Have you learnt what artificial intelligence appears like? When requested to guess, most individuals can’t inform the distinction between AI-generated voices and actual human dialog, in accordance with multiple studies.
This confusion can have disastrous penalties on how we see the world. If you get confused about what’s actual or not on display, you can begin to consider misinformation, and in worst instances, racist stereotypes about individuals being depicted in AI-generated videos.
However there is perhaps one dependable method to suss out what’s AI, particularly on video: Hearken to how the individuals sound.
A variety of AI consultants shared the telltale indicators of why the voices and sounds in an AI video can usually reveal its artificial origin. Right here’s how.
Illustration: HuffPost; Pictures: Getty
Hear for the over-caffeinated tone.
Actual individuals have a pure rhythm to how they converse, in order that some phrases are mentioned extra slowly than others. However AI voices usually sound unnaturally rushed on a regular basis.
Jeremy Carrasco, a video professional who debunks AI movies on social media, mentioned he notices that movies from Sora ― a synthetic intelligence video app owned by OpenAI ― usually have an “overly energetic” high quality. “They’re saying a lot they usually’re not saying a lot in any respect, they’re simply cramming in phrases,” he mentioned.
Even OpenAI is conscious of this telltale signal. Too many em dashes in a textual content reply is known to be a giveaway in OpenAI’s ChatGPT solutions that may reveal when somebody’s cowl letter or first date message acquired AI-generated.
In October, the hosts of video streaming present TBPN requested Invoice Peeples, the pinnacle of Sora, about what the “em sprint of [AI] video” was in an interview. His speedy response was telling.
“I believe proper now the ‘em sprint’ is that this barely wired speech sample in Sora the place it likes to say loads of phrases rapidly,” Peeples said.
Be careful for garbled, slurred voices.
What we would name somebody’s talking rhythm is what linguists would name “coarticulation,” or how our voices bodily go from one sound to a different as air goes via our noses and out our mouths. And loads of AI-generated speech continues to be dangerous at this and makes garbled sounds that seem to flatten out pure sound pitches.
“No human being would ever produce that very same sort of garbled high quality [as an AI-generated voice], as a result of, actually, we will’t,” mentioned Melissa Baese-Berk, a linguistics professor on the College of Chicago. “Our vocal monitor can’t go from one sound to a different with out some blurring of the knowledge between these two sounds.”
Baese-Berk used the instance of an AI subway meet-cute video the place a girl meets a person she instantly calls her “husband.” The video fooled many individuals into believing it was actual. However when the lady says “husband,” the “band” a part of the phrase sounds “tremendous duper bizarre,” she mentioned. The “band” a part of the phrase “is lacking the pure coarticulatory info that occurs whenever you transfer from the tip of your tongue to your lips,” Baese-Berk mentioned.
“Solely a robotic might go from their tongue to their lips with out having any sort of mashing up of these sounds,” Baese-Berk mentioned.
This inhuman mash-up of phrases is by design.
“Textual content-to-speech fashions are educated to foretell the almost definitely pronunciation of a phrase in sequence, however they usually battle to easily mix the sounds that join phrases,” mentioned Migüel Jetté, vice chairman of AI at Rev, a speech-to-text service. “For instance, the place a human would possibly naturally say ‘didja’ as a substitute of ‘did you,’ AI tends to both over-enunciate every phrase, or mix them too abruptly.”
Take note of mispronounced phrases.
If there’s an clearly mispronounced phrase, that can be an indication, Jetté mentioned, as a result of “AI voices can battle with uncommon or distinctive phrases that don’t seem within the coaching information.”
Google’s text-to-video Veo mannequin, for instance, “won’t be cramming in as many phrases, however they may put them out of order, or the flawed particular person will say one thing,” Carrasco mentioned he has noticed.
Discover when emotional reactions don’t match the story of the video.
In a 2025 study that requested contributors to fee which voices had been AI or not, the AI voices created by text-to-speech fashions had been solely recognized precisely 55% of the time. The most important errors occurred with AI voices that sounded offended.
This can be as a result of contributors anticipated AI voices to sound robotic, mentioned Camila Bruder, a co-author of that research and a researcher from the Max Planck Institute for Empirical Aesthetics.
In actuality, AI voices are sometimes too emotional for what the scene requires. If the AI voice is “too stereotypically joyful, like, ‘Wow!’ or it’s stereotypically mad…like a foul actor,” these traits will be indicators that the video is AI, Bruder mentioned.
Carrasco mentioned you must also discover when what’s being mentioned is an odd emotional response. Take one viral AI video of fish falling from the sky. “They’re fish, they’re really fish!” a girl within the video exclaims.
“They’re simply narrating what’s occurring on the display. You wouldn’t do this in actual life,” Carrasco mentioned about this video. “If a bunch of fish had been raining [down], I’d in all probability simply say ‘What the fuck.’”
Evaluate the inappropriate AI feelings to the real-life horror a truck driver not too long ago skilled when he was filmed watching a airplane crash that occurred in entrance of him in Kentucky. On this video, the motive force doesn’t narrate his expertise, his mouth merely drops open. “He’s simply in disbelief. That’s sort of how loads of these can be” in actual life, Carrasco mentioned.
You can even merely take a look at what individuals’s mouths are doing for clues. “The visible giveaways in these movies will be simply as revealing because the audio,” Jetté mentioned. “If the speaker’s lips don’t completely sync with the audio…that’s a robust indicator.”
These clues are useful, however they aren’t all the time assured.
After all, these clues will not be all the time a assured method to reveal an AI-generated voice. ElevenLabs, the AI lab which clones actual voices, is sweet at including vocal fry and human pauses, so listening for a voice that speaks with out breaths isn’t “all the time the case” that it’s AI, Bruder mentioned.
However as a complete, these telltale indicators are a robust indicator that the video you might be watching was in all probability created by a machine. And that’s a useful begin. As AI continues to evolve at breathtaking speeds, we’d like all the assistance we will get to know what’s faux and what’s not.
“If one thing feels off, it in all probability is,” Jetté mentioned. “A wholesome dose of skepticism and an excellent eye and ear for element can go a good distance.”
