By Jennifer Evans with support from ChatGPT
Photo of Minister Champagne from Collision 2022
The emergence of generative AI systems like ChatGPT, DALL·E 2, and Midjourney is revolutionizing fields ranging from content creation to customer service. Their flexibility and wide-ranging applicability underscore AI’s vast potential. However, with this potential comes an equally broad risk profile, emphasizing the importance of responsible development, deployment, and usage. Canada’s proposal for a voluntary code of conduct is a reflection of this reality.
Scope of the Canadian Code of Conduct
The Canadian AI code of conduct covers areas from accountability and safety to transparency and human oversight. By addressing the role of firms and roles in both developing and managing AI, it attempts to provide a holistic framework. The inclusion of fairness, equity, and robustness measures acknowledge societal implications, including bias and discrimination.
Any effectiveness will depend wholly on its implementation, which is voluntary. Generative AI has potentially significant implications for health, safety, individual privacy rights, and societal structures. This makes its wide-ranging deployment both an asset and a liability.
Efficacy in Real-world Scenarios
The hypothetical “DeepTruth Technologies” and “NeuraHealth” scenarios that follow demonstrate potential shortcomings. Whether it’s the misuse of avatars or biases due to non-diverse datasets, they illustrate how even a well-intentioned framework can fall short. Ambiguities in terms, commercial pressures, and inconsistencies in application can result in unintended consequences. These scenarios also underscore the rapid pace of AI development, which can often outstrip the speed of regulatory adaptation.
Strengths and Limitations
While voluntary codes set industry standards and can foster a culture of ethical development, their non-binding nature is a massive limitation. Binding regulations, licensing requirements, mandatory audits, and legal liability offer stronger control. Even these measures can be challenging to implement effectively due to AI’s rapid, decentralized, and global development nature.
Localized vs. Global Efforts
The global nature of AI technology, combined with competitive pressures and the decentralization of AI research, complicates regulatory efforts. If efforts are localized and inconsistently applied, they might have limited global impact. For instance, strict regulations in one country could push companies to relocate their operations to regions with laxer oversight, diluting the efficacy of localized measures. Nevertheless, prominent AI hubs, like Canada, can still influence global practices, especially if their guidelines become industry benchmarks.
Canada, with its strong AI research ecosystem, can exert influence beyond its borders. Its collaborations, partnerships, and leadership in AI research can spread Canadian best practices. Moreover, if Canada’s code becomes a gold standard, demonstrating the balance between innovation and ethics, it could inspire similar global initiatives.
However, given the major roles of larger economies and tech hubs like the U.S. and China, Canada’s direct influence on global AI practices, while significant, remains proportionate to its standing in the global AI ecosystem.
Future Trajectory and Considerations
The intersection of AI’s potential and its risks necessitates a proactive approach to governance. While Canada’s and others’ voluntary codes are a step in the right direction, real-world impact hinges on consistent application, periodic revisions to keep pace with technological advancements, and perhaps most crucially, global collaborations to standardize best practices.
AI’s evolution is inevitable. The challenge lies not in impeding this progress but in guiding it responsibly. As AI systems continue to weave into the fabric of daily life, ensuring they benefit humanity without undue risks remains paramount.
There are several measures that could be more impactful than a voluntary code of conduct in influencing or impeding the speed of AI development:
1. Binding Regulations and Legislation: Governments can enact laws with specific requirements for AI development, deployment, and usage. Non-compliance would result in legal consequences, such as fines, sanctions, or business restrictions. This is the next step for this code.
2. Licensing and Certification: Just as certain professions or industries require licensing, AI systems or AI-driven businesses could require certifications that ensure they meet certain standards. Without this certification, they couldn’t operate.
3. Mandatory Third-party Audits: AI firms could be required to undergo periodic third-party audits to ensure their practices align with established guidelines or regulations. This adds an external layer of scrutiny.
4. Moratoriums: Governments or international bodies could place temporary bans on specific AI applications deemed high-risk until they’re better understood or until appropriate regulations are in place.
5. Trade Restrictions: Countries could impose trade restrictions or tariffs on AI technologies that don’t meet certain ethical or safety standards, discouraging their development or deployment.
6. Public Funding Conditions: Many AI research initiatives benefit from public funding. Governments could make such funding contingent on adherence to certain ethical guidelines or developmental milestones.
7. Liability Laws: Establishing clear liability laws where developers or companies can be held accountable for the harm caused by their AI systems can deter reckless development.
8. Transparency and Openness Requirements: Mandating that AI companies disclose algorithms, training data, or methodologies can deter malicious practices and ensure a broader scrutiny of AI development processes.
9. Ethical Review Boards: Just as medical research often requires review by ethics boards, AI projects, especially those with significant societal implications, could be mandated to undergo ethical review.
10. Education and Training Mandates: Requiring AI practitioners to undergo specific ethical training, much like medical professionals swear by the Hippocratic Oath, can instill a culture of responsible development.
11. Collaborative International Frameworks: Global cooperation can lead to harmonized standards and practices. International agreements or treaties on AI development can provide a unified approach to manage the global nature of AI advancements.
12. Public Awareness Campaigns: Governments or advocacy groups can initiate campaigns to educate the public about potential risks of unchecked AI, leading to consumer-driven demands for responsible development.
