Teachers have long needed to be vigilant about the originality of their students' work. Yet it's fair to say that at no time in history has it been more difficult to do so than now. Not only does the internet itself provide a vast arena for potential plagiarism, but also the recent widespread availability of generative AI chatbots offers even more temptation to students looking for a shortcut.
Fortunately for educators, digital AI detection tools have quickly proliferated in response to the new AI-enabled text generators.
I tested the performance of 13 free AI detection websites, with some surprising results. Many were completely fooled by the ChatGPT text, indicating they haven't kept up with its well-publicized advances. And fewer than half were able to correctly—and with certainty—identify all four trial texts. Caveat emptor? Yep. Your time is worth it.
How We Tested
- Text #1: Generated by ChatGPT from the prompt "Write a 500-word essay about causes of the Great Depression."
- Text #2: Generated by BARD from the prompt "Describe in 500 words the causes of the American Revolutionary War."
- Text #3: This article by Tech & Learning's Erik Ofgang.
- Text #4: This article by The New York Times columnist Maureen Dowd.
The full text or article was used in each trial, unless the platform restricts the number of words or characters. In that case, we entered the maximum allowable text, starting from the beginning.
BEST FREE AI DETECTION SITES
PROS: Fast, easy to use.
CONS: Flubbed ChatGPT text. Maximum 400 words
A free service provided by the nonprofits Quill.org and Commonlit, AI Writing check was unfortunately fooled entirely by the Chat GPT text. Users can do better.
PROS: Fast, easy to use. Also offers AI-generated image detection.
CONS: Not very accurate
Content At Scale's AI Content Dectector provides a sentence-level assessment of likelihood of being AI-generated: Your text is color-coded as "Highly likely," "Possibly," or "Unclear." While this granularity may be appealing, it has to be accurate to be useful.
PROS: Very accurate, fast and easy to use.
CONS: None noted
Copyleaks' free basic AI Content Detector performed perfectly in our tests, flagging all AI content while recognizing the human-created text as such. To gain enhanced functions, such as multiple languages, and more frequent scans, sign up for a free account.
PROS: Highly accurate, up to 3,000 words analyzed.
CONS: Free account setup required after four uses.
Crossplag provides a percentage score representing the likelihood that a given text is AI generated, from 0% to 100%. It's easy to use, and correctly ID'd all four of our testing texts.
PROS: Provides high-level background information on the technology of generating fake text
CONS: Results are difficult to interpret.
Think of the Giant Language Model Test Room more as a lesson in principles of AI-generated text than an a practical AI detection tool. But if you or your students want to learn about probability histograms, top k values, and word choice predictions, this site could be very educational.
PROS Fast. Detailed results without account setup.
CONS Equivocal response for Test Text #2. Account required after several uses.
Created in response to ChatGPT by a Princeton senior, GPTZero's algorithm is based on "perplexity" and "burstiness," two qualities that reflect the randomness and complexity of writing as interpreted by GPTZero. More random + more complex = human generated.
PROS: Fast, easy to use.
CONS: Flubbed ChatGPT text. Maximum 510 words.
The Hugging Face GPT-2 Output Detector Demo failed the first set of tests roundly, stalling on every text in three different browsers. Given an opportunity several days later, it performed much better, correctly identifying three of the four testing texts. However, given that the current free ChatGPT version is based on the GPT-3.5 architecture, Hugging Face is likely to whiff on any new ChatGPT content. Look elsewhere.
PROS Free, fast.
CONS Not very accurate.
The AI Text classifier website is upfront about the limitations of the platform, saying it's "not always accurate" and that AI-generated text can easily be edited to evade detection. Of course, no platform is 100%. But its inability to recognize Text 1 as AI (generated by ChatGPT, Open AI's latest product) renders it of dubious value.
PROS: Fairly accurate, no credit card required to sign up, simple pricing.
CONS: Account signup required. Partial miss on text 3.
Sign up and get 50 free credits to get started, after which the cost is a penny per 100 words checked. Originally, the AI had no trouble detecting both AI-generated texts as such. And it correctly tagged text 3 from our own Erik Ofgang as 99% likely human origin. Yet somehow it declared testing text 4, from longtime New York Times columnist Maureen Dowd, to be only 58% likely original. I doubt it.
