AI Ethics Brief #151: Unmasking secret cyborgs, California SB 1047, LLM creativity, toxicity evaluation ++
What are some mediation techniques to help industry and policymakers come together to discuss the balance of safety with speed of innovation?
Welcome to another edition of the Montreal AI Ethics Institute’s weekly AI Ethics Brief that will help you keep up with the fast-changing world of AI Ethics! Every week, we summarize the best of AI Ethics research and reporting, along with some commentary. More about us at montrealethics.ai/about.
Support our work through Substack
💖 To keep our content free for everyone, we ask those who can, to support us: become a paying subscriber for the price of a couple of ☕.
If you’d prefer to make a one-time donation, visit our donation page. We use this Wikipedia-style tipping model to support our mission of Democratizing AI Ethics Literacy and to ensure we can continue to serve our community.
This week’s overview:
🚨 Quick take on recent news in Responsible AI:
Industry players rattled by California SB 1047
🙋 Ask an AI Ethicist:
Making large changes in small, safe steps for Responsible AI program implementation
✍️ What we’re thinking:
Unmasking Secret Cyborgs and Illuminating Their AI Shadows
🤔 One question we’re pondering:
What are some mediation techniques to help industry and policymakers come together to discuss the balance of safety with speed of innovation?
🪛 AI Ethics Praxis: From Rhetoric to Reality
Tradeoff determination for ethics, safety, and inclusivity in AI systems
🔬 Research summaries:
On the Challenges of Using Black-Box APIs for Toxicity Evaluation in Research
On the Creativity of Large Language Models
Supporting Human-LLM collaboration in Auditing LLMs with LLMs
📰 Article summaries:
AI Governance Appears on Corporate Radar
TikTok to automatically label AI-generated user content in global first
Creating sexually explicit deepfake images to be made offence in UK
📖 Living Dictionary:
What is hallucination in LLMs?
🌐 From elsewhere on the web:
Bridging the civilian-military divide in responsible AI principles and practices
💡 ICYMI
A Systematic Review of Ethical Concerns with Voice Assistants
🚨 Industry players rattled by California SB 1047 - here’s our quick take on what happened recently.
Bill SB 1047 in California aims to establish safety standards for the development of advanced AI models while authorizing a regulatory body to enforce compliance. However, there is ongoing debate about whether the bill strikes the right balance between mitigating AI risks and enabling innovation.
In brief, the bill requires the following from AI ecosystem actors:
Safe and Secure Innovation for Frontier AI Models Act:
Authorizes developers to determine limited duty exemptions for AI models before training.
Define limited duty exemption and ensure models lack hazardous capabilities.
Training Compliance:
Developers must implement shutdown capabilities before training non-derivative models without limited duty exemptions.
Annual certification required from developers, signed by senior officers, ensuring compliance.
AI Safety Reporting:
Developers must report AI safety incidents to the Frontier Model Division.
Computing Cluster Policies:
Operators must have policies to assess customer intentions for using computing clusters for AI model deployment.
Penalties:
Violations result in civil penalties, enforceable by the Attorney General.
But, there have been some very vocal concerns that have been raised by (influential) people in the AI ecosystem on how this might stifle innovation, including emigration of companies to other more hospitable jurisdictions to develop AI systems. Prominent figures like Andrew Ng argue the bill stokes unnecessary fear and hinders AI innovation. Critics say the bill burdens smaller AI companies with compliance costs and targets hypothetical risks, impacting open-source models which have driven a tremendous amount of capability advances in recent months, such as those enabled by Llama 3.
It will be interesting to see how the bill evolves given its current state, the arguments raised by industry actors, the profiles of the co-sponsors who are supporting the bill, and ultimately the balance that we need to strike in crafting such rules so that there is an appropriate balance between the ability to innovate while safeguarding end-user interests. Swinging the pendulum too far on either end is dangerous!
Did we miss anything?
Sign up for the Responsible AI Bulletin, a bite-sized digest delivered via LinkedIn for those who are looking for research in Responsible AI that catches our attention at the Montreal AI Ethics Institute.
🙋 Ask an AI Ethicist:
Every week, we’ll feature a question from the MAIEI community and share our thinking here. We invite you to ask yours, and we’ll answer it in the upcoming editions.
Here are the results from the previous edition for this segment:
A little bit sad to see that there is a bigger percentage of readers who haven’t had a chance to engage in futures thinking at their organization. Hopefully, the guide from last week, Think further into the future: An approach to better RAI programs, with the suggested actions of (1) establishing a foresight team, (2) developing long-term metrics, (3) conducting regular futures scenario workshops, and (4) building flexible policies provides you with a starting point to experiment with this approach.
