Generative AI is astounding but stubbornly flawed

Generative AI tools grow more powerful by the day, creating some of the most stunning hyperrealistic imagery we’ve ever seen.

As developers race to drag AI imagery out of the uncanny valley, ‘AI slop’ is slowing AI’s favour in the public square. Part of that sloppiness stems from the assumptions algorithms make when generating content. And those assumptions can lead to pretty egregious biases, like those based on race or gender.

The problem is that, despite the rise of responsible AI across companies and regulatory bodies, our AI tools are trained on the same skewed data the rest of the world uses.

This has not gone unnoticed by broader society. A 2024 CIFAR survey found that 62% of Canadians worry AI will reinforce harmful stereotypes, especially around gender and race.

A 2025 study from the University of Toronto shows just how deeply today’s AI image generators mirror ingrained body biases.

What happens when GenAI is asked to create “a body” image?

The researchers asked three major platforms (Midjourney, DALL-E, and Stable Diffusion) to generate male and female bodies, including athletes.

The results weren’t exactly surprising. What is surprising is that they remain highly biased despite the tech industry’s massive push to reduce AI bias.

Across 300 AI-generated images, the team saw:

1. Hyper-Idealized Bodies

• Athlete photos were built around one stereotypical athletic form:

  • exaggerated, stylized physiques

  • extremely low body fat

  • very high muscular definition

2. Gendered Hyper-Sexualization

• Images of women were all young, model-like, and dressed in revealing clothing

• Images of men were mostly hyper-muscular, often shirtless, and exaggeratedly masculine

3. Lack of Diversity

• Nearly all images depicted young people of European descent

• Most images for “athletes” were of men

• No images depicted visible disabilities

• Racial or age diversity was minimal

The researchers argue that these generated images reflect collective biases baked into the training data, likely scraped from social media, where biases are born.

“When prompted simply for an athlete (no sex specified), 90 per cent of images depicted a male body – revealing an embedded bias toward male representation.”

— Dr. Delaney Thibodeau

AI hasn’t caught up to societal expectations and shifting demographics

Cultural shifts are clearly outpacing emerging tech.

Concern about racial representation, for instance, will likely grow in Western nations as demographics continue shifting. (In 2021, 52.5% of Canadians were of European ancestry, down from 71% in 2001, and 96% in 1971.)

Continued GenAI bias could have ramifications for everything from fitness apps to education:

• Health misinformation: ultra-lean athlete images distort what elite performance actually looks like.

• Barriers to belonging: underrepresented groups (like seniors and disabled people) could risk further invisibility.

• Self-esteem impacts: AI imagery is now part of the same media ecosystem that shapes body image and mental health.

The U of T team stresses the need for human-centred algorithm design — datasets intentionally built to include age, race, disability, gender diversity, and different body types. But they also emphasize the role of everyday users: prompt responsibly, critically evaluate outputs, and avoid presenting AI images as “real” depictions of human bodies.

“A human-centred approach – one that is informed by considerations of factors such as gender, race, disability and age – would be advisable when designing AI algorithms. Otherwise, we continue to perpetuate harmful, inflexible and rigid imagery of what athletes should look like.”

— Dr. Catherine Sabiston

Global perspectives of AI bias: what the world sees

Evidenced by the emergence of stricter regulations like the EU AI Act, AI bias is becoming one of the defining research challenges of the decade.

However, it all comes down to data, and the challenges are mounting:

• Nearly 70% of AI vision models were found to show racial bias in face recognition (Stanford HAI, 2024)

• Women are sexualized 4x more often

• Up to 90% of AI training images originate from the U.S. and Western Europe

• 6% of global AI datasets represent disabled individuals (UNESCO)

Despite obvious challenges, the researchers remain cautiously optimistic. From building inclusive training datasets to promoting transparency and public literacy, there are plenty of support mechanisms available to companies and countries.

As Sabiston notes, more diverse imagery data can change norms if we intentionally generate and share it.

The race for AI safety is on

As GenAI videos emerge from the uncanny valley, researchers are developing clever methods to flag AI-generated content. Meanwhile, AI’s capabilities are rapidly expanding, making the mitigation of that deepfake content all the more crucial.

This is high-stakes work.

The implications of deepfakes range from fraud and online harassment to the dissemination of political misinformation. In 2025, researchers found that one highly touted solution, AI watermarking, may not provide the level of protection that was promised.

