Help! I’m Nostalgic for the Days of AI Slop

My morning started with the most uncanny-valley video I neither needed, wanted nor asked for.

Scrolling Instagram for a dopamine hit to wake myself up on a Monday — sue me! — I came across a video from Mexico’s Paco De Miguel, a 26-year-old comedian and digital content creator who GQ Mexico named 2024’s comedian of the year.

De Miguel mainly performs impersonations of a stereotypical Mexican mother-daughter relationship where he, of course, plays both, distinguishing each character with chaotically placed wigs.

Every time I see his videos, I can’t help but giggle uncontrollably and share them with my sister, who, like me, grew up listening to our own Mexican mother lecture and embarrass us in Spanish. Of the people in my life, I know that only she can appreciate how deeply relatable these hyper-specific portrayals are.

View this post on Instagram

A post shared by Paco De Miguel (@pacodemiguel)

But the second the video began that Monday morning, something about it made me feel deeply confused and uncomfortable.

De Miguel’s voice was the same, his intonations the same, the plot and jokes the same. But the video was in English instead of Spanish. The Mexican accent, the rolls of the tongue, the slang were all gone.

Within a couple seconds, I saw the label: “Translated with Meta AI.”

View this post on Instagram

A post shared by Adam Mosseri (@mosseri)

Are we cooked?

I’d like to think that I’m decently media literate. I can spot AI — or at least generative AI, which uses machine learning techniques and data it scrapes from the internet to generate text, images and videos — as well and quickly enough as the next guy.

My previous AI litmus tests included checking a subject’s hands (do they have six fingers on each hand and then suddenly four?), their skin (humans have textured skin whereas generative AI makes faces look like a piece of sea glass) and whether objects in the foreground disappear into the background without explanation, and vice versa.

So, what threw me more than the comedian delivering punchlines in English — which did not hit the same, thank you very much — was the fact that I didn’t even realize the content was AI until I saw Instagram’s fine-print disclosure of it in the bottom left corner.

I pondered how De Miguel’s lips moved so perfectly in English, how his English timbre was so identical to that of his Spanish (people sound different in different languages, regardless of fluency, and translations are never one-to-one) and how in the years of following him I’d never once heard him speak English on his platform until now.

“Have you ever wished you could magically speak another language so that your content could reach and resonate with people all over the world? With the help of Meta AI, this is starting to become reality,” Meta boasts on its website about its new AI translation tool, which launched worldwide late this summer.

According to the company which owns Facebook, Messenger, Instagram, Threads and Whatsapp, this new AI translation tool allows creators to have their voices dubbed and lip-synced in Instagram’s vertical videos, also called Reels, from Spanish to English and vice versa. Having found success with these initial languages, Meta announced Oct. 9 that it added Hindi and Portuguese to the list of translatable languages.

“As a Spanish speaker, I am both amazed but also terrified,” said Irving Torres, a San Diego-based digital creator, in an Instagram Reel.

“Are we cooked? Because although this is a very, very exciting feature, are we ready for this? Is Meta the right company to lead the charge, knowing what we know about them and everything that they’ve completely messed up in the past?”

Users do have some control over the feature: when posting a video, you have to turn the AI translator tool on.

Instagram also allows you to review the translation before posting. The app warns that automated translations could have inaccuracies — a potentially harmful implication for users who want to translate their video but don’t speak both languages to fact-check the auto-translations. But Torres notes that in his experience, translations have been fairly accurate.

Automatic translations and lip-syncing are useful tools for creators looking to expand their audiences and render their content accessible to more people. But to what end, and at what cost?

The unceremonious launch and use of a tool that lets us hand over our faces and voices to generative AI with the toggle of a button — agreeing to terms and conditions we likely don’t understand — seems to mark a shift away from any principles we once held around how we use generative AI, and how it, as a result, uses us.

As generative AI becomes more realistic, my ability to tell what’s real isn’t vanishing as much as it’s transforming into a paranoia.

In that moment Monday morning, watching the English translated version of De Miguel’s reel, I felt my brain recalibrate. I realized that even on my highly curated feeds, where I regularly turn off suggested posts and only follow friends or accounts that I really love, I can no longer trust my own eyes.

It’s not that I can’t do a quick fact-check. It’s that we’ve arrived at a moment in the social web’s timeline where I have to question the veracity of even the most ordinary content at all.

Creators, too, are left in the dark

Meta is not the only social media giant to provoke suspicions of generative AI being used on content unbeknownst to consumers or even creators.

One day while scrolling YouTube Shorts, I noticed a smoothing filter over the video’s subjects. It had the smeary quality of an oil painting that some have identified as a telltale sign of AI-generated content.

I still suspected the video was real. After all, what reason would this inconsequential celebrity interview clip have to use generative AI?

“Something strange has been happening on YouTube over the past few weeks,” Alex Reisner wrote for the Atlantic in August. “After being uploaded, some videos have been subtly augmented, their appearance changing without their creators doing anything. Viewers have noticed ‘extra punchy shadows,’ ‘weirdly sharp edges’ and a smoothed-out look to footage that makes it look ‘like plastic.’ Many people have come to the same conclusion: YouTube is using AI to tweak videos on its platform, without creators’ knowledge.”

A spokesperson for Google, which owns YouTube, told Reisner that the platform is “running an experiment on select YouTube Shorts that uses image enhancement technology to sharpen content.”

The spokesperson added that the technology does not include generative AI. Reisner pointed out, however, that “generative AI” has “no strict technical definition” and that “image enhancement technology” can mean anything.

As YouTube and other tech companies both encourage creators to use generative AI and boast how many creators already do, some suspect that this quiet move by YouTube was to get its consumers “used to” the AI aesthetic.

Can you still spot AI?

I’m far from the only one whose AI-spotting skills may no longer be up to snuff. According to a Scientific Reports study published in March, AI-powered voice clones are highly — but not uniformly — convincing.

In January 2024, thousands of U.S. Democratic Party voters received deepfake robocalls in the voice of former president Joe Biden telling them not to vote in the New Hampshire primaries.

In the study about AI-powered voice clones, researchers had participants listen to two audio clips: one featured the voice of a real person and one featured that of an AI clone. Participants overwhelmingly judged the two clips were recordings of the same person’s voice.

In a separate test, researchers asked participants whether an audio clip was an AI clone. Participants correctly identified the AI clone 60.8 per cent of the time.

“There is good reason to believe that AI-generated voices will soon be indistinguishable from real ones both in terms of naturalness and identity,” write the authors of the study. “While this should be considered a triumph for those on the generative side, it raises real concerns for public safety.

“Our results highlight that relying on human perception to detect AI-generated voice clones is no longer consistently reliable.”

This applies to video-based fraud as well, the study notes. Tools that can detect generative AI for users will become “essential,” write the authors.

These kinds of tools are already well underway, popular among educators. But they’re not without controversy. One study found that these detectors “consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified.”

‘Meow meow meow meow’

Back in the day — one year ago, tops? — before generative AI started clogging our social media feeds, I’d indulge in some TikToks before bed and call my partner over to watch the latest AI-generated cat soap opera.

The short films were soundtracked to the tunes of Billie Eilish pop single “What Was I Made For” or Rihanna’s “Diamonds” as comedically timed cat meows replaced the lyrics.

Chortling at the absurdity of it all, I remember thinking that maybe generative AI doesn’t have to be this big, scary, threatening thing after all. I was perfectly aware how naive that is. But maybe I just wanted to be naive for a moment.

@mpwild Stay unique ! #catsoftiktok #cat #aiart #ai #poorcat ♬ son original - Cat’slife - chubby_s_life

The cat soap operas marked the apex of the AI slop era; it was and is a reminder that, even in our technocratic dystopia, where game-changing inventions can both highlight our humanity and undermine it, some deeply powerful tools will be used for deeply unserious reasons. It’s a reminder for the anxious among us that, sometimes, you can opt to just laugh it off.

Barely any time has passed from the heyday of the cat soap opera, and I already have AI whiplash. Browsing the web nowadays feels, more than ever like, an acid trip. I can’t tell what’s real; why does that guy’s face look so off; I wish I was home and sober already.

That massive, multi-platform channels like the hit podcast Diary of a CEO are experimenting with not one but several shows created entirely using generative AI — from script to production to hosting to editing — is somewhat shocking.

But it’s the more discreet and smaller use cases that demonstrate an inching-over of the AI Overton Window.

Our ability to discern reality might soon — and somewhat already does — depend on the willingness of conglomerate tech companies to disclose the altered content that they algorithmically shoved down our throats in the first place.

As the Atlantic’s Reisner wrote, “Viewers who don’t want to be fooled… must now be alert to the subtlest signs of AI modification.”

And it doesn’t affect consumers alone: “For creators who want to differentiate themselves from the new synthetic content,” Reisner wrote, “YouTube seems interested in making the job harder.”

After seeing the English-language video dub for one of my favourite Spanish-language comedians, I frantically searched my settings for the option to turn it off, breathing a sigh of relief when I found it.

I immediately closed the app and moved on with my day. Perhaps the one silver lining.