Spotting AI-Generated Text: Your Guide for the Internet in 2024
“Experts estimate that as much as 90 percent of online content may be synthetically generated by 2026” — Europol
Have you ever wondered if a computer wrote something you read online?
Chances are, it probably did.
AI-generated text is becoming commonplace, and it’s getting harder to tell the difference between human and AI writing.
As someone who’s been building and testing AI language systems since 2019, I’ve witnessed firsthand how quickly this technology is evolving.
What was cutting-edge 5 years ago is basically ancient technology compared to what we have today. It’s exciting and a bit scary.
But why should you care?
Knowing if AI wrote something will help you scrutinize information, spot biases, and appreciate human insights. It’s quickly become a key part of media literacy — just like spotting fake news.
However, only the most perceptive readers can consistently recognize AI-generated content and it’s not easy.
In this article, I’ll show you:
- How much AI-written text is out there (spoiler: it’s a lot!)
- What AI writing looks like
- Some cool tools that can help spot AI text (useful, but flawed)
- How to train your own “AI detector” brain
This guide is for everyone. Whether you’re a researcher, teacher, or just someone who likes to read online, knowing how to spot AI text can be really helpful.
Remember, AI isn’t all bad. It can be pretty awesome when used right. We just want to be smart about what we’re reading.
Let’s start by looking at just how much AI-written stuff is out there today.
The Growing Prevalence of AI-Generated Content
I recently came across a fascinating graph that shows how AI is changing the way we write.
This graph looks at research papers that use the word “delve” in their title or abstract.
Here’s what’s interesting: from 1990 to 2022, the use of “delve” was minimal. But in 2023 (the year ChatGPT was released) it shot up like a rocket! And in 2024, it’s even higher.
Why does this matter?
“Delve” is a word that AI writing tools really like to use. So this jump almost certainly means that more researchers are using AI to help write their papers.
But it’s not just in research papers. AI is writing all sorts of things:
- News articles
- Social media posts
- Marketing emails
- Product descriptions
- Creative stories & scripts
Some experts think that by 2026, up to 90% of online content could be AI-generated.
This doesn’t mean AI is taking over writing completely. Humans are still in charge, using AI as a tool. But it does mean we need to be more aware of what we’re reading.
What AI Writing Looks Like
What exactly does AI-generated text look like? While it can be hard to tell sometimes, there are a few common characteristics:
Lack of personal voice or style
AI-generated text tends to lack a unique voice or writing style.
It can feel generic and formulaic, without the personal flair that makes human writing engaging.
Overuse of common phrases and structures
AI models are trained on vast amounts of existing writing, which can lead to the repetition of common phrases and sentence structures.
If something feels like you’ve read it before, it might be AI-generated.
Inconsistencies and factual errors
While AI models are improving, they can still make mistakes or generate inconsistencies in their writing.
This could be as simple as mixing up character names in a story or as serious as presenting inaccurate facts in a news article.
Overused words and phrases
AI-generated text, especially from models like ChatGPT, tends to rely heavily on certain words and phrases. Some common examples include:
- Specific Words like: “delve,” “explore,” “discover,” “landscape,” “realm,” “tapestry,” “robust,” “pivotal,” “certainly,” “quinoa,” “orchestrate,” “align,” “diverse,” “recommend,” “note,” “employ,” “bustling,” “treasure,” “paradigm,” “captivate,” “song,” “underscore,” “calculate,” “especially,” “climate,” “hedging,” “inclusive,” “exercise”
- Adverbs and adjectives like: “excitingly,” “amazing,” “remarkable,” “revolutionize the way,” “transformative power,” “groundbreaking advancement”
- Phrases indicating importance: “It is important to note that…,” “It’s important to remember/note…”
- Concluding phrases: “In conclusion…,” “Only time will tell…”
- Phrases emphasizing progress: “Bringing us one step closer…,” “Pushing the boundaries…,” “Paving the way…”
- Other common phrases: “rapidly developing,” “cutting-edge,” “significantly enhances,” “dive deep,” “real of,” “digital landscape/age/realm,” “leverage,” “pivotal,” “adhere,” “amplify,” “Certainly!”, “Absolutely”, “However”, “Ultimately”, “Nevertheless”
Overly flowery language and out-of-place metaphors
AI-generated text often tries to sound sophisticated by using overly descriptive, flowery language and elaborate metaphors.
However, these attempts can come across as awkward or slightly out of place. For example:
- “The sun was a golden chariot racing across the azure tapestry of the sky.”
