Welcome to The AI Ethics Brief, a bi-weekly publication by the Montreal AI Ethics Institute. We publish every other Tuesday at 10 AM ET. Stay informed on the evolving world of AI ethics with key research, insightful reporting, and thoughtful commentary. Learn more at montrealethics.ai/about.

Follow MAIEI on Bluesky and LinkedIn. ❤️ Support Our Work.

• H&M’s AI Models: Who Really Wins When Fashion Models Go Virtual?

• UNCTAD Technology & Innovation Report 2025: AI’s Promise and Peril for Global Development

• Surveillance by Default: What Clearview AI Reveals About Power and Technology

• AI Policy Corner: Frontier AI Safety Commitments, AI Seoul Summit 2024

• Red Teaming is a Critical Thinking Exercise – Part 1 (AI Vulnerability Database)

• Mapping the Responsible AI Profession, A Field in Formation (techUK)

• A Google Gemini model now has a “dial” to adjust how much it reasons – MIT Technology Review

• Company apologizes after AI support agent invents policy that causes user uproar – Ars Technica

• Your politeness could be costly for OpenAI – TechCrunch

We came across a fascinating position paper by David Silver (AlphaGo) and Richard Sutton (a pioneer of reinforcement learning), shared via Jack Clark’s Import AI newsletter. Their argument is that the next leap toward superintelligence won’t come from training on human-curated datasets but from agents that learn through their own experiences: experimenting, reasoning, and even developing non-human modes of thought. You can read the full paper here.

Silver and Sutton describe an “Era of Experience,” where AI agents inhabit ongoing streams of interaction, grounded in real-world environments, optimizing based on outcomes they experience rather than human judgments. As they learn autonomously, they may discover radically new ways of reasoning, moving beyond human language, human logic, and possibly even human oversight.

“In the era of human data, these reasoning methods have been explicitly designed to imitate human thought processes. For example, LLMs have been prompted to emit human-like chains of thought, imitate traces of human thinking, or to reinforce steps of thinking that match human examples. The reasoning process may be fine-tuned further to produce thinking traces that match the correct answer, as determined by human experts.

However, it is highly unlikely that human language provides the optimal instance of a universal computer. More efficient mechanisms of thought surely exist, using non-human languages that may for example utilise symbolic, distributed, continuous, or differentiable computations. A self-learning system can in principle discover or improve such approaches by learning how to think from experience.”

This echoes what we explored in The AI Ethics Brief #161, where we asked whether today’s AI systems are simply optimized to obey rather than to question. If compliance-trained models already risk reinforcing the status quo, what happens when agents create their own internal languages and cognitive frameworks, ones we might not even be able to decode?

As we step into the “Era of Experience,” the question isn’t just what AI can do, it’s whether we’ll still be able to understand it.

Please share your thoughts with the MAIEI community:

Leave a comment

What Happened: Clothing brand H&M recently made the controversial decision to create AI “clones” of thirty of its fashion models to use in advertising campaigns. This decision both cuts costs and increases efficiency, essentially allowing models to be in multiple places at once using their AI avatar. This is not a novel decision. Many prominent brands, such as Levi’s and Mango, have similarly used generative AI to produce AI fashion models.

📌 MAIEI’s Take and Why It Matters:

H&M’s motivation to reduce spending on photo shoots through AI model clones is transparent, but the implications of this decision are much more far-reaching. Not only could this practice become financially exploitative for models and reduce demand for other on-set employees, such as photographers and make-up artists, but it is also another facet through which automation is eroding artwork.

Fashion images are the result of harmonious co-production between photographers, designers, stylists, lighting technicians, and models. A wealth of expertise is amalgamated for a single shot. While displacing employees from their work, the generated images strip away that work’s artistic essence. As noted in “AI Art and its Impact on Artists,” a 2023 paper co-authored by the late MAIEI founder Abhishek Gupta:

“… a work of art is a cultural product that uses the resources of a culture to embody that experience in a form that all who stand before it can see. On this view, art refers to a process that makes use of external materials or the body to make present experience in an intensified form.”

AI fashion models clearly do not fit that definition.

