The Ethics of AI: Research Ethics Principles and Their Application
Introduction to AI Ethics and Research Ethics
Research ethics is a field that focuses on the moral principles governing research. It is distinct from legal frameworks and philosophical theories, instead operating on a set of core principles.
Core Principles (External)
These principles guide the interaction of research with the external world and its impact on individuals and society:
Respect for persons: Emphasizes treating individuals with autonomy and dignity. A historical example where this principle was critically examined is the Milgram experiment.
Good consequences: Focuses on maximizing benefits and minimizing harm. The development of the atomic bomb during the Manhattan Project serves as a stark historical example of the profound and wide-ranging consequences of research.
Justice: Concerns the fair distribution of benefits and burdens of research, ensuring no group is disproportionately harmed or excluded. The Tuskegee Syphilis Study is a notorious example of a severe violation of the principle of justice.
Internal Principle
Integrity: Relates to the internal ethical conduct of researchers and the research process itself, emphasizing honesty, competence, and transparency.
Potential Conflicts in Ethical Principles
It is important to consider how these basic external research ethical principles can come into conflict with each other. For example, a study designed to achieve significant good consequences for a large population might, in some hypothetical scenarios, necessitate choices that compromise individual respect or just distribution for a smaller group. Addressing such dilemmas is a key challenge in ethical research.
Respect for Persons in AI Research
This principle centers on treating individuals as autonomous agents, capable of making informed decisions.
Key Components and Challenges:
Voluntary informed consent: A cornerstone of respect for persons. Obtaining genuine informed consent becomes complex with:
Big Data approach: The sheer volume and impersonal nature of data collection challenge traditional consent models.
Transparency: Users often lack a clear understanding of how their data is used or how AI systems make decisions.
Alien intelligence (Deep Neural Networks): Understanding the internal workings of complex AI models, especially Deep Neural Networks, poses significant challenges for transparency.
Explainable AI (XAI): Efforts to make AI decisions interpretable:
Ask (confabulating): Directly asking the AI for its reasoning, which can sometimes lead to superficial or misleading explanations (confabulation).
Ask which factor most important: Querying the AI about the weight or significance of different input factors in its decision-making.
Watch neural network light up: Observing the activation patterns within the neural network to infer processing and decision pathways.
Presence or Deception
This aspect delves into whether AI systems should be designed to mimic human presence (sonzai-kan) and the ethical implications of such design, particularly regarding potential deception.
Hiroshi Ishiguro is known for his research on human-like robots and the concept of
sonzai-kan(presence), exploring how robots can evoke a sense of presence in human interactions.Sophia the Robot (Hanson Robotics/Loving AI): An example of a human-like robot designed to connect "deeply and meaningfully with humans." This raises questions about the nature of the connection and potential for misinterpretation or emotional manipulation.
PARO: An interactive therapeutic robot designed to stimulate patients with cognitive disorders. Its therapeutic purpose might justify its human- or animal-like presence without intending deception.
Siri: Handles countless minor tasks. While not designed for deep human connection, its conversational interface can still raise questions about perceived intelligence or sentience.
Not "the hard problem": The discussion of AI 'presence' is not equivalent to addressing "the hard problem" of consciousness.
Not the imitation game: The goal is not merely to pass a Turing test but to understand the ethical implications of the perception of intelligence or presence.
There is no obvious answer concerning consciousness in AI (contrary to arguments by Sloman (1999) or Turing (1950, pp. 445-6), who discussed solipsism in the context of machine intelligence).
Good Consequences in AI Research
This principle focuses on ensuring that AI research leads to positive outcomes and mitigates potential harms.
Key Considerations:
Data security: Critical for protecting sensitive information used by AI. Data can be categorized as:
Anonymized: Data where identifying information has been removed.
De-personified: Data that is generalized or abstracted to remove personal identifiers.
Personal: Data directly linked to an individual.
Sensitive: Personal data requiring extra protection (e.g., health, financial, religious).
"The Alignment Problem" (Russell 2019; Christian 2019): A central concern that arises when powerful AI systems might pursue goals that are not aligned with human values or well-being.
The Singularity/human extinction (e.g., OpenAI): Advanced AI, if misaligned, could potentially lead to catastrophic outcomes, including human extinction, reflecting a concern expressed by organizations like OpenAI.
