Writing | Technology
6 Signs Something Was Written By AI
An analysis.
I can spot AI-written articles 80–90% of the time.
There’s no silver bullet that tells you — usually. It’s more a collection of signs.
It’s fluffy and stuck on the surface
GPT-3 uses predictive text, meaning it combs the internet for what usually comes next in a sequence.
Here’s the problem: Most blogs are poorly written. For every quality New York Times piece, there are 50 SEO fluff pieces (1500-word articles that only needed 700 words). Articles about recipes and home improvement are the worst offenders.
For example, I entered a prompt into the AI writer on how to wake up earlier. Here’s what came back:
Wake up on time. This is the most important step to getting your day started on the right foot. The best way to wake up on time is with a morning routine. You can set an alarm for a specific time, but if you want to wake up early, you should create a routine of some kind that will help get you out of bed at the same time each day.
Note the repetition of words and redundant ideas. That entire paragraph could be compressed to “A good morning routine helps you wake up faster and be more productive.”
Daphne Ippolito, a senior researcher at Google Brain, said “If you have enough text, a really easy cue is the word ‘the’ occurs too many times.”
I’ve noticed this as well with demonstrative pronouns (this, that, these, and those). AI writing is littered with them (experienced writers know you can often remove “this” and “that” from sentences).
There are no typos
If AI writers have any virtues, it’s that they rarely misspell and make stupid grammatical mistakes.
They’ll spew complete nonsense ideas— while also passing a spellcheck.
And that makes sense from a design perspective. How can you sell an AI writing program if it generates typos and embarrasses the user?
GTP-3 will, however, use words in the wrong context.
It is loaded with blogging cliches
By design, predictive text often produces phrases that are overused.
I’ve spent my 10,000 hours in the blogging world and can spot blogging cliches fairly easily.
I plucked these common culprits out of the engine’s results:
- “If you’re like me,”
- “But what if there was another way?”
- “Everyone wants to ____”
- “It doesn’t happen overnight”
- “Anyone can live their dream.”
- “There are many benefits to”
- “But for most people”
- “Have you ever wondered___?”
- “There are many ___ out there”
They will often appear multiple times in one post.
There are no segues or context
I do a head tilt when I see something wonky in an article.
If the writer changes the subject and there’s no explanation, it’s a sign.
There are subtleties to human language that are missed with these programs.
They have no heart. There’s this vanilla flavor they develop that’s devoid of emotion and nuance.
It usually lacks metaphors, similes, and other literary devices. And when they are used, it is out of place.
GPT-3 also tends to produce longer sentences. Even if your inputs are short and punchy, the results eventually spiral out like a loose firehose you are turning up.
The voice isn’t personal (nor does it change)
Most programs churn out 3rd and 2nd person voice and it isn’t entirely the program’s fault.
GPT-3 can generate a first-person voice just fine. But most writers don’t put the prompts in with that voice.
Additionally, experienced human writers intentionally break conventional writing rules all the time.
The rules we learn in school aren’t technically hard fast rules — they are meant to guide students to be better writers. Any rule can be broken so long as you have a strategy in mind.
For example, many online writers swap between 1st/2nd/3rd voice because it helps keep the reader’s attention. The alternation is just random enough to keep them guessing, but not enough to make their head spin.
It’s boring
There’s one thing an AI article and false positives have in common: they’re bad.
AI writing programs are a massive time suck if you’re attempting to write something of even mediocre quality. Rather than hitting the generate button and combing through the results, you can accomplish much more with a blank draft screen.
AI articles tend to lack depth and are highly generic, going on and on without getting to the point.
The takeaway: I’m not worried
An AI writing program kills the incentive of being a writer in the first place.
It robs you of the dopamine hit that comes with having your ideas validated by readers.
The good writers you enjoy reading will still stand out amongst the bots.
What I do know: people despise reading AI text without being notified it is AI.
On Quora, they launched a Quora Question Prompt Generator.
It went over like a lead balloon. There was an avalanche of posts attacking the prompt for its lack of authenticity. Quora eventually took it down.
Lastly, it’s often said that writing is editing, and vice-versa. If you don’t have a decent editorial eye, you’ll struggle to make good use of the AI-generated text.
To make a pottery analogy, if you were forced to write with GPT-3 every time, it would be like starting with a quasi-molded block of clay — except that block was extra firm already — and required much more work than a fresh and untouched block.
I’m not worried about AI writers. Ultimately, it’s a form of machine learning.
So we shouldn’t be surprised that they sound like a machine.
Subscribe to my newsletter for more free content.