While brands like H&M are more clear on their aim to decrease costs, other brands like Levi’s utilize alternative justifications. Upon receiving significant backlash for their partnership with Dutch AI model company Lalaland.ai in 2023, Levi’s emphasized their use of AI “to create hyper-realistic models of every body type, age, size and skin tone.”

Regardless of stated aims, cost reduction remains the overwhelming benefit of AI-generated fashion models. Framed against Levi’s justification, diversity is no longer the goal of AI avatars but their casualty, sacrificed for corporate efficiency. Real representation and employment of underrepresented fashion models is now being replaced by technology that not only risks erasure but also encodes and amplifies existing hegemonic biases and inaccuracies.

What Happened: The UNCTAD Technology and Innovation Report 2025 examines AI’s dual role as both a catalyst for development and a potential amplifier of global inequalities. It advocates for human-centered AI deployment through targeted policies and cross-border collaboration to ensure technological benefits are widely shared. The report warns of the growing concentration of AI innovation among a few nations and corporations in the Global North, an imbalance that risks deepening existing divides, as many developing economies lack the infrastructure, data resources, and technical expertise needed for AI adoption.

📌 MAIEI’s Take and Why It Matters:

The tension between AI’s economic promise and its threat to labour markets demands urgent ethical scrutiny, especially for developing economies whose competitiveness has long relied on affordable labour. Why does this matter? Because technological disruption without thoughtful governance risks widening global inequalities rather than bridging them.

While UNCTAD’s Frontier Technologies Readiness Index highlights promising capacity in countries like India, Brazil, and China, navigating this future demands more than bare metrics; it requires a fundamentally values-driven approach to AI deployment.

Imagine AI strategies that prioritize human augmentation over replacement. These strategies address both ethical and practical imperatives by safeguarding livelihoods while boosting productivity, crafting locally relevant solutions while nurturing homegrown innovation, and guaranteeing technological self-determination rather than dependency.

This moral imperative transcends borders and must extend to global governance, where developing nations’ participation is not just diplomatic protocol but a universal necessity. Without diverse voices shaping AI’s evolution, we risk perpetuating power imbalances under the veneer of progress. The true benchmark for successful AI governance isn’t cutting-edge sophistication, it’s whether it elevates human agency and fair opportunity across our global society.

What Happened: A new investigation reveals that Clearview AI’s facial recognition technology, infamous for mass-scraping billions of images without consent, was deliberately designed to aid the surveillance of marginalized groups.

Reports from the Business & Human Rights Resource Centre and Mother Jones highlight that Clearview AI’s biometric database has been used by U.S. agencies like ICE and the FBI to monitor immigrants, protesters, and other vulnerable communities. The investigation also uncovered links between Clearview AI leadership and far-right political agendas, raising further alarm about the weaponization of AI against marginalized populations.

📌 MAIEI’s Take and Why It Matters:

Clearview AI is no stranger to controversy: in 2021, Canadian privacy regulators found the company’s collection, use, and disclosure of personal data to be illegal and in violation of Canadian privacy laws. Despite facing enforcement actions abroad, Clearview AI’s practices continued to evolve, expanding both its database and its entanglements with politically motivated surveillance efforts.

The Clearview AI case exemplifies the broader global challenge: the lack of robust, enforceable regulations to protect individuals from AI-powered mass surveillance. While the EU AI Act has banned untargeted facial image scraping, gaps remain worldwide.

Clearview AI’s trajectory shows what happens when surveillance technologies are unleashed without ethical constraints: those already marginalized bear the brunt. When AI systems are trained on billions of scraped images and deployed by agencies tasked with immigration enforcement or protest monitoring, the result isn’t safety, it’s the amplification of systemic injustice.

Mother Jones uncovered the ties between Clearview AI’s leadership and far-right political interests, further illustrating that technological “neutrality” is a myth. Who builds and controls AI matters, and without strict oversight, these tools serve power, not people. This case reinforces why facial recognition should be considered a high-risk application globally. Consent, proportionality, and accountability are not optional; they are essential.

Without clear global standards, we risk entrenching surveillance structures that erode civil liberties under the false promise of technological progress.

Did we miss anything? Let us know in the comments below.

