How to Spot (and Stop) AI in the Wild

#Technology Read time:
We know two things for sure: AI is everywhere, and AI detectors aren’t great at their one job.

What’s a human to do?

Learn the ways of AI identification and take a good, strong look at our own media literacy.

 

Pictures too perfect

The scourge of social media feeds: Misleading and downright silly AI generated images. Often, these are enjoyed and shared without a shred of suspicion. Use three “C’s” to catch clues that an image has been artificially created and doesn’t represent reality.

Colors: Pay attention to the palette, the hues, and the saturation of colors. This can be super subtle (especially for folks who don’t work with art or illustration), but something is just a bit off with the colors in an image. Too harmonious or bright? Too sharp in the foreground and blurred in the background?

Contrast: The lighting and shadows are often too harsh in contrast. The bright spots are super bright, while the shadows are quite dark, and the light source may be from an unexpected (but oh-so-pleasant!) angle. Think of where the lamp or sunlight would have to originate from and compare it to where the shadows fall.

Cartoonish: Too-round and too smooth. As a professional photographer points out, most faces are not symmetrical. Having nice, even, aesthetically pleasing features sounds good to a robot, but human flaws are what make us, well, human.

 

Words too wordy

Now, your girl doesn’t use generative AI to write. No, this is 100% from my brain (and it honestly bugs me a lot talking in first person on my blog so I’m going to stop now). But there are plenty of clues someone has used a program like ChatGPT to create their text (I’m not going to call it writing. Even if there’s a fire.).

The first clue is that it’s longer than it needs to be. Generative AI takes the five-paragraph form incredibly seriously and will shoehorn in an extra six 50-cent words just for funsies. Some other super common gen AI markers:

Bullets: One of gen AI’s strengths is summarizing things. When a person begins to write in exclusively bullet points, that’s a clue they’re actually running thoughts through AI to extract their insight.

Rule of threes: The rule of threes refers to the tidy way our brains enjoy reading lists. We’re not a huge fan of one or two points. We love seeing a list of three things.

I took the bus and swam in the pool.

I took the bus, swam in the pool, and enjoyed the sunshine.

Guess who else loves listing three examples? It’s one of those really obvious outcomes of teaching a robot to be a word calculator. If we love something, it will too. The difference lies in making a stylistic choice to use the rule of threes versus peppering triads of examples across an already lengthy piece of text.

Transitional phrases: “Firstly,” “secondly,” “thirdly,” and “however,” “moreover,” and “furthermore,” especially when added to EACH paragraph at the beginning, quickly add up to overkill. A human writer is more likely to either use deliberate repetition to create cadence or vary the beginning of their paragraphs. The size of the paragraphs usually varies when written by humans as well.

Em dashes: This one is really getting under writers’ skins, because writers love to use a good em dash—nothing makes them happier, in fact. But human writers know that they can have a little em dash as a treat, so if you see these meaty punctuation marks in pairs or triplicate, you’re probably reading a text from a large language model (LLM).

“Worth noting”: Boy howdy is stuff worth noting to generative AI. This phrase might as well be a bullseye on a generated piece of text.

 

Videos: Probably the most insidious of all sneaky AI

By now your juicy human brain is solidifying the patterns generative AI leaves in its wake as it creates an approximation of art or writing. Less common (for now) but probably most apocalyptic is the advent of artificially generated videos of real people, also called deepfakes. How on earth do we differentiate between what’s real and what’s not if it appears to be a moving, breathing human?

Here are a couple clues:
Choppiness: Video editing does leave some clues, so look twice at videos that appear to glitch, skip, or jump from frame to frame.

Jerky movements: Your brain may subconsciously recognize robotic or uncanny movements, leading to the gut feeling that something’s not quite right. If something feels off with how someone moves, talks, blinks, or fidgets, take notice.

Lighting and shadows: Similar to images, identify the light source and see if the shadows fall where they should. Does the light source appear to come from multiple angles? Is someone’s face lighted differently than their body parts?

Multiple ratios and sizes in the same frame: Look for unnatural or heavy-handed blurring or a strange depth of field a camera cannot replicate.

Too bright and too cartoony: Consider the way colors born on a screen compare to those we see first with our eyes and capture with our tools. They are totally different. Life isn’t a cartoon, so videos of real life shouldn’t feel like a cartoon.
 

Qs to ask

As for stopping the spread of misinformation, even “harmless” things like fake cartoony pictures of a supposed sunken statue or astonishing fiber art, what can you do when faced with whether or not something is generated by artificial intelligence? Ask yourself these three questions:

1. Do I WANT this to be real? (Probably the most important question of all.)
2. What is my bias? What is the post’s bias?
3. Can other sources corroborate?


You’re in control of your algorithms. Reject AI slop and embrace humanity.
Share this story: