2 weeks ago

AI Slop Detector Excels at Text, Struggles With Images

AI Slop Detector Excels at Text, Struggles With Images
AI slop detector works well on text but much less so on images: Here’s why · indianexpress.com

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.

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

Sources

Related news