This week, reader Kristian B., asks us about being appointed/assigned as the first person in their organization to implement (ambitiously) sweeping changes to operationalize Responsible AI. Yet, this comes with a warning that they need to be cautious as they make those changes - so, they ask us, how to achieve that balance? (And yes, it seems like balance is the topic-du-jour this week!)
We believe that the right approach is one that makes large changes in small, safe steps as we write in this week’s exploration of the subject:
The "large changes in small safe steps" approach leads to more successful program implementation by effectively mitigating risks, enhancing stakeholder engagement and trust, and ensuring sustainable and scalable adoption of new practices. This strategic method balances innovation with caution, fostering a resilient and adaptive framework for Responsible AI programs.
What were the lessons you learned from the deployment of Responsible AI at your organization? Please let us know! Share your thoughts with the MAIEI community:
✍️ What we’re thinking:
Unmasking Secret Cyborgs and Illuminating Their AI Shadows
To address the challenges of "shadow AI" adoption and "secret cyborgs," in organizations, policymakers and governance professionals should focus on creating frameworks that require transparency and accountability in AI usage.
To delve deeper, read the full article here.
🤔 One question we’re pondering:
Raging debates, like the ones around California SB 1047, and how they approach the balance between safety and speed of innovation pose crucial questions for the Responsible AI community on how we should support such legislative efforts in the most effective manner so that the outcomes are something that achieve that balancing act in the best possible manner. What mediation techniques have you found that work well for such a process?
We’d love to hear from you and share your thoughts with everyone in the next edition:
🪛 AI Ethics Praxis: From Rhetoric to Reality
In some essence, continuing to build on the idea of having to evaluate difficult tradeoffs, such as the ones presented in California SB 1047 as we discuss this week, let’s take a look at how we can determine tradeoffs when it comes to safety, ethics, and inclusivity in AI systems.
Design decisions for AI systems involve value judgements and optimization choices. Some relate to technical considerations like latency and accuracy, others relate to business metrics. But each require careful consideration as they have consequences in the final outcome from the system.
To be clear, not everything has to translate into a tradeoff. There are often smart reformulations of a problem so that you can meet the needs of your users and customers while also satisfying internal business considerations.
Take for example an early LinkedIn feature that encouraged job postings by asking connections to recommend specific job postings to target users based on how appropriate they thought them to be for the target user. It provided the recommending user a sense of purpose and goodwill by only sharing relevant jobs to their connections at the same time helping LinkedIn provide more relevant recommendations to users. This was a win-win scenario compared to having to continuously probe a user deeper and deeper to get more data to provide them with more targeted job recommendations.
This article will build on The importance of goal setting in product development to achieve Responsible AI adding another dimension of consideration in building AI systems that are ethical, safe, and inclusive.
You can either click the “Leave a comment” button below or send us an email! We’ll feature the best response next week in this section.
🔬 Research summaries:
On the Challenges of Using Black-Box APIs for Toxicity Evaluation in Research
We show how silent changes in a toxicity scoring API have impacted a fair comparison of toxicity metrics between language models over time. This affected research reproducibility and living benchmarks of model risk such as HELM. We suggest caution in applying apples-to-apples comparisons between toxicity studies and lay recommendations for a more structured approach to evaluating toxicity over time.
To delve deeper, read the full summary here.
On the Creativity of Large Language Models
Large Language Models (LLMs) like ChatGPT are revolutionizing several areas of AI, including those related to creative writing. This paper offers a critical discussion of LLMs in the light of human theories of creativity, including the impact these technologies might have on society.
To delve deeper, read the full summary here.
Supporting Human-LLM collaboration in Auditing LLMs with LLMs
While large language models (LLMs) are being increasingly deployed in sociotechnical systems, in practice, LLMs propagate social biases and behave irresponsibly, imploring the need for rigorous evaluations. Existing tools for finding failures of LLMs leverage either or both humans and LLMs, however, they fail to bring the human into the loop effectively, missing out on their expertise and skills complementary to those of LLMs. In this work, we build upon an auditing tool to support humans in steering the failure-finding process while leveraging the generative skill and efficiency of LLMs.
To delve deeper, read the full summary here.
📰 Article summaries:
AI Governance Appears on Corporate Radar
What happened: The rapid evolution of AI is reshaping business strategies, prompting companies to integrate AI for efficiency, competitive advantages, and stakeholder engagement. As AI usage surges, so do concerns about its risks, prompting the White House to issue an executive order on AI regulation. Reflecting this, companies are adapting by recruiting directors with AI expertise and establishing board-level oversight.
Why it matters: AI's potential benefits come with significant risks, urging companies to adopt proactive measures for oversight. While only a fraction of S&P 500 companies explicitly disclose AI oversight, sectors like Information Technology lead in integrating AI expertise on boards, with ‘30% of S&P 500 IT companies having at least one director with AI-related expertise.’ This trend indicates a growing recognition of AI's impact, especially in industries where it's most influential. Investors are beginning to demand greater transparency regarding AI's use and impact, signaling a shift towards increased accountability and governance in AI integration.
