Artificial intelligence (AI) content generators are rapidly transforming how we produce written content. In the past few years, advanced language models like ChatGPT, Google's BARD, and Jasper have enabled anyone to generate high-quality text on demand for a wide range of purposes, sparking the desire to check if content contains AI-written text.
In this guide, we'll provide an overview of the leading AI content tools, discuss key indicators to identify AI-generated text, and review emerging attribution standards and detection methods. With AI poised to fundamentally disrupt content creation, it's crucial to understand how to validate if a piece of writing comes from humans, or is directly copy/pasted from an AI generator.
Using AI to kickstart writing is fine, so long as you put in the work to confirm stuff and polish it up. The aim of programs that detect AI isn't to call all computer-written stuff unethical. It's more about spotting AI content that someone was lazy and just copied without credit.
AI detectors help maintain honesty by telling apart stuff that was carefully put together from thoughtless copying. They aren't meant to punish using AI, but to encourage thoughtful use of computer-generated text. With some elbow grease and being upfront, AI material can boost human writing while keeping it real.
* this section was written with AI 😱, and edited by a human.
It's actually quite easy the tell if something was written by AI, especially if you read enough of the outputs that you know were written by it. In fact, LLMs (large language models, what all generative AI is based on in 2023) are essentially statistical machines! They identify the most likely next word or letter based on the patterns they've learned from a massive dataset of text.
For example, we've hire a lot for this company, and it's incredibly apparent when a cover letter is written by ChatGPT and directly copy/pasted without any edits. It's very impersonal and is a bad impression. When you're able to compare an AI created piece of writing with a human written one, especially if they have the same instructions (essay, cover letter, schoolwork, etc...) its VERY easy to tell that AI created it.
ChatGPT is a free AI content writer, which can be used as a rough AI content detector. For example, here's a prompt we use a lot:
Was this content written with Al or by ChatGPT? Consider the structure, wording, and general content before answering.ChatGPT Prompt
From the example below, you can see that ChatGPT notes that it was "written in a way that could be written by AI".
However, this is a relatively imperfect method, as noted in the response itself. Also, the prompt is a bit leading and may get you a false positive. This is a worthwhile read discussing the limitations of OpenAI-derived content detection methods:
These detectors are only partially accurate, as we'll demonstrate in this section.
GPTZero is BY FAR the best AI content detector online today, mainly due to the fact that they raised $3 million from VC and actually run their own backend models.
It's not perfect - here you can see us inputing 100% AI content, but it only returns a 50% probability of being written by AI. However, it won't typically give you a false positive like many others will, meaning that it's one of the more accurate tools online.
The longer the content, the more accurate the output, and we've had great success in identifying AI content that is copied directly out of ChatGPT
Here's another example of an AI detector, which isn't as great, but is 100% free and works sometimes. In this example, however, it's completely wrong. This is a great example of why you shouldn't rely on free AI Detectors, especially ones that use other companies APIs and aren't backed by investment and innovation.
Paid AI detectors fare much better and can typically sniff out direct AI content with ease. Perhaps the best AI detector on the market is AI is called Originality AI which recently updated their detection to a 2.0 model, which works well.
The UI is simple and it will give you the probability that it's AI vs original content. You can also hover over specific sentences and paragraphs to get a more granular look into the AI content within writing.
Despite the accuracy, there are still some false positives, so be mindful of everything.
GPT Zero also has it's own paid plan, where there's a much higher character limit, batch file uploads, 1 million words per month, and a better finetuned model specific for use cases like Education or other things. There's also an API for developers.
Educators typically want to determine if AI was used to plagiarize somebody else's content. Most tools have plans specifically for this use case.
If your institution uses TurnItIn, they've released AI Writing and ChatGPT detection in beta to some users subscribed to specific plans. Many institutions get this for free so it's worth checkin out, but there's no public version of this tool.
This tool/detector can be bypassed as we have seen first hand.
GPTZero offers a discounted plan for educators, with a 50k character limit per document, the ability to upload multiple documents, and an even better detection model.
It's important to stress that AI content is the future, and educators should be learning to implement it in their teaching, not restricting it. By restricting it, you'll just put your students at a disadvantage once they enter the professional world, where AI is everywhere for good reason (save time and money).
Yes, all AI detectors can be beat. It's simple, just heavily edit AI generated content, adding your own information and knowledge to it, and it'll be relatively undetectable.
Or, even better, just write the content yourself!
If you're looking to pass AI detectors at scale, leveraging the efficiencies of AI but still wanting a humanlike, undetectable tone, you can use a tool like HideMyAI, which will also bypass AI detectors. Here's an example with the Originality tool, with the same input used for the first test.
First, we take the output, directly generated by ChatGPT and place it in the HideMyAI AI Humanizer tool. We've selected "English" as the language, "Blog Post" as the type, "Professional - Regular" as the level, "Third Person" perspective, and "ChatGPT" as the generator.
The process takes about 15 seconds and outputs highly readable content that will bypass even the best (paid) detectors (seen below).
You can try it out for free at HideMyAI.
In the past, many detectors made false positives which got a lot of people who wrote legitimate content in deep trouble. Now, detectors are getting better and will typically find AI content, while skewing to the side of caution (ie if they're not 100% sure, they'll let you know).
Even if you try and paraphrase AI content or lightly edit it, detectors will still likely figure out that it was written by AI.
However, there are tools out there that will bypass AI detectors by using AI of their own, HideMyAI, demonstrated in the section above, being one of them.