<p>How can a computer program like GPTZero can tell if a text was AI generated or now. These class of softwares are guided by 2 ratios:</p><ul><li><strong>perplexity</strong>: a function of "how surprising" is this language given the training data. Given the phrase <em>"I want a cup of _". </em>Continuing with a word like "<em>tea</em>" means low perplexity. "a cup of spiders" however is more surprising, and thus human. </li><li><strong>burstiness:</strong> evaluates the variability in sentence length and structure throughout a text. AI tends to be more omogenous aka boring and rapid changes in sentences length, the use of adjectives here but not there are more human.</li></ul>

The tech for detecting AI generated text

<p>The US constitution and most parts of the Bible fail the test and most tools think they were AI generated. It's because these texts are so fundamental that they are overly-present in the training data. Not just that:</p><ul><li>Most formal writing appears AI generated.</li><li>Text created by foreign writers also appears AI generated because most of them tend to be more calculated and formal in style.</li></ul><p>Given the training data challenges and potential discrimination implications, it may be the case AI detection will never gain a lot of support. </p>

Why does the US Constitution seem to be AI generated

<figure><img src="https://cdn.arstechnica.net/wp-content/uploads/2023/05/madison_chatgpt_hero_1-800x450.jpg" alt="An AI generated image of James Madison writing the U.S. Constitution using AI." width="725" height="408"><figcaption class="caption" dir="ltr"><div class="caption-text" dir="ltr"><a href="https://cdn.arstechnica.net/wp-content/uploads/2023/05/madison_chatgpt_hero_1.jpg" class="enlarge-link" data-height="675" data-width="1200" data-uri="bdfb3a696e893700359f70c08d74cad9" dir="ltr">Enlarge</a> <span class="sep" dir="ltr">/</span> An AI-generated image of James Madison writing the US Constitution using AI.</div><div class="caption-credit" dir="ltr">Midjourney / Benj Edwards</div></figcaption></figure><p dir="ltr">If you feed America's most important legal document—the <a href="https://www.archives.gov/founding-docs/constitution-transcript" data-uri="97d2feb715fd8da6d3958fb028685a9a" dir="ltr">US Constitution</a>—into a tool&nbsp;designed to detect text written by AI models like <a href="https://arstechnica.com/information-technology/2022/12/openai-invites-everyone-to-test-new-ai-powered-chatbot-with-amusing-results/" data-uri="9756a9c78da6e1b8e379a8d2bb2f4e1f" dir="ltr">ChatGPT</a>, it will tell you that the document was almost certainly written by AI. But unless James Madison was a time traveler, that can't be the case. Why do AI writing detection tools give false positives? We spoke to several experts—and the creator of AI writing detector GPTZero—to find out.</p><p dir="ltr">Among news stories of <a href="https://www.washingtonpost.com/technology/2023/05/18/texas-professor-threatened-fail-class-chatgpt-cheating/" data-uri="c89af3700ac19b084c6e007ed6b69a65" dir="ltr">overzealous professors</a> flunking an entire class due to the suspicion of AI writing tool use and kids <a href="https://www.reddit.com/r/ChatGPT/comments/132ikw3/teacher_accused_me_of_using_chatgpt/" data-uri="1ca3f663db0465a645e8eec1a9091b12" dir="ltr">falsely accused</a> of using ChatGPT, generative AI has education in a tizzy. Some think it represents an <a href="https://www.theatlantic.com/technology/archive/2022/12/chatgpt-ai-writing-college-student-essays/672371/" data-uri="ec97fe51204a227885f2299175c0a97e" dir="ltr">existential crisis</a>. Teachers relying on educational methods developed over the past century have been scrambling for ways to <a href="https://www.reddit.com/r/Teachers/comments/zkguxg/my_frustrations_with_chatgpt/" data-uri="29b831b33ec5be19685430dbe4ddd170" dir="ltr">keep</a> the status quo—the tradition of relying on the essay as a tool to gauge student mastery of a topic.</p><p dir="ltr">As tempting as it is to rely on AI tools to detect AI-generated writing, evidence so far has shown that they are <a href="https://techcrunch.com/2023/02/16/most-sites-claiming-to-catch-ai-written-text-fail-spectacularly/" data-uri="3a69ca3a663f81741d0e4ac39bc9894d" dir="ltr">not reliable</a>. Due to false positives, AI writing detectors such as <a href="https://gptzero.me/" data-uri="3c781528a1c5e68b61f26160d0240d8f" dir="ltr">GPTZero</a>, <a href="https://www.zerogpt.com/" data-uri="17a65dbc198c1a4dc372c934a19f8f8e" dir="ltr">ZeroGPT</a>, and OpenAI's <a href="https://platform.openai.com/ai-text-classifier" data-uri="003b2c25ca3419df58292f0c628826ee" dir="ltr">Text Classifier</a> <a href="https://theconversation.com/we-pitted-chatgpt-against-tools-for-detecting-ai-written-text-and-the-results-are-troubling-199774" data-uri="c8018a5eeea8cabe588c12481c8ed13b" dir="ltr">cannot</a> be trusted to detect text composed by large language models (LLMs) like ChatGPT.</p><UL><LI><figcaption dir="ltr">
                                  <div class="caption" dir="ltr">
                    A viral screenshot from April 2023 showing GPTZero saying, "Your text is likely to be written entirely by AI" when fed part of the US Constitution.                  </div>
                                                  </figcaption><div dir="ltr">
                    
                                          Ars Technica                                      </div></LI><LI><figcaption dir="ltr">
                                  <div class="caption" dir="ltr">
                    When fed part of the US Constitution, ZeroGPT says, "Your text is AI/GPT Generated."                  </div>
                                                  </figcaption><div dir="ltr">
                    