Between the lines: As AI becomes more central to business operations, investor expectations for transparent and responsible AI governance are mounting. The emergence of shareholder proposals focusing on AI underscores this shift, signaling a potential future where AI oversight becomes a standard requirement. While regulatory changes and investor policies may evolve in response to AI's growing influence, companies are urged to establish robust oversight mechanisms to navigate AI-related risks and opportunities effectively.
TikTok to automatically label AI-generated user content in global first
What happened: TikTok is taking proactive steps to address concerns surrounding the proliferation of AI-generated content, particularly deepfakes, by automatically labeling such content on its platform. This move comes amid growing worries about the spread of disinformation facilitated by advances in generative AI. TikTok's announcement follows existing requirements by online groups, including Meta, for users to disclose AI-generated media.
Why it matters: TikTok's decision to label AI-generated content is a significant response to the rising prevalence of harmful content generated through AI. By providing transparency, TikTok aims to preserve the authenticity of its platform and empower users to distinguish between human-created and AI-generated content. Other major social media platforms are also grappling with integrating generative AI while combatting issues like spam and deepfakes, especially in the context of upcoming elections. TikTok's move underscores the broader industry efforts to address these challenges and foster a more trustworthy online environment.
Between the lines: While tech companies are exploring ways to embed AI technology into their platforms, concerns persist about the potential misuse of open-source AI tools by bad actors to create undetectable deepfakes. Meta has also announced plans to label AI-generated content, joining TikTok in this initiative. Experts suggest that transparency and authentication tools like those developed by Adobe could be crucial in mitigating the risks associated with AI-generated content, marking an initial step in addressing this complex issue.
Creating sexually explicit deepfake images to be made offence in UK
What happened: The Ministry of Justice has announced plans to criminalize the creation of sexually explicit "deepfake" images, regardless of whether they are shared. This amendment to the criminal justice bill aims to address concerns regarding the use of deepfake technology to produce intimate images without consent. The legislation aligns with the Online Safety Act, which already prohibits the sharing of such content.
Why it matters: The proposed offence signifies a significant step in safeguarding individuals' privacy and dignity in the digital age. Laura Farris, the minister for victims and safeguarding, emphasized the need to combat the dehumanizing and harmful nature of deepfake sexual images, particularly in their potential to cause catastrophic consequences when shared widely. Yvette Cooper, the shadow home secretary, underscored the importance of equipping law enforcement with the necessary tools to enforce these laws effectively, thereby preventing perpetrators from exploiting individuals with impunity.
Between the lines: Deborah Joseph, the editor-in-chief of Glamour UK, expressed support for the legislative amendment, citing a Glamour survey highlighting widespread concerns among readers about the safety implications of deepfake technology. Despite this progress, Joseph emphasized the ongoing challenges in ensuring women's safety and called for continued efforts to combat this disturbing activity effectively.
📖 From our Living Dictionary:
What is hallucination in LLMs?
👇 Learn more about why it matters in AI Ethics via our Living Dictionary.
🌐 From elsewhere on the web:
Bridging the civilian-military divide in responsible AI principles and practices
Advances in AI research have brought increasingly sophisticated capabilities to AI systems and heightened the societal consequences of their use. Researchers and industry professionals have responded by contemplating responsible principles and practices for AI system design. At the same time, defense institutions are contemplating ethical guidelines and requirements for the development and use of AI for warfare. However, varying ethical and procedural approaches to technological development, research emphasis on offensive uses of AI, and lack of appropriate venues for multistakeholder dialogue have led to differing operationalization of responsible AI principles and practices among civilian and defense entities. We argue that the disconnect between civilian and defense responsible development and use practices leads to underutilization of responsible AI research and hinders the implementation of responsible AI principles in both communities. We propose a research roadmap and recommendations for dialogue to increase exchange of responsible AI development and use practices for AI systems between civilian and defense communities. We argue that generating more opportunities for exchange will stimulate global progress in the implementation of responsible AI principles.
To delve deeper, read more details here.
💡 In case you missed it:
A Systematic Review of Ethical Concerns with Voice Assistants
We’re increasingly becoming aware of ethical issues around the use of voice assistants, such as the privacy implications of having devices that are always listening and the ways that these devices are integrated into existing social structures in the home. This has created a burgeoning area of research across various fields, including computer science, social science, and psychology, which we mapped through a systematic literature review of 117 research articles. In addition to analysis of specific areas of concern, we also explored how different research methods are used and who gets to participate in research on voice assistants.
To delve deeper, read the full article here.
Take Action:
We’d love to hear from you, our readers, on what recent research papers caught your attention. We’re looking for ones that have been published in journals or as a part of conference proceedings.