2 weeks ago
AI Slop Detector Excels at Text, Struggles With Images
Computers can now write stories and draw pictures that look like a person made them.
Sometimes these robot-made things fill the internet with low-quality posts called 'slop.'
A new tool called Pangram works like a detective that sniffs out robot writing.
You paste in some words, and it tells you how much was probably written by a person or by a computer.
A reporter tested Pangram for a week with lots of tricky examples, and it almost never made a mistake with text.
The tool learns each robot's special style of choosing words, kind of like recognizing a friend's handwriting.
But Pangram is not very good at spotting robot-made pictures — in tests, it and another tool called Hive Detect missed several fake photos, including some showing famous people.
Some companies hide tiny 'watermarks' in robot-made images to help detectors, and one company, Anthropic, plans to do the same for writing.
Experts say the safest thing is to get your news from trusted places and remember that a lot of what you see online might not be real.
A week of hands-on testing found Pangram, a $20-a-month AI detector, was near-perfect at distinguishing AI-generated text from human writing.
Pangram claims its text detection is accurate 9,999 times out of 10,000, with a University of Chicago study also calling the tool near-perfect on text.
The tool works by reverse-engineering the 'stylistic fingerprint' of AI models such as ChatGPT, Claude, and Gemini.
AI image detection lagged: Pangram misidentified 2 of 20 viral AI images and Hive Detect misidentified 8, though both caught several famous fakes.
Anthropic announced it will embed invisible watermarks in Claude's writing in response to EU transparency regulations, and experts urge getting news from trusted media outlets.
- Who
- The story centers on Pangram, an AI detector founded by Max Spero, which was reviewed by a reporter and compared with rival tool Hive Detect; Dartmouth professor Hany Farid also tested it.
- What
- A review of Pangram found it near-perfect at detecting AI-generated text but unreliable at flagging AI-generated images.
- Where
- Online, using a web tool and posts from LinkedIn and Reddit; one debunked image test involved a street sign in San Francisco.
- When
- This month, during a roughly one-week testing period.
- Why
- Because AI-generated 'slop' is proliferating online — some studies estimate 50% of online articles are AI-generated — creating demand for detection tools.
Key facts
- Product
- Pangram — an AI 'slop' detector web tool
- Price
- $20 per month
- Text detection accuracy
- 9,999 of 10,000 correct per Pangram; University of Chicago study called it near-perfect on text
- Image detection test
- Pangram misidentified 2 of 20 viral AI images and Hive Detect 8, including a Zendaya deepfake and an image of Sen. Mitch McConnell; both correctly flagged a fake Zendaya-Tom Holland wedding photo, a photo of President Donald Trump in China, and an image of the Clintons with Jeffrey Epstein
- Pangram's own image test
- 41 of 43 AI images correctly identified
- AI content online
- Some studies say about 50% of online articles are artificially generated
- Watermarking
- Anthropic will embed invisible watermarks in Claude's writing in response to European Union transparency regulations; OpenAI, Anthropic and Google already embed watermarks in some AI images
- Expert advice
- Hany Farid of Dartmouth recommends relying on trusted media outlets over social media for news
Quotes
Max Spero
Founder of Pangram, the AI text‑detection company
“We’re reverse engineering their stylistic fingerprint.”
indianexpress.com
Hany Farid
Professor at Dartmouth College and founder of GetReal Security
“Stop getting your news from social media. An assumption that most of what you are seeing is fake is probably pretty good right now.”
indianexpress.com