                                          Ars Technica                                      </div></LI><LI><figcaption dir="ltr">
                                  <div class="caption" dir="ltr">
                    When fed part of the US Constitution, OpenAI's Text Classifier says, "The classifier considers the text to be unclear if it is AI-generated."                  </div>
                                                  </figcaption><div dir="ltr">
                    
                                          Ars Technica                                      </div></LI></UL><p dir="ltr">If you feed GPTZero a section of the US Constitution, it says the text is "likely to be written entirely by AI." Several times over the past six months, screenshots of other AI detectors showing similar results have <a href="https://twitter.com/0xgaut/status/1648383977139363841?s=20" data-uri="9ae25ee202fc2676b99fbee0dad326b7" dir="ltr">gone viral</a> on social media, inspiring confusion and plenty of jokes about the founding fathers being robots. It turns out the same thing happens with selections from The Bible, which also show up as being AI-generated.</p><p>To explain why these tools make such obvious mistakes (and otherwise often return false positives), we first need to understand how they work.</p><h2>Understanding the concepts behind AI detection</h2><p>Different AI writing detectors use slightly different methods of detection but with a similar premise: There's an AI model that has been trained on a large body of text (consisting of millions of writing examples) and a set of surmised rules that determine whether the writing is more likely to be human- or AI-generated.</p><p dir="ltr">For example, at the heart of GPTZero is a neural network trained on "a large, diverse corpus of human-written and AI-generated text, with a focus on English prose," according to the service's <a href="https://gptzero.me/faq" data-uri="962a8e3069f8d7b881f0613cfd08648c" dir="ltr">FAQ</a>. Next, the system uses properties like "perplexity" and burstiness" to evaluate the text and make its classification.</p><figure><img src="https://cdn.arstechnica.net/wp-content/uploads/2023/07/we_the_people-640x425.jpg" width="725" height="481" srcset="https://cdn.arstechnica.net/wp-content/uploads/2023/07/we_the_people.jpg 2x"><figcaption class="caption" dir="ltr"><div class="caption-text" dir="ltr"><a href="https://cdn.arstechnica.net/wp-content/uploads/2023/07/we_the_people.jpg" class="enlarge-link" data-height="797" data-width="1200" data-uri="a190b21559b5fff0d568facf939df63d" dir="ltr">Enlarge</a></div><div class="caption-credit" dir="ltr">Bonnie Jacobs / Getty Images</div></figcaption></figure><p>In machine learning, perplexity is a measurement of how much a piece of text deviates from what an AI model has learned during its training. As Dr. Margaret Mitchell of AI company Hugging Face told Ars, "Perplexity is a function of 'how surprising is this language based on what I've seen?'"</p><p>So the thinking behind measuring perplexity is that when they're writing text, AI models like ChatGPT will naturally reach for what they know best, which comes from their training data. The closer the output is to the training data, the lower the perplexity rating. Humans are much more chaotic writers—or at least that's the theory—but humans can write with low perplexity, too, especially when imitating a formal style used in law or certain types of academic writing. Also, many of the phrases we use are surprisingly common.</p><p>Let's say we're guessing the next word in the phrase "I'd like a cup of _____." Most people would fill in the blank with "water," "coffee," or "tea." A language model trained on a lot of English text would do the same because those phrases occur frequently in English writing. The perplexity of any of those three results would be quite low because the prediction is fairly certain.</p>

Can AI writing detectors be trusted? We dig into the theory behind them.

cybersecurity

Cybersecurity

computerscience

Computer Science

artificialintelligence

Artificial Intelligence

<p>{
    "Time": 15.902237892150879,
    "Prompt tokens": 1024,
    "Completion tokens": 442,
    "Total tokens": 1466,
    "Config": {
        "model": "gpt-3.5-turbo",
        "max_tokens": 1000
    }
}</p>

AI writing detectors cannot be trusted to detect text composed by large language models (LLMs) like ChatGPT

<p>The article discusses how AI writing detectors, such as GPTZero and OpenAI's Text Classifier, give false positives when identifying text composed by large language models (LLMs) like ChatGPT. The author highlights that relying on these tools to detect AI-generated writing is not reliable.</p>

AI writing detectors mistakenly classify the US Constitution and the Bible as AI-generated

<p>The article mentions that when fed sections of the US Constitution and the Bible, AI writing detectors like GPTZero and ZeroGPT falsely classify the text as AI/GPT generated or likely to be written entirely by AI. This leads to confusion and jokes about the founding fathers and the Bible being created by AI.</p>

AI writing detectors use methods like measuring perplexity to evaluate the text

<p>The author explains that AI writing detectors use methods like measuring perplexity to evaluate the text. Perplexity is a measurement of how much a piece of text deviates from what an AI model has learned during training. The closer the output is to the training data, the lower the perplexity rating.</p>

AI models like ChatGPT rely on their training data when generating text

<p>The article describes how AI models like ChatGPT tend to reach for what they know best when generating text, which comes from their training data. The output of AI models with similar content to the training data would have lower perplexity ratings, while humans can also write with low perplexity, especially when imitating formal writing styles.</p>

Educators are concerned about the impact of AI-generated writing on traditional educational methods

<p>The article mentions that educators are concerned about the impact of AI-generated writing on traditional educational methods. Some worry that relying on AI tools for writing detection represents an existential crisis for the traditional use of essays as a method to gauge student understanding.</p>

AI writing detectors have sparked viral screenshots and confusion on social media

<p>The article highlights that screenshots of AI writing detectors classifying sections of the US Constitution and the Bible as AI-generated have gone viral on social media, leading to confusion and jokes about the founding fathers being robots. The article also mentions similar false positives with other selections of text.</p>

How AI detectors work

The tech for detecting AI generated text

Why does the US Constitution seem to be AI generated

Similar ideas