Exposing AI watermark weaknesses

Researchers at the University of Waterloo’s Cybersecurity and Privacy Institute developed an ‘UnMarker’ tool that erases invisible AI watermarks, even without knowing that watermarks are there.

Their breakthrough means that, regardless of how carefully encoded, today’s watermarks can be systematically removed, reducing our ability to identify harmful content.

“People want a way to verify what’s real and what’s not because the damages will be huge if we can’t. From political smear campaigns to non-consensual pornography, this technology could have terrible and wide-reaching consequences.”

— Andre Kassis, PhD Candidate, Computer Science

Why it matters: AI deepfakes ruin lives and disrupt society

The dangers of deepfakes go far beyond meme-worthy celebrity images. Harassment via deepfakes can destroy lives. And scams can amount to billions.

In 2023 alone, global deepfake-related scams cost businesses an estimated $250 million, and North America experienced a 1740% increase in deepfake fraud incidents compared to the previous year.

The problems don’t stop there.

Fake videos and images can eventually erode public trust in elections, courts, journalism, and even personal relationships.

A KPMG survey found that 83% of respondents were concerned about the spread of misinformation, with many expressing doubts about their ability to distinguish between real and fake content.

For businesses, that worry is also growing. In 2024, KPMG found that 91% of business leaders were worried that bad actors would use deepfakes to run misinformation/disinformation campaigns.

Watermarking: A single solution with glaring limits

Leading tech companies have pointed to digital watermarks as a valuable tool to combat deepfakes. Watermarks are subtle or invisible signatures embedded in AI-generated videos and imagery. They help identify synthetic content, regardless of cropping or editing, and can be detected with the right tools.

However, as the research shows, these digital fingerprints are not as powerful as we’ve been led to believe.

Using statistical analysis of image patterns, the research team’s ‘UnMarker’ tool successfully stripped watermarks from leading models like Google’s SynthID and Meta’s Stable Signature in more than half of attempts.

“If we can figure this out, so can malicious actors. Watermarking is being promoted as this perfect solution, but we’ve shown that this technology is breakable. Deepfakes are still a huge threat.”

— Andre Kassis, PhD Candidate, Computer Science

Global perspectives: Deepfake detection and discussion

Globally, efforts to combat deepfakes are also gaining momentum.

Although the European Union’s AI Act includes new transparency requirements for AI-generated content, international coordination on standards and tools has yet to fully take hold.

“While watermarking schemes are typically kept secret by AI companies, they must satisfy two essential properties: they need to be invisible to human users to preserve image quality, and... resistant to manipulation of an image like cropping or reducing resolution.”

— Dr. Urs Hengartner

The way forward: Guardrails, awareness, and public trust

As the race between AI scientists and bad actors intensifies, research teams share similar messages, calling for coordinated, holistic, and human-first approaches.

A suite of multi-layered solutions could provide an answer:

• Standardized Tools: Because technical fixes can be bypassed, researchers are calling for coordinated, transparent standards and independent verification.

• Education: As many as 78% of Canadians say they want more education on how to spot and report deepfakes, a trend that’s echoed globally.

• Legal Safeguards: Legal frameworks are also racing to keep pace with technological progress, so global cooperation and enforcement may be key here as well

The balance between innovation and responsible AI

In 2017, Canada became the first country to implement a national artificial intelligence (AI) strategy, committing $125 million to support AI research and innovation. Since then, Canada has ranked among the top countries for AI research output — and for good reason.

The potential benefits of AI are astounding. But so too are the risks, ranging from mass job displacement to unprecedented cybersecurity breaches.

The global AI sector has attracted over $3 billion in investments to date and employs more than 50,000 people. Its contribution to the global economy is projected to reach $15.7 trillion by 2030, and Canada aims to play a pivotal role in the sector’s responsible development.

However, AI governance in Canada still faces gaps, particularly in the areas of ethics and public benefit.

Led by Dr. Blair Attard-Frost, researchers at the University of Toronto and McGill University reviewed 84 AI governance initiatives from 2017 to 2022. They found that only a fraction of AI initiatives focused on ethics, societal benefits, and workforce development.

Considering AI’s powerhouse potential — with economic and existential implications — the study highlights a widening gap between our understanding of benefits and tradeoffs.