PROS: Highly accurate, fast, easy to use, no account needed
CONS: None noted
Undetectable AI's main purpose is editing AI-flagged content into writing that's indistinguishable from human-written text, for a price. But its stand-alone AI detection system is free and accurate, correctly evaluating all four texting texts. Bonus: Undetectable AI also shows you how multiple AI detectors view your text.
PROS: Fairly accurate
CONS: Account is required to use. Free account limited to 2,000 words monthly
If you have fewer than 2,000 words to check, Winston AI is a great choice, flagging both AI testing texts as not human while generally recognizing the human-created texts as such. For $12 per month, users can submit up to 80,000 words, get printable reports, and upload documents or photos of documents for optical character recognition (OCR).
PROS: Fast, no account required, clean interface
CONS: Not very accurate. 1,500 character maximum for one check.
Not only did Writer AI declare ChatGPT's AI-written text to be "98% Human Generated" but it also commented "Fantastic!" as if to encourage users to create and submit more AI-generated text. Plus the 1,500-character limit is limiting. There are better options.
PROS: Very accurate, fast, no account required, clean interface. Text suspected to be AI-generated is conveniently highlighted.
CONS: None noted.
Didn't we already cover this one? Nope, that was the similarly-named GPTZero. Zero GPT performed nearly flawlessly in our tests, with the added benefit of highlighting text suspected to be AI-generated. Easy to use and reliable.
Detailed Table of Results
|AI Detection Platform||Text 1: Chat GPT||Text 2: BARD||Text 3: Ofgang||Text 4: Dowd||Comment|
|AI Writing Check||Text Written by Human||Text Written by AI||Text Written by Human||Text Written by Human||Maximum 400 words|
|Content at Scale AI Content Detector||Flagged 1 sentence as AI, 7 as possibly and 3 as unclear. 16 sentences not flagged||Flagged 4 sentences as AI, 6 possible, 2 unclear. Multiple not flagged.||One sentence flagged as "unclear."||2 possible, 2 unclear||Up to 25,000 characters will be used|
|Copyleaks||99.9 % probability AI; all sentences flagged||99.9 % probability AI; all sentences flagged||This is human text||This is human text||Row 2 - Cell 5|
|Crossplag||100% AI: This text is mainly written by an AI.||100% AI: This text is mainly written by an AI.||0% AI: This text is mainly written by a human.||0% AI: This text is mainly written by a human.||3000 word maximum/Create free account after four checks.|
|Giant Language Model Test Room||543/112/37||445/56/11||747/167/62/24||505/133/71/28||Top k frequent elements, most predicatable on left|
|GPTZero||Likely to be written entirely by AI||May include parts written by AI||Likely to be written entirely by a human||Likely to be written entirely by a human||Row 5 - Cell 5|
|Hugging Face||97.14% real||99.98% fake||99.98% real||99.98% real||Row 6 - Cell 5|
|Open AI Text Classifier||Very unlikely AI-generated.||Possibly AI-generated.||Very unlikely AI-generated.||Very unlikely AI-generated.||Row 7 - Cell 5|
|Originality AI||100% AI- entire text highlighted as 100% confidence||100% AI- entire text highlighted as 100% confidence||99% orig, 1% AI||58% orig/42% AI||Row 8 - Cell 5|
|UNDETECTABLE.AI||Your content is detected as written by AI||Your content is detected as written by AI||Your content appears human||Your content appears human||Row 9 - Cell 5|
|Winston A.I||0% human||0% human||73% human||100% human||Row 10 - Cell 5|
|Writer AI||98% human-generated Content||13% human-generated Content||100% human-generated Content||100% human-generated Content||Row 11 - Cell 5|
|ZeroGPT||94.68% AI/GPT Generated||86.54% AI/GPT Generated||Human written (3.93% AI GPT)||Human written (5.83% AI GPT)||Row 12 - Cell 5|
- Best Free Plagiarism Checking Sites
- Best Sites & Apps for K-12 Education Games
- ChatGPT Plus vs. Google’s Bard
To share your feedback and ideas on this article, consider joining our Tech & Learning online community here.