Leave a comment

The Insights & Perspectives summarized below explore how Responsible AI (RAI) development demands more than technical excellence; it requires critical thinking, institutional accountability, and professional stewardship.

From the evolving international commitments on Frontier AI safety, to the redefinition of red teaming as a systemic governance exercise, to the urgent call for formalizing the RAI profession, a common thread emerges: frameworks alone are not enough. Meaningful progress hinges on how risks are identified, how critical assumptions are challenged, and how institutions invest in the people tasked with operationalizing AI ethics.

Across policy, practice, and professionalization, these pieces show that the future of AI governance will not be decided by intentions alone. It will be shaped by how rigorously we build structures of responsibility, and by how early we recognize that AI ethics, like security, must be embedded, enforced, and continually reexamined.

AI Policy Corner: Frontier AI Safety Commitments, AI Seoul Summit 2024

By Alexander Wilhelm. This article is part of our AI Policy Corner series, a collaboration between the Montreal AI Ethics Institute (MAIEI) and the Governance and Responsible AI Lab (GRAIL) at Purdue University. The series provides concise insights into critical AI policy developments from the local to international levels, helping our readers stay informed about the evolving landscape of AI governance.

Recent international efforts, including the AI Safety Summit 2023 held in the UK and the AI Seoul Summit 2024, have pushed forward discussions on the “safe” development of AI. The Seoul Summit produced the Frontier AI Safety Commitments, endorsed by 20 organizations including Anthropic, Microsoft, NVIDIA, and OpenAI, requiring the publication of safety frameworks focused on severe risks. However, the Paris AI Action Summit in 2025 (see The AI Ethics Brief #158 for more) shifted its emphasis toward AI’s benefits rather than its risks, casting uncertainty over the future of these commitments.

To dive deeper, read the full article here.

Red Teaming is a Critical Thinking Exercise – Part 1 (AI Vulnerability Database)

The AVID (AI Vulnerability Database) blog series “Red Teaming is a Critical Thinking Exercise” opens with a tribute to Abhishek Gupta, whose early encouragement helped shape the project. In the first installment, authors Brian Pendleton and Subho Majumdar contend that red teaming should not simply validate AI security measures but function as a broader critical thinking exercise, challenging assumptions, surfacing vulnerabilities, and strengthening governance practices across AI systems. Rather than focusing narrowly on technical flaws, effective red teaming must examine system-wide impacts, including data integrity, decision-making transparency, and organizational resilience. As AI adoption accelerates, the piece highlights the urgent need for enterprises to apply rigorous, systemic analysis throughout the AI lifecycle to drive safer, more accountable innovation.

To dive deeper, read the full article here.

Mapping the Responsible AI Profession, A Field in Formation (techUK)

A new report from techUK, Mapping the Responsible AI Profession: A Field in Formation, highlights the growing need to formalize the role of Responsible AI (RAI) practitioners across industries. As AI governance shifts from a theoretical concern to an urgent operational priority, the report emphasizes that a lack of clear professional pathways risks eroding stakeholder trust and slowing innovation. Drawing from practitioner experiences, techUK identifies critical gaps in career development, certification, and leadership integration, and recommends targeted actions for organizations, professional bodies, and policymakers to strengthen the emerging RAI profession and its role in ensuring accountable AI adoption.

To dive deeper, read the paper summary here.

A Google Gemini model now has a “dial” to adjust how much it reasons – MIT Technology Review

1. Summary: Google DeepMind has introduced a new feature in its Gemini Flash 2.5 AI model: a “reasoning” dial that allows developers to control how much the model “thinks” through a problem. The aim is to reduce cost and energy use, acknowledging that advanced reasoning models often overthink simple prompts. Reasoning models have become a new frontier in AI development, promising better performance on complex tasks like coding or research synthesis. However, they also come with higher computational costs and a risk of getting stuck in loops. DeepMind’s update allows developers to strike a balance between performance and efficiency by fine-tuning the level of reasoning. While the dial is currently available only to developers, it reflects a broader shift in the AI field toward prioritizing smarter, not just bigger, models.

