The use of AI-generated content is growing rapidly across the internet. Along with this growth, a new problem is emerging, which could be described as AI Slop Fatigue. Social media feeds are increasingly filled with posts featuring repetitive bullet points, excessive emojis, overly polished language and frequent use of em dashes. As a result, it can be difficult to determine whether a piece of content was written by a human or generated by an AI chatbot.
Amid this trend, Pangram and other AI detection tools are gaining popularity. Pangram is an AI-powered service that analyses text and estimates whether it was created with the help of artificial intelligence. Users can paste text into the platform to see how much of the content may have been AI-generated.
Pangram was tested for around a week using different types of content, including personal writing, AI-generated passages and social media posts. Across dozens of tests, the tool performed well at distinguishing between human-written and AI-generated text.
Pangram is one of several AI detection tools currently available, but its reported accuracy has attracted considerable attention. The company claims its technology can correctly identify AI-generated text in 9,999 out of 10,000 cases. An independent University of Chicago study also found that Pangram performed extremely well in text detection.
Older AI detectors, by comparison, have made serious mistakes. Some tools have incorrectly classified authentic writing by famous authors such as Charles Dickens as AI-generated.
If the amount of AI-generated content online continues to increase, AI detection tools could eventually become as common as antivirus software. They could help users identify fake, misleading and low-quality content.
However, AI detection technology still faces a major challenge. While these tools can perform well when analysing text, they are considerably less reliable when it comes to identifying AI-generated images. This is particularly concerning because fake AI images can be an effective way of spreading misinformation online.
How Does AI Detection Work?
AI chatbots analyse patterns in human-created content and use those patterns to generate new material. AI detection systems take the opposite approach. They study the patterns and characteristics associated with content produced by different AI models to determine whether a piece of content was generated by AI.
Detecting AI-written text is not simply a matter of looking for em dashes, bullet points or a particular writing style.
According to Pangram co-founder Max Spero, AI models such as ChatGPT, Claude and Gemini follow complex decision-making processes when selecting each word. The choice of one word can influence the words that follow.
AI detectors analyse large amounts of data produced by different AI models to understand these patterns. This allows them to identify what can be considered a “stylistic fingerprint” associated with each AI model.
Detecting AI-generated images works differently. Potential indicators can include unusual pixel textures, unnatural lighting, overly saturated colours and other visual inconsistencies.
Some AI image-generation companies, including OpenAI, Anthropic and Google, are also working on invisible watermarks and other digital identification technologies. These systems are intended to make it easier to determine whether an image was generated using AI.
Similar technology is being developed for AI-generated text as well. Anthropic, for example, has announced plans to introduce invisible watermarking for text produced by its Claude chatbot. Such technologies could make AI-generated content easier to identify while also helping companies comply with emerging transparency regulations.
Testing AI Text Detection
Around 50 tests were conducted using Pangram’s text detection system. Some of the tests involved inserting AI-generated sentences into passages written by humans.
In most cases, the tool correctly identified which sections had been written by a person and which had been generated by AI.
The reverse test was also conducted. Human-written passages were inserted into AI-generated paragraphs, and Pangram was generally able to identify the human-written sections while classifying the remaining content as AI-generated.
Several publicly available passages written by Charles Dickens were also tested. Pangram classified all of the samples as human-written.
The tool was also used to analyse several LinkedIn posts that appeared to have been generated using AI. In one case, Pangram suggested that the author had written the opening line themselves but had used AI to generate the rest of the post. The author later confirmed that this was the case.
These tests suggest that AI text detectors can be useful for identifying content that initially appears to be entirely human-written.
AI Image Detection Still Has Major Limitations
The results were very different when AI-generated images were tested.
Twenty AI-generated images that had previously been shared online and debunked by news organisations were tested using Pangram and another AI image detection tool, Hive Detect.
Both tools showed several weaknesses during the tests.
Hive Detect incorrectly classified eight AI-generated images as real. These included a deepfake showing actress Zendaya as pregnant, an image depicting Senator Mitch McConnell in a hospital bed and a fake San Francisco street sign that falsely claimed stealing goods worth less than $950 from stores was legal.
Pangram also incorrectly classified two AI-generated images as real, including the fake McConnell image and the fabricated San Francisco street sign.
In addition, Pangram refused to analyse four images. Some were considered too violent, while others were of insufficient quality for the tool to provide a reliable result.
However, both detectors successfully identified several widely circulated fake images. These included a fabricated photo allegedly showing Zendaya’s wedding to her Spider-Man co-star Tom Holland, an image showing Donald Trump holding a girl during his visit to China and a fake photo depicting the Clintons partying with Jeffrey Epstein.
Hive Detect noted that in cases such as the McConnell image, inaccurate results could be caused by the lower quality of copies shared on social media. Social platforms frequently compress or resize original images, which can remove important information from the file.
This is one of the biggest challenges facing AI image detection. By the time an image reaches ordinary users, it may have already been compressed, cropped, resized or turned into a screenshot.
If such small modifications are enough to prevent an AI detector from correctly identifying a fake image, the technology becomes significantly less useful in real-world situations.
Pangram co-founder Max Spero also acknowledged that the company’s image detection feature is still in its early stages and continues to be developed. According to him, Pangram correctly identified 41 out of 43 AI-generated images in the company’s internal tests.
Why Is It So Difficult to Detect AI-Generated Images?
AI-generated images can easily be edited, manipulated, cropped, compressed and re-uploaded across different platforms. This makes it difficult to determine an image’s authenticity using automated detection alone.
Hany Farid, a Dartmouth professor and founder of digital content authentication company GetReal Security, also tested Pangram’s image detector. He uploaded five AI-generated wartime images, and the tool correctly flagged three of them.
According to Farid, distinguishing between real and fake images is difficult because photographs can be modified, distorted and manipulated in many different ways.
His research suggests that visual AI detectors are still not reliable enough for their results to be treated as definitive proof that an image is real or fake.
Conclusion
AI detection technology is developing rapidly, but its accuracy depends heavily on the type of content being analysed.
For now, text detection appears to be considerably more reliable than image detection. Tools such as Pangram can potentially help identify AI-generated emails, fake reviews, AI-written social media posts and other forms of low-quality online content.
AI-generated images, meanwhile, remain a major challenge. An image can be compressed, edited, cropped or manipulated multiple times before reaching users. As a result, even advanced AI detectors can struggle to determine its true origin.
For now, it is better to rely on reputable news organisations and trusted sources when verifying potentially fake or misleading content.
As AI-generated content continues to spread, users should remain cautious about suspicious posts, images and claims. Rather than treating an AI detector’s result as definitive proof, it is safer to use such tools as an additional layer of verification.
About the Author