While these measures can be effective in impeding or guiding the speed of AI development, they need to be applied judiciously. Overly restrictive measures could stifle innovation, hinder economic growth, or result in a competitive disadvantage on the global stage. The challenge lies in balancing the need for safety, ethics, and responsibility with the benefits of innovation and progress.
The following scenarios focus around circumstances not covered by a voluntary code of conduct:
Scenario: “NeuraHealth” and the “Biased Health Assistant Debacle”
Background: NeuraHealth, a pioneering Canadian AI firm, unveils “MediBot”, an advanced AI system intended to provide instant medical information to users. Marketed as a “personal health assistant”, MediBot is designed to analyze a user’s symptoms and provide potential diagnoses, suggest treatments, and recommend when to see a doctor.
Act 1: Dataset Limitations:
1. NeuraHealth, adhering to the code of conduct, curates a vast dataset using medical literature and patient data. However, most of this data is derived from studies predominantly featuring middle-aged Caucasian subjects.
2. Despite the code’s emphasis on fairness and equity, NeuraHealth overlooks this dataset bias, believing their broad dataset is comprehensive.
Act 2: Unforeseen Biases:
1. As MediBot gains popularity, reports emerge about its inconsistent accuracy. People of certain ethnic backgrounds find that the AI’s diagnosis often misses conditions prevalent in their communities.
2. For instance, MediBot frequently misdiagnoses conditions in individuals of Asian descent because certain diseases manifest differently in them compared to the majority Caucasian data the AI was trained on.
Act 3: Ambiguous Accountability and Transparency:
1. Affected users demand explanations. NeuraHealth, while acknowledging the oversight, points to their adherence to the code’s guidelines on assessing and curating datasets.
2. The company claims they acted within the “reasonably foreseeable” framework, emphasizing that they couldn’t have predicted every potential bias.
3. Users also criticize NeuraHealth for lack of transparency. While the code suggests publishing “a description of the types of training data”, NeuraHealth’s descriptions were generic, not detailing the ethnic or demographic distribution of their datasets.
Act 4: Consequences and Lessons:
1. Several users, based on MediBot’s inaccurate advice, suffer health complications, leading to lawsuits against NeuraHealth.
2. The public starts questioning the reliability of AI in sensitive sectors like healthcare. Medical professionals warn against over-reliance on such tools without human oversight.
3. NeuraHealth faces significant reputational damage. They embark on a mission to diversify their datasets, but the trust once lost is hard to regain.
4. The incident serves as a case study illustrating that even with comprehensive guidelines, real-world complexities can lead to unintended biases and harms.
This scenario emphasizes the importance of recognizing biases in datasets and highlights how even well-intentioned guidelines might fall short when confronted with the nuances of real-world application.
Scenario 2: “DeepTruth Technologies” and the “Digital Avatar Crisis”
Background: DeepTruth Technologies, a Canadian AI firm, has developed an advanced AI system named “Persona” that can create hyper-realistic digital avatars of humans. These avatars can mimic real-life individuals, including their voice, facial expressions, and mannerisms. The primary intent of Persona is to help businesses create lifelike virtual customer service representatives.
Act 1: Loopholes and Ambiguity:
1. Exploiting the distinction between “general use” and “publicly available” systems, DeepTruth markets Persona as an “enterprise solution”, thereby skirting some of the more stringent guidelines for public systems.
2. The term “reasonably foreseeable adverse impacts” is interpreted loosely by DeepTruth. They focus more on technical glitches and obvious misuse but do not deeply consider the broader societal implications.
Act 2: Misuse and Commercial Pressures:
1. A third-party company, ShadowSoft, acquires Persona for “customer service” purposes. However, ShadowSoft secretly starts offering a service where anyone can create digital replicas of real people for a fee.
2. This service becomes an underground hit. Malicious actors start creating misleading videos of politicians, celebrities, and even private individuals, leading to misinformation campaigns, blackmail, and personal harassment.
Act 3: Delayed Response:
1. As the crisis unfolds, victims demand accountability. DeepTruth claims they followed the code of conduct by assessing “reasonably foreseeable” risks. They argue that ShadowSoft’s specific misuse wasn’t anticipated.
2. The code’s emphasis on governing human conduct rather than technological use means that while DeepTruth’s ethical responsibilities are questioned, there’s no direct regulation preventing the creation of such avatars.
3. The recommendation in the code for “Human Oversight and Monitoring” falls short. DeepTruth had outsourced their post-deployment monitoring to a third-party, which failed to notice ShadowSoft’s misuse until it was too late.
Act 4: Repercussions:
1. The public’s trust in AI technologies takes a hit. There’s an outcry for stricter regulations.
2. Victims of the misleading videos struggle to combat the spread of fake content. Even with debunking, the “seeing is believing” nature of video content sows doubt and mistrust.
3. DeepTruth faces lawsuits and potential bankruptcy. The code of conduct’s guidelines, while comprehensive, failed to anticipate this specific misuse scenario or prevent its fallout.
This scenario illustrates how ambiguity in guidelines, combined with commercial pressures and rapid technological advancements, can lead to unintended and harmful consequences.