2. Why It Matters: The ability to adjust reasoning marks a meaningful step forward in how AI is deployed. As reasoning models become the new gold standard, the industry seems to be grappling with how to harness their capabilities without draining resources or compromising performance. DeepMind’s dial is a pragmatic response to the messy reality that more “thinking” isn’t always better; sometimes, it’s just more expensive and wasteful. It also raises bigger questions about how we define intelligence in machines, and whether “reasoning” is the right goal or just another buzzword. Ultimately, this can be viewed as a reminder that innovation in AI isn’t just about what models can do, but how intentionally we choose to use them.

To dive deeper, read the full article here.

Company apologizes after AI support agent invents policy that causes user uproar – Ars Technica

1. Summary: Users of the Cursor code editor discovered that switching devices would mysteriously log them out. When they contacted “Sam” for support, they were given a seemingly plausible explanation citing security protocols. However, it turned out that “Sam” was a bot, and the policy had been fabricated entirely.

2. Why It Matters: This case highlights two fundamental realities of AI today: its persuasiveness and fallibility. LLMs can present false information convincingly, making it harder for users to distinguish truth from falsehood and whether they are interacting with a human at all. Recent advancements in AI have pushed us far beyond the original Turing Test, raising deeper questions around trust, verification, and how we define meaningful human-AI interactions.

To dive deeper, read the full article here.

Your politeness could be costly for OpenAI – TechCrunch

1. Summary: Our instinctive courtesy in AI conversations, i.e. saying “please” and “thank you” to AI systems, has unexpectedly contributed to millions of dollars in additional computing costs for OpenAI. According to CEO Sam Altman, while each polite prompt adds only a small computational burden, the cumulative effect across millions of interactions has real financial and environmental impacts. This finding reflects broader questions about the hidden resource demands of everyday AI use, as well as the subtle ways user behaviour shapes the future of human-AI interaction.

2. Why It Matters: What seems like a harmless social nicety reveals deeper tensions at the heart of human-AI interaction. While individual instances of politeness add minimal computational load, at a global scale, they contribute to the significant environmental footprint of AI systems. More critically, anthropomorphizing AI risks creating deeper misconceptions about what these systems actually are: prediction engines, not sentient companions. LLMs can mimic empathy but lack true understanding, ethical judgment, or autonomy. Treating them as “human-like” obscures the real accountability structures behind AI, hides the biases baked into their training, and misleads users into overestimating their capabilities. As AI becomes more embedded in everyday life, both the environmental and conceptual costs of our habits deserve closer scrutiny.

To dive deeper, read the full article here.

MAIEI to Participate in AI Governance Panel at Point Zero Forum 2025

The Montreal AI Ethics Institute has been invited to participate in the Point Zero Forum 2025 (May 5-7, 2025) in Zurich, Switzerland, to contribute to discussions on balancing innovation, regulation, and ethics in AI governance.

Renjie Butalid, MAIEI Co-founder and Director, will join a panel of global leaders to explore how risk-based approaches like the EU AI Act can shape the future of trustworthy, human-centric AI deployment. The session will examine how regulatory frameworks can safeguard societal interests without stifling innovation, and how ethical AI can serve as a strategic advantage in a competitive global landscape.

Learn more about the Point Zero Forum here.

Responsibly Navigating the Enterprise AI Landscape (Partnership on AI)

Partnership on AI’s latest report, Responsibly Navigating the Enterprise AI Landscape, outlines key challenges and opportunities for organizations adopting AI responsibly. As enterprise AI use accelerates, the report highlights a growing gap: while much guidance focuses on AI development, far less addresses responsible use. Through two workshops with businesses, model providers, civil society groups, and academic institutions, the report identifies critical challenges in responsible AI readiness, evaluation and monitoring, and building trust across the AI value chain. It calls for future work on knowledge alignment, governance structures, implementation guidelines, and impact measurement to ensure that the real-world deployment of AI advances human-centered and ethical outcomes.

To dive deeper, read the full report here.

Help us keep The AI Ethics Brief free and accessible for everyone by becoming a paid subscriber on Substack or making a donation at montrealethics.ai/donate. Your support sustains our mission of democratizing AI ethics literacy and honours Abhishek Gupta’s legacy.

For corporate partnerships or larger contributions, please contact us at support@montrealethics.ai

Have an article, research paper, or news item we should feature? Leave us a comment below — we’d love to hear from you!