AI videos just got very real: Here’s how to spot them - What in the World podcast, BBC World Service

Key Concepts

• AI-Generated Content: Content created using artificial intelligence, specifically machine learning algorithms trained on vast datasets to mimic human-like work.
• Vibes (Meta): A new AI video development tool from Meta, described as a creative tool for expressive, AI-generated videos from artists and creators. It utilizes tools like Midjourney and Black Forest.
• Sora (OpenAI): A new AI video generation tool from the makers of ChatGPT. It aims for a more social media-like experience, allowing users to create content by scanning their faces and placing themselves in various scenarios, or by using celebrity likenesses.
• Disinformation: False or inaccurate information, especially that which is deliberately intended to deceive.
• Erosion of Trust: The decline in people's willingness to believe or rely on information, particularly visual media, due to the prevalence of AI-generated content.
• Rage Bait/Reaffirming Prejudices: Content designed to provoke strong emotional reactions (anger, outrage) or to reinforce existing biases and beliefs.
• Uncanny Valley: A phenomenon where AI-generated content, while appearing realistic, exhibits subtle imperfections that create a sense of unease or revulsion in the viewer.

AI-Generated Content: The Evolving Landscape

The internet is becoming increasingly saturated with AI-generated content, a trend amplified by the launch of new tools like Meta's Vibes and OpenAI's Sora. This raises significant questions about authenticity and the ability to distinguish real from synthetic media.

What is AI-Generated Content?

AI-generated content encompasses anything created by artificial intelligence, specifically through machine learning algorithms. These algorithms are trained on extensive datasets, enabling them to produce new content that imitates human-created work. This technology is not limited to social media but is infiltrating various aspects of life, including coding ("code slop") and presentations ("work slop").

The Rise of AI Content on Social Media

While precise figures are difficult to ascertain, AI-generated content is steadily increasing on social media platforms. In the US, it's particularly noticeable, with AI-powered algorithms influencing 80% of content recommendations. Notably, AI videos are often favored by these algorithms. One analysis suggested that over half of English-language posts on LinkedIn might be AI-generated. The prevalence is also growing on platforms like TikTok, Instagram Reels, and YouTube Shorts, though exact numbers are elusive.

Evolution of AI Video Realism

The quality of AI-generated videos has dramatically improved. Early examples, like the "Will Smith eating spaghetti" video from around three years ago, were noticeably rough, with distorted facial features and unnatural object interactions. However, with newer tools, such content has become significantly more realistic, making it harder for viewers to discern its artificial origin. This advancement has led to instances where viewers, like the speaker, have been fooled into believing AI-generated content is real.

New AI Video Tools: Vibes and Sora

Two prominent new applications are pushing the boundaries of AI video generation:

• Vibes (Meta): Described as a creative tool, Vibes uses existing AI video-making technologies like Midjourney and Black Forest. Mark Zuckerberg envisions it as a platform for expressive, AI-generated videos from artists and creators.
• Sora (OpenAI): This tool aims for a more social media-centric experience, resembling TikTok's "for you" page. Sora is currently invite-only and encourages user-generated content. Users can create videos by:
  • Scanning their own face: Taking a video of their face, performing actions, and speaking numbers to be integrated into various scenarios.
  • Using celebrity likenesses: This feature has led to copyright issues, with instances of characters like SpongeBob and Mickey Mouse being depicted in inappropriate or uncharacteristic ways. Sora faces challenges in managing the ethical implications of using real people's likenesses and creating content they would never endorse.

Examples of Viral AI Content

A typical example of a viral AI video from Sora involves a woman falling backward onto a glass bridge, breaking it, and a child falling through, only to be rescued by a dog. This type of content often incorporates elements designed for virality:

• Rage Bait: Provokes anger and strong reactions in viewers (e.g., criticism of the woman's actions).
• Fear: Creates suspense and concern for the child's safety.
• Resolution: Offers a positive and often heartwarming conclusion, such as an animal rescue.
• Animal Content: Videos featuring animals, especially in rescue scenarios, tend to perform exceptionally well with AI-generated content.

Listener Perspectives on AI Content

Listener feedback highlights a mixed reception to AI-generated content:

• Polina: A self-proclaimed "big fan" of AI-generated videos, particularly for comedy shows and historical POV videos. She finds them engaging, a form of "brain rot" after work, and a vibrant way to learn history, making social media more immersive.
• Maria: Expresses a dislike for AI videos, preferring "real people with real emotions and real faces." She finds AI videos unsettling due to unexpected morphing or melting faces and considers them impressive but creepy rather than art.

The Impact on Social Media Usage and Trust

There are indications that social media usage may have peaked in 2022, with a subsequent steady decline, particularly among Gen Z. While this trend could be influenced by various factors, including the post-COVID return to normalcy, the increasing prevalence of AI-generated content is a potential contributing factor.

The growing difficulty in distinguishing real from AI-generated content leads to an erosion of trust. Viewers may feel compelled to scrutinize every piece of content, potentially leading to skepticism towards even genuine videos.

Identifying AI-Generated Content: Evolving Challenges

The methods for identifying AI-generated content are becoming less reliable:

• Past Indicators: Previously, common tells included incorrect fingers, unnatural blinking, or poorly rendered ears.
• Current Challenges: The "Will Smith eating spaghetti" example, despite its improvements, still shows subtle flaws like unnatural blinking and overly smooth spaghetti consumption. For simpler videos of people speaking directly to the camera, especially well-known figures, these older indicators are often insufficient.

Strategies for Identifying AI Content

The speaker, Jacqui Wakefield, offers a practical approach:

• Likely Scenario: Assess whether the content depicts a plausible situation. For instance, a video of a typically "macho" figure like Jake Paul doing elaborate makeup tutorials might be a red flag, as it deviates from known behavior.
• Contextual Knowledge: Utilize existing knowledge about individuals and the world to gauge the authenticity of the content.
• Subtle Imperfections: While harder to spot, look for subtle inconsistencies in movement, facial expressions, or object interactions that feel "not quite right."

However, these methods are becoming more time-consuming, and the "hacks" that were effective a few years ago are no longer reliable.

Broader Implications of AI-Generated Content

The proliferation of AI-generated content carries significant real-world implications:

• Pure Disinformation: AI can be used to create fabricated events, making people believe things that never happened.
• Erosion of Trust: As mentioned, the constant need to verify content can lead to a general distrust of online information.
• Rage Bait and Reaffirming Prejudices: AI content can be weaponized to amplify existing biases. For example, AI-generated videos of protests can depict participants in stereotypical ways that align with pre-existing negative views of certain groups (e.g., LGBTQ+ or anti-immigration protests). This can reinforce prejudices by presenting biased narratives as factual.

Conclusion and Call to Action

The increasing sophistication and accessibility of AI video generation tools like Vibes and Sora mean that AI-generated content will become even more prevalent. This necessitates a heightened awareness and critical approach from users. The ability to discern real from synthetic content is becoming more challenging, underscoring the importance of checking sources and applying critical thinking to online information. The speaker emphasizes that "it's really important people check their sources, now, more than ever."