{"slug": "ai-slop-detector-excels-at-text-but-falters-on-images", "title": "AI Slop Detector Excels at Text but Falters on Images", "summary": "Pangram, a $20-per-month web tool, correctly identified AI-generated text in dozens of tests by a technology journalist, including flagging AI sentences inside human essays, but it and Hive Detect failed on several widely shared AI images, such as a deepfake of Zendaya and a fake photo of Sen. Mitch McConnell. Pangram claims its technology correctly identifies 9,999 out of 10,000 text samples, and an independent University of Chicago study found near-perfect performance, but image detection remains flawed, with Pangram flagging only 41 of 43 AI images in its own tests and missing two of five in a Dartmouth professor's experiment.", "body_md": "**August 16, 2026**, (Inside AI) — A new fatigue is spreading through social media feeds: slop-induced burnout. Posts riddled with bullet points, emoji, and em dashes increasingly read like chatbot output. A web tool called **Pangram** claims to detect exactly that, and early testing suggests it is remarkably good at spotting AI-generated text.\n\nPangram costs **$20** a month and lets users paste text to receive a percentage score indicating how much was bot-written. In dozens of tests by a technology journalist, the detector never failed to distinguish AI prose from human writing. The tool correctly identified AI sentences sandwiched inside personal essays and flagged human writing sprinkled within AI paragraphs. It even judged passages from **Charles Dickens** as **100%** human.\n\nThe stakes are rising. Some studies say **50%** of online articles are now artificially generated. Accurate detectors could become as essential as antivirus software. Pangram says its technology correctly identifies **9,999** out of **10,000** text samples, and an independent study by the **University of Chicago** found near-perfect performance. That is a sharp contrast to earlier AI text scanners, which made glaring errors including misidentifying famous authors.\n\n## How Pangram Reverse Engineers a Chatbot's Fingerprint\n\nDetecting AI writing is not as simple as looking for em dashes and bullet points. Every AI chatbot, such as **Claude**, **ChatGPT**, and **Gemini**, follows a decision tree where each word choice leads to another. Accurate detection requires understanding how each model's decision tree works, which means gathering enormous data on each model.\n\n**“We're reverse engineering their stylistic fingerprint,”** **Max Spero, a founder of Pangram**\n\nFor AI-generated images, the task is harder. Telltale signs include rough pixel textures, oversaturated colors, and inconsistent lighting. Some companies behind image generators, including **OpenAI**, **Anthropic**, and **Google**, embed invisible watermarks in metadata. Anthropic announced this week that it plans to embed invisible watermarks in Claude's writing, a response to European Union transparency rules.\n\n## Image Detectors Stumble on Widely Shared Fakes\n\nThe journalist tested **20** AI-generated images that had been widely shared and debunked. Pangram and **Hive Detect**, a similar tool, both failed quickly. Hive incorrectly identified eight AI images as real, including a deepfake of actress **Zendaya** appearing pregnant, a photo of **Sen. Mitch McConnell** on a hospital bed, and a fake San Francisco street sign about theft under **$950**.\n\nPangram incorrectly identified two AI images, including the McConnell photo and the phony sign. It also declined to scan four images that looked violent or were too low quality. Both detectors correctly flagged some famous fakes, such as a wedding photo of Zendaya and **Tom Holland**, a photo of **President Donald Trump** holding a girl in China, and a photo of the **Clintons** with **Jeffrey Epstein**.\n\nHive said some failures may have occurred because the journalist scanned copies shared on social media that lacked original details. That rationale is concerning. By the time most people see AI slop online, the image is often a screenshot or automatically compressed by the platform. If that is enough to thwart detectors, the technology is unhelpful in real-world conditions.\n\nSpero said he was surprised Pangram failed to flag the McConnell image but noted that AI image detection is a new, unfinished feature. In Pangram's tests of **43** AI images, it correctly identified **41**. **Hany Farid**, a Dartmouth professor and founder of **GetReal Security**, ran his own quick experiment. He uploaded five AI-generated wartime photos, and Pangram flagged only three.\n\nFarid said distinguishing fake and real photos is extremely hard because images can be distorted and manipulated in many ways. Visual AI detectors are generally still too flawed. For now, he recommends relying on trusted media outlets.\n\n**“Stop getting your news from social media,”** **Hany Farid, Dartmouth professor and founder of GetReal Security**\n\n**“An assumption that most of what you are seeing is fake is probably pretty good right now.”** **Hany Farid, Dartmouth professor and founder of GetReal Security**\n\nDespite image detection gaps, Pangram's text proficiency makes it useful for sniffing out AI-written email scams, phony online reviews, and uninteresting LinkedIn posts. The journalist scanned a handful of LinkedIn posts that appeared machine-written. Pangram revealed that one tech worker wrote only the introductory sentence and used AI for the rest. The worker sheepishly confirmed it.", "url": "https://wpnews.pro/news/ai-slop-detector-excels-at-text-but-falters-on-images", "canonical_source": "https://insideai.news/news/ai-tools/ai-slop-detector-excels-at-text-but-falters-on-images/7910/", "published_at": "2026-08-16 05:06:06+00:00", "updated_at": "2026-08-16 05:10:52.334610+00:00", "lang": "en", "topics": ["artificial-intelligence", "generative-ai", "ai-tools", "ai-ethics"], "entities": ["Pangram", "Max Spero", "University of Chicago", "Hive Detect", "Zendaya", "Mitch McConnell", "Hany Farid", "GetReal Security"], "alternates": {"html": "https://wpnews.pro/news/ai-slop-detector-excels-at-text-but-falters-on-images", "markdown": "https://wpnews.pro/news/ai-slop-detector-excels-at-text-but-falters-on-images.md", "text": "https://wpnews.pro/news/ai-slop-detector-excels-at-text-but-falters-on-images.txt", "jsonld": "https://wpnews.pro/news/ai-slop-detector-excels-at-text-but-falters-on-images.jsonld"}}