How to spot a deepfake video
Your how-to guide for the Sora 2 era
At the end of September, OpenAI launched Sora, a groundbreaking app for AI-powered video generation. Think TikTok, but every clip is created by artificial intelligence (AI) – from Olympic gymnastics routines to backflips on a paddleboard.
Sora2 launch video demonstrating the platform's deepfake technology.
It’s positioned as a fun way to co-create and connect with friends, but it also raises serious questions: how do we tell fact from fabrication in a world where AI can mimic reality with astonishing precision?
We asked Dr Priyanka Singh, UQ Lecturer in Cyber Security, to get us up to speed on this phenomenon and give us her practical tips on how to spot a fake.
Dr Priyanka Singh
Dr Priyanka Singh
At the end of September, OpenAI launched Sora, a groundbreaking app for AI-powered video generation. Think TikTok, but every clip is created by artificial intelligence (AI) – from Olympic gymnastics routines to backflips on a paddleboard.
Sora2 launch video demonstrating the platform's deepfake technology.
It’s positioned as a fun way to co-create and connect with friends, but it also raises serious questions: how do we tell fact from fabrication in a world where AI can mimic reality with astonishing precision?
We asked Dr Priyanka Singh, UQ Lecturer in Cyber Security, to get us up to speed on this phenomenon and give us her practical tips on how to spot a fake.
Dr Priyanka Singh
Dr Priyanka Singh
What exactly is a deepfake?
A deepfake is a synthetic audio, image or video created using AI, usually by training deep learning models – like Sora 2 – to mimic a person’s face, voice, expressions or mannerisms. It can look extremely real, even though the person never actually said or did what’s depicted.
Why are they becoming harder to detect?
AI models are improving rapidly. Tools that once required specialised skills are now publicly available, and they generate content with smoother facial movements, better lip-syncing and more realistic lighting. The gap between real and fake is narrowing so quickly that even experts sometimes struggle to discern between the 2 without using technical tools.
This is an example of a convincing deepfake, designed to look like security camera footage.
What are some practical tips for determining that a video might be AI-generated or manipulated?
Look for small inconsistencies like:
- Unnatural blinking
- Mismatched lighting or shadows
- Skin that looks overly smooth
- Slightly distorted teeth or eyes
- Odd reflections on glasses
- Audio that is out of sync with lip movement
- Emotional expressions that feel ‘flat’ or robotic.
How reliable are current detection methods for everyday users?
Detection tools are improving but none are perfect, especially for high-quality deepfakes. Most tools work best on older or low-quality manipulations. For advanced, real-time AI-generated videos, human judgment plus source verification, by identifying the original poster and cross-referencing with trusted sources, is still the most reliable combination.
What are the risks associated with deepfakes?
Deepfakes can fuel misinformation, impersonation scams, political manipulation, reputational damage and gender-based violence – the most harmful example being non-consensual synthetic pornography. They erode public trust by making it harder to distinguish truth from fabrication in the videos we see online.
Are there any positive use cases for deepfake technology?
Yes, when used ethically, deepfake-style technologies can help in filmmaking, heritage preservation and personalised digital assistants. The challenge is ensuring these innovations are used responsibly and transparently.
Image: Richard O'Donoghue via Adobe Stock.
Image: Richard O'Donoghue via Adobe Stock.
Image: TechAnimationStock via Adobe Stock.
Image: TechAnimationStock via Adobe Stock.
What exactly is a deepfake?
A deepfake is a synthetic audio, image or video created using AI, usually by training deep learning models – like Sora 2 – to mimic a person’s face, voice, expressions or mannerisms. It can look extremely real, even though the person never actually said or did what’s depicted.
Why are they becoming harder to detect?
AI models are improving rapidly. Tools that once required specialised skills are now publicly available, and they generate content with smoother facial movements, better lip-syncing and more realistic lighting. The gap between real and fake is narrowing so quickly that even experts sometimes struggle to discern between the 2 without using technical tools.
This is an example of a convincing deepfake, designed to look like security camera footage.
What are some practical tips for determining that a video might be AI-generated or manipulated?
Look for small inconsistencies like:
- Unnatural blinking
- Mismatched lighting or shadows
- Skin that looks overly smooth
- Slightly distorted teeth or eyes
- Odd reflections on glasses
- Audio that is out of sync with lip movement
- Emotional expressions that feel ‘flat’ or robotic.
How reliable are current detection methods for everyday users?
Detection tools are improving but none are perfect, especially for high-quality deepfakes. Most tools work best on older or low-quality manipulations. For advanced, real-time AI-generated videos, human judgment plus source verification, by identifying the original poster and cross-referencing with trusted sources, is still the most reliable combination.
What are the risks associated with deepfakes?
Deepfakes can fuel misinformation, impersonation scams, political manipulation, reputational damage and gender-based violence – the most harmful example being non-consensual synthetic pornography. They erode public trust by making it harder to distinguish truth from fabrication in the videos we see online.
Are there any positive use cases for deepfake technology?
Yes, when used ethically, deepfake-style technologies can help in filmmaking, heritage preservation and personalised digital assistants. The challenge is ensuring these innovations are used responsibly and transparently.
Image: Richard O'Donoghue via Adobe Stock.
Image: Richard O'Donoghue via Adobe Stock.
Image: TechAnimationStock via Adobe Stock.
Image: TechAnimationStock via Adobe Stock.
Built with Shorthand