. In this process, where computer systems must execute the rules, we need similar mechanics for review, evaluation, and possibly verification. The difference is that our “output” will not be another legal document written in, inherently ambiguous, natural language. Instead, it must be code or knowledge representation that works with the algorithm that needs the correct structure.
The concept of representing legal rules and decisions in computer code is called Jurisprudence as Code. It is a process that involves translating legal concepts and principles into a form that computer systems can process and interpret. This is the core issue that we will have to tackle. We must make legal rules and decisions very accessible, consistent, and transparent. By representing legal rules and decisions in data structures or code embedded in a programming language, it becomes possible to automate certain decision-making tasks, such as compliance analysis and dispute resolution where complexity is high. Jurisprudence as code has the potential to improve the interpretation and application of legal rules. The question is, can this be done at a legislative level with the assistance of AI and computer science communities? It would be ideal if all parties involved decided to disambiguate the AI regulations and come up with an elegant way to infuse those rules into all required areas. As you can see, we are clearly stepping on the legal turf. Even if these ideas gain traction in the legal community, it will be a long process.
Source: https://research.cornell.edu/news-features/coding-justice-law-digital-age
Fortunately, we can tackle the representation issue from the technical side as well. It will be the topic for the next post. If I were going to bet, I would put my money on technology to deal with the abovementioned issues.
Source: https://api.semanticscholar.org/CorpusID:229722844
Conclusion
Regulations and frameworks for curbing AI might be the most important task for AI scientists. Unfortunately, there is a good chance that alarmist voices about "runaway AI" or AI being our "final invention" are true.
