How Much Should You Trust AI Detection Tools?

AI detection tools such as Pangram and Originality.ai claim high accuracy in flagging AI-generated text, but experts like Dartmouth professor Hany Farid warn that these tools are imperfect and face a moving target as generative models evolve. Pangram, founded by Stanford grads in 2023, claims 99.98% accuracy and was rated most effective in a University of Chicago study, yet its CEO Max Spero concedes that 100% accuracy is unattainable. Farid, who founded deepfake-detection company GetReal on the same day ChatGPT launched in November 2022, notes that generative AI is being weaponized, citing a voice-cloning scam that targeted him.

In the space of a historical heartbeat, the world has been inundated with AI slop. Granted, the internet was never a beacon of integrity and trustworthiness. But before ChatGPT launched nearly four years ago, you could at least read an article, watch a video, or listen to a song online and be almost certain a human created it. Not so anymore. We’re living in the era of uncertainty, when half the published text https://graphite.io/five-percent/research/ai-now-writes-as-many-online-articles-as-humans-do on the internet was written by bots, Trump’s a doctor https://www.theguardian.com/us-news/2026/jul/02/trump-ai-video-doctor , and virtually every shocking or outlandish sight on social media sparks the same reaction: Is this real? Dr. Hany Farid has devoted much of his life to answering that question, one AI deepfake at a time. Currently a professor of computer science at Dartmouth, Farid founded a deepfake-detection company called GetReal—by a bizarre coincidence—on the very same day ChatGPT was publicly released in November 2022. GetReal specializes in digital forensics, a relatively young field devoted to identifying the telltale presence of AI the way detectives scour crime scenes for DNA and other signs of human activity. These have become increasingly subtle as generative models have matured, so much so that Farid now mistrusts much of what he sees on a screen. A couple of years ago, he was advising a lawyer on a high-profile case when a scammer used AI to clone his voice, attempting to extract sensitive information from the lawyer—a new form of scamartistry https://gizmodo.com/ai-powered-vishing-attacks-reportedly-targeted-top-hedge-funds-2000794923 known as “voice phishing,” or “vishing.” He and his wife, a neuroscientist and fellow Dartmouth professor, recently agreed upon a safeword to use whenever they speak remotely, to confirm they’re not speaking to a machine imitating their spouse. “Generative AI is growing up very, very quickly,” Farid told me. “And it is being weaponized.” “It’s a moving target” GetReal is one of several young companies aiming to bring transparency, accountability, and trust to a media landscape shrouded in a fog of deepfakes. It focuses primarily on images and videos, but there are also platforms designed to flag AI-generated text—which is arguably the most difficult form of AI-generated content to detect, since the statistical patterns it follows are much simpler than those required to create photorealistic images or videos, for example. Pangram https://www.pangram.com/?gad source=1&gad campaignid=24124423800&gbraid=0AAAABBiSmkh9KI7tWSb9awzCuowkzRIqR&gclid=CjwKCAjw7p UBhBlEiwAhpIs7 EeTQUjqHBZ85e2sSQFm c3YuZCRpCllOgshOLOYmuo5HqPut31-xoCyQ0QAvD BwE is one of the most popular. Founded by two twenty-somethings Stanford grads in 2023, it claims to be able to flag AI-generated text with near-perfect 99.98% accuracy, and an independent study https://bfi.uchicago.edu/wp-content/uploads/2025/09/BFI WP 2025-116.pdf from the University of Chicago comparing multiple AI-detection tools found Pangram to be the most effective. But Max Spero, the company’s CEO and one of its cofounders, told me detection tools like Pangram are doomed to forever play catch-up with generative models themselves. “I don’t think we’ll ever get to one hundred percent accuracy,” Spero told me. The best that the company can do, he adds, is “accumulate more certainty using more data.” Jon Gillham, the CEO of another AI detection tool called Originality.ai http://originality.ai , told Gizmodo something similar. He argued determining whether a piece of text has been generated by AI is analogous to predicting the weather: If an algorithm tells you there’s a hundred percent chance of rain, it’s probably going to rain, and you’d be wise to bring an umbrella if you leave the house. But as we’ve all experienced, such confident weather predictions don’t always pan out. Likewise, says Gillham, the art of detecting AI-generated text is “highly accurate, not perfect… and I’d bet on that continuing to be true in the future.” Like any other kind of statistical forecasting, AI detection tools deal in probabilities, not certainties. Part of the challenge, of course, is that the capabilities of the large language models like ChatGPT, Claude, Gemini, and Grok are constantly evolving as they consume more and more data. Almost as soon as users start noticing certain recurring themes within AI-generated content and complaining about them online, the companies behind them get to work trying to eradicate them. Em dashes are one famous example https://x.com/sama/status/1989193813043069219?s=46 . What are reliable tells of AI on one day become obsolete the next. “It’s a moving target,” Spero told me. At the same time, detection tools also have to contend with so-called “humanizers”—websites designed to rework AI-generated text in order to make it look more convincingly human. Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do — Sam Altman @sama November 14, 2025 https://x.com/sama/status/1989193813043069219?ref src=twsrc%5Etfw The sorts of indicators that humans look for https://gizmodo.com/the-biggest-signs-that-ai-wrote-a-paper-according-to-a-professor-2000634580 when reading a piece of suspicious text are very different from those that a tool like Pangram looks for. “When I’m trying to tell if something’s AI-generated, I’m looking for a few major red flags,” says Spero. “But what Pangram is doing is accumulating hundreds of small, weak signals together. That ability to analyze the text comprehensively—looking at not only the choice of particular words and symbols but their precise placement relative to one another—is what gives detection tools their power. They extract intricate mathematical patterns invisible to the human eye, and not even their builders can look directly under the proverbial hood to understand how this is being done. “Like all AI systems, it remains a little bit of a black box,” Gillham, the Originality.ai CEO, told me. What about watermarks? Of course, such high-tech detection mechanisms probably wouldn’t be necessary if developers were required by law to clearly label AI-generated content as such. At the time of this writing, that’s not the case in the U.S., although some states have begun moving in that direction. The European Union’s AI Act, however, does include a measure https://artificialintelligenceact.eu/article/50/ requiring tech developers to clearly label “text which is published with the purpose of informing the public on matters of public interest” if it “has been artificially generated or manipulated.” The Act also mandates that any AI-generated content, including text, must come with an invisible watermark that can be flagged by a machine. Like AI detection tools, however, watermarking is not at all a silver bullet. Anthropic recently announced https://www.anthropic.com/news/claude-text-watermark that it would begin using watermarks for content generated by Claude to comply with the EU’s AI Act , but developers quickly found ways to bypass them https://www.wired.com/story/coders-say-they-already-found-workarounds-to-claudes-invisible-watermarks/ . Google more recently introduced an option to let users remove the visible label that identifies AI-generated images, video, and music made by the company’s tools though the content will still come with an invisible SynthID watermark and C2PA Content Credentials embedded in its metadata, which means its AI origins can be flagged by a machine . You might also remember the controversy around people figuring out how to splice out the moving watermark on AI-generated videos created by Sora, the controversial app that OpenAI shuttered earlier this year https://gizmodo.com/r-i-p-sora-2024-2026-2000737664 . Still, it’s a start. According to Farid, the Dartmouth computer science and deepfake expert, we shouldn’t dispense with detection tools like watermarks just because they’re flawed. “You lock your door when you leave. Does that prevent a lockpicker from going in? No, but you still lock your door,” he says. “Things that are imperfect are still part of a solution. And then we adapt, and we adapt… that’s cybersecurity in a nutshell.” Another issue with watermarks is that they’re a black-and-white solution in a world of gray. Much of the content one encounters online today has not been fully created either by AI or a human, but rather by a mix of the two—it’s more AI- assisted than AI-generated. Pangram is increasingly focused on trying to determine the extent to which a piece of text has been modified by AI as more humans start leaning on the technology for both writing and editing. As Spero told me: “Watermarks are useful for telling you, Did an AI touch this? But what a watermark can’t tell you is, What was the degree of AI use? That’s the gap that Pangram fills.” Pangram 4 https://www.pangram.com/blog/introducing-pangram-4 , the latest version of the model released last month, is “much more granular” in this regard, he says. The company also recently released an AI-generated image-detection tool, and plans to release another tool for AI-generated videos later this year. Looking ahead Given the rate at which AI is developing and the amount of capital being poured into the technology, there’s every reason to expect that AI-generated sentences, images, videos, songs, and voices are going to be much trickier to detect than they are today. That’s going to lead to new forms of scams, and potentially also to a media landscape so confusing and untrustworthy that our deepfake problems today could look quaint in retrospect. And given the fact that detection tools like Pangram will always be swimming in the wakes of generative models, deepfake detection should be thought of more as a personal skill—to be learned and carefully practiced over time—than as something that can just be completely outsourced to a machine. That includes thinking critically about where information is coming from. Towards the end of our conversation, I asked Farid if there was any advice he liked to give his students about how to stay psychologically afloat in a world flooded by AI slop. “Here’s what you’re going to do,” he told me in the firm but slightly exasperated tone of one who teaches undergrads. “You’re going to stop pretending that you can become a digital forensics expert. What you can do is understand what a trusted source is. Understand that when you read something in Gizmodo, or the New York Times, or the BBC, or CBC, there are serious people behind the scenes who have had these conversations and are fact-checking with editorial standards and oversight, and—they’re not perfect, I understand that—but they are at least trying to bring you reliable information. You cannot say that about Elon Musk and Mark Zuckerberg. They are not trying to bring you trusted, honest information. They are trying to suck every ounce of your soul to keep you addicted to their product, so they can become the next level of trillionaire.” Platforms like GetReal, Pangram, and Originality.ai can help to clear the fog in a world where it’s increasingly difficult to tell the real from the fake. But finding the path forward ultimately comes down to each of us.