Recently, as I was scrolling through my phone, I realized I was suffering from a new kind of fatigue: slop-induced burnout. Almost all the posts I was reading on social media were riddled with bullet points, emojis and em dashes. They clearly looked like the work of a chatbot.
This month, my suspicions were confirmed by a web tool that is getting buzz: Pangram, which uses artificial intelligence to detect when words were generated with chatbots. Users can simply paste in a block of text for Pangram to scan, and the tool calculates a percentage indicating how much of the text was churned out by a bot or written by a person.
I tested Pangram, which costs $20 a month, for about a week and tried all sorts of experiments to challenge the system. I fed it personal essays alongside text generated with chatbots instructed to imitate my writing. I also uploaded social media posts from LinkedIn and Reddit that I suspected were A.I.-generated. In dozens of tests, the detector never failed at distinguishing slop from human writing.
Though Pangram is one of many A.I. detectors that have hit the market in recent years, the product stands out because of its accuracy. Pangram says its technology can accurately identify 9,999 out of 10,000 times whether text was A.I.-generated; an independent study by the University of Chicago also found that the tool was near-perfect with text. In contrast, past A.I. text scanners made glaring errors, including the misidentification of writing by famous authors like Charles Dickens.
At the rate that slop is proliferating on the web — some studies say 50 percent of online articles are now artificially generated — A.I. detectors could become the next staple app, akin to antivirus software, that people will need to shield themselves from phony, low-quality information online.
Technology for scanning A.I. content, however, still has a long way to go. Like other A.I. detectors, Pangram isn’t very good at its other job: identifying A.I.-generated images, which are a more potent medium for spreading misinformation online. (More on this later.)
Here’s what you need to know.
How A.I. Detection Works
In the same way that A.I. chatbots excel at analyzing patterns in our creations to replicate writing and imagery, detectors like Pangram analyze data collected from A.I. models to decipher the characteristics of A.I.-generated writing and pictures.
For detecting A.I. writing, the methodology is not as simple as looking for em dashes and bullet points. Every A.I. chatbot, such as Claude, ChatGPT and Gemini, follows a “decision tree,” where each word chosen is a decision that leads to another, said Max Spero, a founder of Pangram. Accurate detection of A.I. writing is rooted in understanding how each A.I. model’s decision tree works, which involves gathering an enormous amount of data on each model, he added.
“We’re reverse-engineering their stylistic fingerprint,” Mr. Spero said.
With A.I.-generated images, there are telltale signs, like textures inside pixels that look rough when they should be smooth, oversaturated colors or inconsistencies in lighting.
Some companies behind A.I. picture generators, including OpenAI, Anthropic and Google, also embed invisible fingerprints, known as watermarks, in an image’s metadata to reveal to detectors that something is generated with A.I.
Watermarks are also coming soon to A.I.-produced writing. Anthropic announced this week that it planned to embed an invisible watermark in writing produced by its Claude chatbot to help people detect its A.I.-generated text. The move was a response to European Union regulations requiring transparency of A.I. content; other A.I. companies are expected to follow suit.
Testing Words
I ran about 50 tests on Pangram’s word detector. I tried pasting passages of my personal writing with some A.I.-generated sentences sandwiched in between. Pangram correctly detected that the prose was mostly human generated and highlighted the few sentences that a chatbot had produced.
I also tried the opposite: sprinkling some of my writing within paragraphs of A.I. slop. Again, Pangram correctly called out my writing as human and said the rest of it was A.I.
I tried uploading publicly available passages from Dickens. Pangram concluded they were 100 percent human-written.
I scanned a handful of LinkedIn posts that appeared to have been written by a machine. Pangram revealed that one tech worker had written the introductory sentence of his post but used A.I. to generate the rest. (I messaged the worker, who sheepishly confirmed this.) The experience felt empowering — for a moment I felt like Nada in the movie “They Live” when he donned sunglasses to detect aliens.
Picture Imperfect
I took a different approach to testing A.I. photo detection: I uploaded 21 A.I.-generated images that had been widely shared online and debunked by news outlets into Pangram and Hive Detect, a similar A.I. image detector. My sample size was small, but Pangram and Hive Detect failed so quickly that I didn’t see a point in continuing. (My colleague Stuart Thompson did a more comprehensive test of A.I. image scanners this year.)
Hive Detect incorrectly identified nine of the A.I.-generated images as real, including a recent deepfake that made the actress Zendaya appear to be pregnant, a picture of Senator Mitch McConnell on a hospital bed and a photo depicting a street sign in San Francisco that suggested it was OK to steal goods under $950 from stores.
Pangram incorrectly identified three A.I. images, including the one of Mr. McConnell and the phony sign in San Francisco. But the tool also declined to scan four images that looked violent or were too low quality to make a guess.
Both detectors correctly flagged some famous fakes, including a photo portraying the wedding of Zendaya and her “Spider-Man” co-star Tom Holland, a photo of President Trump holding a girl during his visit to China and a photo of the Clintons partying with Jeffrey Epstein.
Hive said that in some of my tests, such as the photo of Mr. McConnell, I may have scanned a copy of the image shared on social media lacking details of the original, which could lead to an incorrect result. My concern with this rationale is that by the time most people see A.I. slop online, the image is either a screenshot of the original or has been automatically shrunk down by the social media platform. If that’s all it takes to thwart A.I. detectors, the technology is pretty unhelpful.
Mr. Spero of Pangram said he was surprised that his tool had failed to flag the image of Mr. McConnell but added that A.I. image detection was a new, unfinished feature for the product. In his company’s tests of 43 A.I. images, he said, Pangram correctly identified 41.
The Bottom Line
Despite Pangram’s issues with spotting A.I. imagery, its proficiency in detecting bot-produced words will make it very useful for sniffing out annoyances like A.I.-written email scams, phony online reviews and uninteresting LinkedIn posts.
A.I.-produced photos, which spread rapidly online, remain a much greater problem. Hany Farid, a Dartmouth professor and a founder of GetReal Security, a company that verifies the authenticity of digital content, said he had run his own quick experiment with Pangram’s image scanner. He uploaded five A.I.-generated wartime photos, and Pangram flagged three.
Dr. Farid said distinguishing fake and real photos was extremely hard because an image could be distorted and manipulated in many different ways. In his research, visual A.I. detectors generally were still too flawed.
For now, he suggested that people rely on trusted media outlets for real information.
“Stop getting your news from social media,” Dr. Farid said. “An assumption that most of what you are seeing is fake is probably pretty good right now.”
The post I Tested a Popular A.I. Slop Detector. It Felt Empowering. appeared first on New York Times.
