Deepfake. What used to be a word relegated to far corners of the internet and conspiracy theorists has become an everyday word in the news. In just a matter of years, AI (artificial intelligence) has flooded our daily lives, being championed as humanity's savior and derided as humanity's destroyer.

While many people are embracing AI tools for productivity or for pleasure, plenty of others are simply ignoring them. However, the results of AI are bombarding us day by day - and some of those results are intentionally created to lie to us about reality.

What is a deepfake?

The term deepfake was, according to MIT, coined in 2017 by a user on Reddit who used face swapping technology on pornographic videos. While MIT goes on to talk about swapping real people and voices with ones generated by technology, our recent models of AI have made it possible for even more complex generation and mimicking of real people and voices.

Kissin' 97.7 logo
Get our free mobile app

However, the concept is older than that. Some of you may remember the movie The Running Man, which features a similar concept as part of its core plot.

In fact, the idea of "fake" people has been around for a long time - photomanipulation has been in use for over 100 years. It's just gotten more complex alongside our technology.

How good are deepfakes, really?

In 2021, a sensational case hit international media as a case was brought a cheerleader's mother for allegedly creating "deepfake" videos of some of the cheer squad girls vaping and doing other activities that would get them kicked off the team. The panic was evident: deepfakes were everywhere, and could even be used to trick mothers that their own children were doing bad things.

However, that case turned out to be more complex - and ultimately, no deepfakes were found to have been created. Experts agreed: the videos were real. The parents had been duped by their own children into believing they were created by AI.

The fact of the matter is, AI tools aren't really 'that' good. They're certainly capable of fooling the average person, but a critical eye can find the flaws created by our imperfect technology. However, they are being used more frequently by the day - especially in a contentious election year in the United States.

The University of Washington will teach you to spot deepfakes for free

No, this isn't some scam or secret online course. Professors at the University of Washington put together an online project called "Which Face is Real?". The game is aimed at helping people learn how to spot fakes so they can be more critical when reviewing media. (If you are interested, though, the University of Washington does offer a course in Calling Bullshit.)

Before you head off to play the game, let's give you a quick quiz of our own!

Deepfake Quiz

Think you can spot a deepfake? Here are five simple tests. Can you tell which one is real or fake?

If not, don't worry - we've got the answers too, as well as some useful clues to spotting those AI generated images.

Gallery Credit: Jaime Skelton

How'd you do? Keep getting more practice with the "Which Face Is Real?" game until you feel like you can't be fooled.

More tips for spotting deepfakes:

Experts at AIPRM have offered the following tips:

1. Verify the source and context. 

  • Examine the credibility of the sources.
  • Cross-check the authenticity of the source.
  • If the content provokes emotions, "pause and re-assess" to avoid being played.

2. Observe facial expressions and movements.

  • Double check 'microexpressions' around the eyes and mouth.
  • Watch for unusual movement patterns and mismatched facial expressions.
  • With videos, pause and examine closely for irregularities.

3. Use reverse image and video search tools

  • Utilize tools like Google Reverse Image Search and InVID to look for the original source.
  • Check to make sure that in the context, the image is not being misused.

4. Look for digital artifacts and inconsistencies.

As seen in the quiz and game above, flaws can give away the AI game.

  • Check for blurriness or pixelation, especially around the edges of faces and objects.
  • Look for inconsistencies in lighting, shadows, reflections, and distortions.

5. Check for audio-visual synchronization

  • Watch lips closely - AI is not good at replicating complex speech movements.
  • Red flags on movements that don't match the sounds you hear, or oddities in tone and pitch.

Will Washington State implement rules against deepfakes? We don't know yet, but until then, stay vigilant against misinformation.

This Evidence Of A Dinosaur in Boise is Definitely NOT Fake (VIDEO)

It's easy to dismiss a video and say it's fake but what if this is actual evidence of a dinosaur flying over the Treasure Valley?

Gallery Credit: Chris Cardenas

AI-Generated Pokemon Based on Washington Towns

Love Pokemon? So do we! So we asked AI to come up with unique "Fakemon" (fake Pokemon) that represent ten major Washington towns.

Gallery Credit: Jaime Skelton

 

More From Kissin' 97.7