How do AI Detectors work?

Perplexity: Perplexity measures how predictable the next word of a sentence is. AI-generated writing tends to be more predictable because it’s designed to flow better. This predictability comes from machine learning models being trained on vast amounts of text, learning which words are most statistically likely to follow others based on patterns in that data. This means it uses words and phrases that seem most likely to come next, making the text easier to follow. In contrast, MOST people tend to write with less perplexity than an AI. But that's where a potential issue is introduced that I will talk about later. 

Burstiness: Burstiness refers to the variety in sentence length and structure. AI writing is a lot more uniform with its patterns and lengths, while humans often mix things up when writing. A human might follow a short sentence with a long one. An AI, less so. When burstiness is low, It’s another clue that the text might have been machine-written. 

For example this sentence generated by ai, has low burstiness: “The sun rose over the horizon, casting a warm glow across the quiet countryside.” While this human written text has higher burstiness: “The sun exploded into view, painting the sky with fiery streaks of orange and gold, while shadows stretched and danced across the dew-covered fields like restless spirits awakening from slumber.”

False Positives: False positives are the issue I was speaking about earlier; not everybody writes with less perplexity than AI. Some people write very similar to ChatGPT or Gemini, which can trick AI detectors into believing that human written text is machine-written. This is a huge issue since people can get accused of something they didn’t do, which can ruin their credibility. Now this is rare, but it's important to always document the writing and thinking process of your work, so you can defend yourself if you face this issue. 

Manual Spotting is still important: You don’t have to rely only on AI detectors to spot AI writing though; you can train yourself to notice the signs. Look out for things like language that’s too polite, sentences that sound stiff or robotic, strange or confusing logic, and a writing style that doesn’t match the person you think wrote it. The more you read AI-generated text, the easier it gets to tell the difference.