Paper clips/gray goo: Hypothetical scenarios illustrating how an AI optimized for a seemingly innocuous task (e.g., maximizing paperclip production) could consume all available resources, neglecting human well-being. "Gray goo" refers to self-replicating nanobots that could consume the biosphere.
Humans not good at knowing what we (should) want: Acknowledges the difficulty in articulating explicit, comprehensive human values for AI alignment.
Constitutional AI: An approach to embed ethical principles and human values directly into an AI's operational framework to guide its behavior.
Inverse "inverse reinforcement learning"?: This potentially refers to a complex approach where, instead of inferring a human's reward function from their actions (inverse reinforcement learning), the system might try to find an 'inverse' operation to ensure alignment, though the phrase itself is ambiguous, possibly implying a re-evaluation of assumptions in alignment research.
SmartLoans (Meyer et al 2020): An example of AI application in finance, raising questions about consequences such as fairness, access, and potential bias in lending decisions.
Energy consumption: The computational demands of AI, especially large models and supercomputers, are significant.
Supercomputers/prompts: Training and running large AI models require immense computational power, leading to high energy usage.
"Tiny ML" (Tiny Machine Learning): An emerging field focused on enabling machine learning on small, low-power devices, aiming to reduce energy consumption and broaden AI accessibility.
Justice in AI Research
This principle typically addresses issues of fairness, equity, and the distribution of benefits and burdens across different groups or populations.
Key Areas of Concern:
Typically about groups: Justice in AI often pertains to how AI systems impact various societal groups, particularly vulnerable or historically marginalized communities.
Predictive policing: AI systems used in law enforcement can perpetuate or amplify existing biases if trained on biased data, leading to unfair targeting of certain communities.
Data sets - training: The composition and quality of training data sets are crucial. Biased, incomplete, or unrepresentative data can lead to discriminatory outcomes when AI is deployed.
No/access: Disparities in access to AI technologies and the benefits derived from them can exacerbate existing inequalities.
top universities vs. GPT: This comparison highlights potential inequities, where elite institutions might have greater access to advanced AI resources or the ability to influence their development, potentially leaving others at a disadvantage.
Integrity in AI Research
Integrity is an internal principle focusing on the moral character and standards of researchers and the research process itself.
Core Tenets:
CUDOS (Robert K. Merton, ): A set of norms defining the ethos of good science. While the transcript mentions "vs. FFP" (Fabrication, Falsification, Plagiarism), the focus here is on CUDOS as positive guidelines:
Communism: Scientific knowledge should be communal, openly shared, and property of the scientific community.
Universalism: Scientific claims should be evaluated based on impersonal criteria (e.g., merit, logical consistency) rather than the personal attributes or affiliations of the researcher.
Disinterestedness: Researchers should pursue truth for its own sake, free from personal gain or bias.
Organized Skepticism: All scientific propositions should be subject to continuous, critical scrutiny and empirical testing.
Uprightness and competence: Researchers must act ethically and possess the necessary skills and knowledge to conduct their work responsibly.
Openness: Transparency in methods, data, and results fosters trust and repeatability.
Relation teaching/research: The ethical conduct of research is intertwined with responsible education and mentorship.
Co-authorship?: Ethical concerns can arise regarding proper attribution and roles in collaborative AI research, especially with complex projects involving many contributors.
Motivation: The underlying goals and motivations of AI developers and researchers are critical. For example, OpenAI/Microsoft explicitly states a goal of achieving Artificial General Intelligence (AGI), which has significant ethical implications.
Ethical Guidelines for AI Research
Discussion Points:
Looking at the NENT guidelines: A critical question is whether existing ethical guidelines, such as those from the Norwegian National Research Ethics Committees (NENT), are sufficient for AI research.
Do we need separate research ethical guidelines for AI research?: Due to the unique characteristics and rapid evolution of AI, it is debated whether specific guidelines are necessary.
If we do: Why? It is argued that existing guidelines may fall short concerning AI research because of:
The unprecedented scale of data collection and processing.
The opaque nature of complex AI models (the "black box" problem).
The potential for autonomous AI agents to make decisions without human intervention.
The broad societal impact and potential for systemic bias and discrimination.
Challenges in ensuring accountability and responsibility for AI actions.
The rapid pace of technological development, which can outstrip ethical and regulatory frameworks.
Specific issues like the alignment problem, potential for superintelligence, and the philosophical implications of creating artificial consciousness or presence.