- “The city was a pulsating heart, its streets the veins through which the lifeblood of commerce flowed.”
- “The computer’s CPU was a maestro, orchestrating a symphony of data in perfect harmony.”
While these metaphors might sound impressive at first glance, they often don’t quite fit the context or feel a bit forced.
Example: Recognizing an AI generated paragraph
Here’s an example of AI-generated text (written by ChatGPT)
Toronto: Canada’s Dynamic Urban Gem
Toronto, Canada’s largest city, is a vibrant metropolis renowned for its cultural diversity and economic prowess. Home to iconic landmarks like the CN Tower and bustling neighborhoods like Chinatown and Little Italy, Toronto offers a rich tapestry of arts, entertainment, and green spaces. With a thriving economy, top-tier cultural institutions, and high quality of life, Toronto stands as a beacon of multiculturalism and urban vitality in North America.
While this text is grammatically correct and includes some relevant facts about Toronto, it feels a bit generic and lacks a strong voice or perspective.
The overuse of flowery language, metaphor related to music and terms like “rich tapestry” and “vibrant” are usually a dead giveaway that something is AI generated. (Also anything like “delve” or “deep dive” for some reason)
Tools to Detect AI-Generated Text
While no tool is perfect, there are some decent AI content detectors out there. These look at your text paragraph by paragraph and tells you if it thinks AI wrote it.
Notable AI Content Detectors
- Scribbr’s Premium AI Detector: This paid tool claims to identify 84% of AI-generated, human-written, and paraphrased texts in tests.
- QuillBot AI Detector (Free): This free tool claims 78% accuracy.
- Scribbr’s Free AI Detector: This one too.
- Originality.AI: This one got 76% accuracy in tests.
- Copyleaks AI Detector: Claims over 99% accuracy in spotting AI-generated text, even if it’s been paraphrased or plagiarized.
- TraceGPT: Claims to be one of the few tools that can accurately detect text written by the Claude language model, which stumps many other detectors.
^While some of these claims may hold up with specific tests, these are mostly just marketing ploys. In actual testing and I found dozens of incorrect classifications using all of these tools.
Limitations of AI Detectors
While these tools can be helpful, they’re not perfect:
- False positives: Sometimes, they flag human-written text as AI-generated by mistake. I found this to be quite common across all detectors.
- Trouble with edited AI text: If someone manually edits or paraphrases AI-generated content to make it sound more human-like, the detectors are more likely to miss it.
- Different results for different AI models: Detectors trained mainly on older GPT models might not work as well with newer models like Claude 3, GPT-4o, Gemini, etc.
Here’s the thing: it’s impossible to be 100% sure if text is AI-generated or not, and all these tools can give false positives.
So, use AI detectors as a helpful guide, but don’t treat them as the ultimate truth. It’s best to use them along with other clues and your own critical thinking.
How to Train Your Own “AI Detector” Brain
You’ve seen how AI-generated writing is becoming more common, learned some of its characteristics, and looked at tools that can help detect it.
But here’s the thing: AI is getting better every day and beating detectors.
What can you do?
Train your own “AI detector” brain!
Here’s how:
- Get familiar with AI writing: The best way to spot AI-generated content is to know what it looks like. Play around with tools like ChatGPT and Claude to get a sense of how they write. Pay attention to the common characteristics we talked about. Look for overuse of certain words and phrases, lack of personal voice, and weird metaphors.
- Read critically: Whenever you’re reading something online, ask yourself, “Could an AI have written this?” Look for those telltale signs of AI-generated content, but also trust your gut. If something feels a bit off or too generic, it might be AI-written.
- Use detection tools, but don’t rely on them: AI content detectors like Scribbr and QuillBot can be helpful, but remember, they’re not perfect. Use them as a starting point.
- Stay up-to-date: AI technology is advancing faster than ever,what we know about AI writing today might not be true year form now. Keep learning about the latest developments in AI language models and how they’re being used to generate content.
- Embrace the uncertainty: There may come a day when we simply can’t tell the difference anymore. And that’s okay! The key is to approach all online content with a healthy dose of skepticism and to always think critically about what you’re reading.
In the end, the rise of AI-generated content doesn’t have to be a scary thing. By understanding how it works, staying aware of its presence, and honing your own “AI detection” skills, you can be prepared.
And who knows? You might even find yourself appreciating the amazing capabilities of AI language models — just remember that human insight and creativity will always have a vital role to play.