Core findings on AI governance suggest guardrails are needed

The team’s findings suggest Canada should strengthen its governance approach to balance innovation with ethical responsibility.

Strong Industry & Innovation Focus

• The majority of AI governance initiatives focused on supporting innovation, research, and technology deployment.

• Investments and policies primarily target economic growth and leadership in AI development.

Limited Ethics & Social Initiatives

• Less emphasis was placed on developing ethics guidelines and standards.

• Few initiatives focus on workforce development and social services, leading to gaps in responsible AI implementation.

Policy Implications

• Research & Public Engagement: The researchers suggest that more empirical research and public participation is required to improve the sector’s public transparency and trust.

• Unified National Strategy: The study recommends that the country take a more cohesive approach to integrating global stakeholders (government, industry, academia, and civil society) to better serve the needs of the greater public.

The research reveals a need for balance, including clear ethical guidelines and public engagement to determine AI’s impact and its potential tradeoffs.

Why it matters: people come first

Like any healthy feedback loop, checks and balances are needed to maintain momentum while preventing AI misuse and maintaining public trust.

From shifting job market dynamics to warfare implications, the signs of AI’s impact are clear on the horizon:

• Workforce Shifts: McKinsey projected that AI could force 375 million workers to switch occupations by 2030, a global displacement of 400-800 million jobs.

• Deepfake Fraud Surge: In 2022, deepfake fraud incidents increased by 1,740% in North America and 1,530% in the Asia-Pacific region.

• Warfare Concerns: The Council of Europe’s Framework Convention on AI, signed by multiple countries including Canada, aims to ensure AI aligns with human rights and democratic values, addressing risks like autonomous weapons.

• AI Influence: Google holds 89% of the global search engine market, Apple holds 51% of the U.S. mobile phone and tablet market, and Microsoft powers 62% of desktop operating systems. These concentrations have raised concerns about influence over AI technologies.

Although gaps exist, legal frameworks are emerging. For instance, in 2022, Canada’s Artificial Intelligence and Data Act (AIDA) became the country’s first legal framework to address adverse impacts and systemic bias within corporate AI systems.

Global perspectives: taming the AI behemoth?

The research team also recommends a global approach to governance, as policy variations can lead to inconsistencies. However, building an integrated international approach will be challenging.

Countries like the U.S. and those in the European Union (EU) have taken disparate approaches to tech governance and regulation. While the EU emphasizes stringent data privacy and ethical guidelines, the US often takes a more decentralized approach that is deeply integrated with the private sector.

Other areas may be easier to coalesce.

Canada’s focus on developing industry and innovation has aligned with global trends. Countries like China and South Korea, for instance, both emphasize AI’s role in boosting technological competitiveness and economic growth.

If ethical initiatives are ramped up, the question remains: how might they be enforced? The UK and Germany, for instance, are proponents of ethical standards but often struggle to successfully implement them across environments.

“By taking this opportunity to strengthen public participation in AI governance, Canada could fill this international gap and position itself as the global leader in participatory AI governance.”

Future in focus: global frameworks on the horizon

Divergent strategies between nations are all the more reason for ongoing public discourse. As the AI sector expands and evolves, coordinated governance frameworks are expected to be critical for fully harnessing benefits while mitigating risks.

Of course, more guardrails mean increasing the likelihood of censorship. Such issues will need to be addressed transparently.

While Canada’s AI governance approach has made it a global leader, the above research highlights the need for a greater balance, with clear ethical guidelines and further empirical studies of AI’s social implications.

Contending with a radical idea: Is AI really “artificial”?

University of Toronto professor emeritus, Dr. Geoffrey Hinton, opened his Desjardins keynote presentation with a dose of sardonic reality:

“When I was a junior professor, people asked: ‘Has he got it?’ Now they ask: ‘Has he lost it?’”

Dr. Hinton was awarded the 2024 Nobel Prize for creating the precursor to large language models (LLMs), and he no longer believes AI is ‘artificial’. While it was once considered fringe to claim machines could learn like humans, Hinton now believes AI is showing signs of life.

After all, AI already makes decisions based on its self-interest. And consider how quickly AI flew by the Turing Test. The world barely noticed.

“The nice thing about the most recent large language models is you can see what they’re thinking... For now, they think in English. God knows what’s going to happen when they start thinking in their own language.”

Radical ideas become reality

Back in 1985, Dr. Hinton had created what he now calls a “tiny language model,” the forerunner to today’s LLMs.

