The deepfake dilemma: When you can’t trust your eyes

The inference is clear — it’s hard for anyone to trust anything if AI-powered tools can create videos, photos and text that appear accurate on first glance. Social media platforms like TikTok, Instagram and X make matters worse by providing limited safeguards for users to know what’s real and what’s fake. ”

Rob Tornoe, digital editor and writer, The Philadelphia Inquirer

Recently, a friend sent me a profane video featuring a former NFL head coach he thought I would find funny. The clip was amusing, but it also gave me a headache, because for the first time, I genuinely couldn’t instantly determine if the video was a deepfake or not. 

Deepfakes have come a long way over just the past few years, from laughable attempts to show Will Smith eating spaghetti to convincing videos of celebrities and politicians doing just about anything. 

Recently, former New York Governor Andrew Cuomo turned to tools powered by artificial intelligence (AI) to convincingly show himself as a subway conductor and a window cleaner in an ad for his mayoral campaign. A small disclaimer at the bottom of the ad, required by New York law, notes the ad “was created with the assistance of artificial intelligence.” 

He’s not the first. Current New York City Mayor Eric Adams turned to AI photos and videos before dropping his reelection campaign, and President Donald Trump has been sharing AI-altered videos of his political opponents for months, most notably putting a sombrero on the head of House Minority Speaker Hakeem Jeffries. However, these campy attempts at mimicking online humor fall short of the realism that AI-powered tools can produce. 

That’s why Cuomo’s video feels like a “cat out of the bag” moment. Companies like Google and OpenAI used to limit or prohibit users from creating AI-doctored videos featuring public figures for obvious reasons. Those guardrails are quietly being dismantled, with OpenAI facing widespread criticism (and the threat of lawsuits) for allowing users on its new social video app Sora to use celebrities, politicians and copyrighted material freely. 

The result? An endless stream of AI slop featuring everything from realistic videos of deceased celebrities to shockingly realistic clips of real people as Nazis to disturbingly accurate but completely fabricated videos of historic events. Vox’s Bryan Walsh aptly compared the mashup of a TikTok-style social media platform with AI-created content to taking heroin and mixing it with heroin — a brain-rotting, highly-addictive flood of garbage content predictably seeping onto more popular social media networks.

OpenAI’s Sam Altman ultimately reversed course to create more controls for copyright holders over their intellectual property, which likely staved off the lawyers but did nothing to alter the downstream effects of his misinformation machine. 

There are real consequences. In 2024, a principal at a suburban Baltimore school was placed on administrative leave after an audio recording using racist and anti-Semitic slurs went viral on social media. The clip turned out to be an AI-generated fake created by the school’s athletic director, who was under investigation for allegedly stealing hundreds of dollars from the school. 

Thankfully, incidents such as this remain rare. A study conducted by professors Soubhik Barari, Christopher Lucas and Kevin Munger and published in the Journal of Politics earlier this year revealed that manipulated videos themselves are unlikely to deceive the public on a massive scale.

However, they did find that the existence of so much potentially fake content has the power to continue an erosion of media trust that has occurred over the past two decades. After all, if we can’t believe our eyes and ears any longer, what can we trust?

As part of the study, Republicans were far more likely to think a real video of President Donald Trump getting Apple CEO Tim Cook’s name wrong was a deepfake. The same went for Democrats shown a real video of former President Barack Obama appearing to make a post-election deal with Russian President Vladimir Putin. 

“In sum, our work suggests that deepfakes are not uniquely deceptive, but their existence may discredit real media,” Lucas, an associate professor of political science at Washington University in St. Louis, said in a statement. “Our research also shows that people often misjudge authentic news as fake, especially when it depicts a political figure from their own party in a negative light.” 

The inference is clear — it’s hard for anyone to trust anything if AI-powered tools can create videos, photos and text that appear accurate on first glance. Social media platforms like TikTok, Instagram and X make matters worse by providing limited safeguards for users to know what’s real and what’s fake. 

How to figure out if a video is real or a deepfake

In my own case, I had nagging thoughts about the authenticity of a video of former New Orleans Saints head coach Jim Mora cursing at reporters during practice in an almost-too-good-to-be-true summation of what the NFL was like more than two decades ago. While it seems real, older videos are also more susceptible to deepfakes due to low-resolution cameras and a lack of visual fidelity that can obscure some of the trademarks of AI meddling. 

One thing I discovered is that there aren’t really any great, reliable options for journalists attempting to figure out if a video is authentic or a deepfake. Although numerous free tools can detect fake images or photos, there is no equivalent tool for videos that’s easy to use on a deadline. Those touted elsewhere were either glitchy, performed poorly or continually shot back error messages. 

Researchers at Cornell referred to this as a new “battleground,” with AI detection tools unable to keep up with the tech companies pumping out increasingly convincing slop. Experts at the Tow Center for Digital Journalism argue that detection tools are unreliable for catching content created or manipulated by AI and may even contribute to a false sense of security by weakening “journalistic rigor.”

One tip that worked for me as an initial gut check comes from online research expert Henk van Ess, who suggested using Hiya Deepfake Voice Detector. It’s a Chrome plugin that analyzes audio in real time to determine if what someone’s saying is real or artificially created and provides an authenticity score between one and 100. 

Obviously, it only works if the clip has audio, but I tested Hiya with 10 real videos and 10 deepfakes, and it nailed all 20. For one fake video — an AI-altered campaign ad featuring former Vice President Kamala Harris, famously shared by Elon Musk — Hiya gave it an authenticity score of 1/100, noting the “sampled voice is likely a deepfake.” 

Likely isn’t certainty, though, so it remains on journalists to be vigilant and skeptical about the content that comes across social media. The best advice is to employ the “Sift” method as pioneered by digital literacy expert Mike Caulfield, which stands for stop, investigate the source, find better coverage and trace the claim (or video) to its original context. 

And for the record, the Mora video I was sent was real. I couldn’t verify its authenticity visually (though Hiya did note it was “likely authentic”). However, I was able to find reports about the scrum from trusted sources deep in the archives of Newspapers.com. 

Sometimes the old ways are still the best. 

Rob Tornoe is a cartoonist and columnist for Editor and Publisher, where he writes about trends in digital media. He is also a digital editor and writer for The Philadelphia Inquirer. You can reach him at robtornoe@gmail.com.