OTTAWA — The footage captures an actual historical event: a speech given by Nazi dictator Adolf Hitler in 1939 at the start of World War II.
But there's one big difference: the viral video has been altered by artificial intelligence, and in it, Hitler makes anti-Semitic remarks in English.
A far-right conspiracy influencer shared the content on X (formerly Twitter) earlier this year, and it quickly garnered more than 15 million views, Wired magazine reported in March.
It's just one example of what researchers and organizations that monitor hateful content call a worrying trend.
It is said that AI-generated hatred is on the rise.
“I think anyone who studies hate content and hate media is going to see more and more AI-generated content,” said Peter Smith, a journalist who works for the Canadian Anti-Hate Network.
Chris Tennove, associate director of the Centre for the Study of Democratic Institutions at the University of British Columbia, said white supremacist and other hate groups “historically have been quick to adopt new internet technologies and techniques”.
It's a concern that a UN advisory body flagged in December, saying it was “deeply concerned” about the possibility that anti-Semitic, Islamophobic, racist and xenophobic content “may be amplified by generative AI.”
Sometimes, the content seeps into real life.
Smith said the AI was used to generate “extremely racist Pixar-style movie posters,” which some people then printed out and pasted on the sides of cinemas.
“Anything that's available to the public, that's popular, or that's new, especially when it comes to technology, can be very quickly used to generate hate propaganda.”
Generative AI systems can create images and videos almost instantly with just a few simple prompts.
Instead of an individual spending hours creating a single image, Smith said, they could create dozens of images “with just a few keystrokes in the same amount of time.”
B'nai Brith Canada pointed to the problem of AI-generated hate content in a recent report on anti-Semitism.
The report said that last year there was “an unprecedented increase in anti-Semitic images and videos created, altered or fabricated using AI.”
Richard Robertson, director of research and advocacy, said the organisation had seen “some truly horrifying and graphic imagery being produced using AI, linked to the denial, minimisation and distortion of the Holocaust”.
He gave the example of doctored images that showed an amusement park inside a concentration camp.
“It looks like Holocaust victims are riding in vehicles and having fun in a Nazi concentration camp, which could probably only be created using AI,” he said.
The organization's report also said AI had a “profound impact” on the spread of Israeli-Hamas propaganda in the wake of the war.
AI can be used to create deepfakes – highly realistic videos of celebrities, politicians and other public figures.
Tennove said deepfakes, set in the context of the Israel-Hamas war, have spread disinformation about events and attributed false claims to both the Israeli military and Hamas officials.
“That is to say, these are actions that are aimed at inciting anger and fear in people against the other side, and then engaging in deception to achieve that.”
Jimmy Lin, a professor in the University of Waterloo's computer science department, agreed that “we're seeing an increase in fake content that's specifically designed to provoke people on both sides.”
Amira Elghawaby, Canada's special representative against Islamophobia, says both anti-Semitic and Islamophobic rhetoric has increased since the conflict began.
She said the issue of AI and hate content needed further research and discussion.
There is no dispute that AI-generated hate content is an emerging problem, but experts have yet to agree on the scope of the problem.
Tennove said “there is a great deal of speculation at this point,” as well as broader societal questions about “harmful or problematic content spreading on social media platforms.”
Lin said systems like ChatGPT have safeguards built into them, and an OpenAI spokesperson confirmed that the company trains its models to reject generating hate speech before releasing new systems.
But Lin said there are ways to jailbreak AI systems, noting that certain prompts could “trick the model” into creating so-called objectionable content.
David Evan Harris, a professor of public policy at the University of California, Berkeley, said it's hard to know where AI content comes from unless the companies developing these models ensure they're watermarked.
He said some AI models, like those created by OpenAI and Google, are closed-source models, while others, like Meta's Llama, are more openly available.
If the system were opened up to everyone, bad actors could remove safety features and create hate speech, scams and phishing messages in ways that are very difficult to detect, he said.
Meta said in a statement that it has built safeguards into its system and is not open sourcing “everything.”
“Open source software is typically safer and more secure due to ongoing feedback, scrutiny, development and mitigations from the community,” the company said.
In Canada, there are federal laws the Liberal government says will help address the problem, including Bill C-63, Addressing Online Harms.
Chantal Obertun, a spokeswoman for Justice Minister Arif Virani, said the bill's definition of hateful content includes “any kind of content, such as images or videos, as well as any artificially generated content, such as deepfakes.”
Innovation Canada said its proposed Bill C-27, which would regulate artificial intelligence, would require AI content to be made identifiable through watermarking or other means.
A spokesperson said the bill would also require companies “responsible for high-impact general-purpose AI systems to assess the risks, test and monitor the systems to ensure they are working as intended, and implement appropriate mitigations to address the risks of harm.”