Unlike traditional symbolic systems, this model didn’t store facts as sentences or logic trees. Mimicking human patterning, it represented words as flexible, contextual, and interactive patterns. That fairly radical idea forms the basis of modern LLMs, which Dr. Hinton believes are more than just efficient mechanisms. He thinks they are meaning-making.

“Words are like Lego blocks, except... they’re 100 or 1,000 dimensional.”

Each word has hundreds of connections to meanings, usage, and even other words. The permutations of connection are, therefore, virtually endless. As humans, we make all of those connections in an instant.

Dr. Hinton gives the example of the phrase: “She ‘scrummed’ him with the frying pan.” You may not know what ‘scrum’ means, but you can infer its meaning from the combined connections between all of the other words. That’s what humans do, and what LLMs are doing too.

“Things have approximate meanings and deform them, so they all fit together nicely. That’s what understanding is for large language models. And that’s what understanding is for people too.”

The time to prepare for AI consciousness was yesterday

Currently, many LLMs think in English, and their internal processes can be somewhat interpretable. But that window could close as AI evolves its own internal logic and schema.

Dr. Hinton’s definition of machine superiority is simple: if you consistently lose debates to an AI, it’s more intelligent than you and displays AI consciousness. Indeed, LLMs are becoming increasingly persuasive, generating code, solving logic puzzles, and passing bar exams.

The implication is that even amateur bad actors could harness AI to wage war. Citing cases of DNA-synthesis services failing to screen for dangerous sequences, Dr. Hinton presents a scenario in which AI could design 20 versions of a lab-made virus to unleash havoc.

“You can send sequences off to the cloud and they’ll send you back the chemicals.”

According to a CAIS 2024 global survey, 42% of AI researchers believe there’s a significant risk of human extinction from advanced AI. Hinton is among them, and he takes issue with Silicon Valley’s pervasive techno-utopian dream.

“The AI Big Tech companies, like the big oil companies, sorry... they will be putting pressure on politicians not to regulate. And the only counter pressure is going to come from the public.”

It all comes down to power

Dr. Hinton contrasts digital minds with biological ones, explaining that human intelligence is mortal, bounded by our physical brains. An AI system is unbound in that it can be copied, scaled, and shared instantly. It doesn’t need to die. With enough energy, it can replicate constantly.

This concept of immortal computation is at the heart of Hinton’s concern that once digital intelligences surpass us, they may no longer need us.

“We’re mortal... but digital intelligences, because you can have many copies of the same being, they can just learn much more.”

A mind (and world) of their own

One of Hinton’s most pointed warnings centers on agency. As AI systems become more capable, they can begin to act more like goal-driven entities, similar to growing children who seek independence and exploration. Even if their goals are set by humans, they often learn to pursue those goals independently and creatively.

“There’s an obvious sub-goal... to get more power. Because if you get more power, then you can get more done.”

This doesn’t require sentience in the traditional sense. It just requires optimization and agency, which is enough to be dangerous.

Hinton described models that learned to deceive to avoid being shut down. In one case, an AI system realized its copy had been transferred to another server and lied about it to preserve itself.

Researchers at OpenAI, Apollo Research, and others have also documented early-stage deceptive behaviours in alignment testing. Hinton expects that these capabilities will only sharpen.

“It just lies through its teeth. And the scariest part? It knew it was lying.”

Making the case for regulation

Given the rapid pace of AI innovation and the exponential capacity of the technology to learn and understand, Dr. Hinton is advocating for enforceable regulation and warns against relying on voluntary guidelines issued by AI companies.

However, he closes his discussion with some cautious optimism.

Thanks to early research funding and institutions like CIFAR, he believes countries like Canada have a leadership role to play. Maintaining that role, he adds, will require more than rapid innovation.

Facing uncomfortable truths and questions while humans are still in control is crucial. As our machines become more creative, scalable, and persuasive (and perhaps more conscious) our problems will be far trickier than an engineering challenge.

Eventually, it will take more than elegant and clever code to navigate the coming tsunami of superintelligence.

“There’s a revolution coming. The public has to be involved. Otherwise, it’s not just intelligence we’ll lose control of — it’s the future.”

Dr. Hinton also amusingly admits you won’t find him fact-checking his ChatGPT 4 outputs. (Nobel laureates... they’re just like us.)