Understanding AI Detectors: A Guide for Instructors

‍

‍

Regardless of where you sit on the “should you use an AI detector or not” debate, it’s important to understand how it works.

‍

This might be surprising to some but how they work is very similar to how an instructor reads and analyzes a paper to see if it is plagiarized. After all, we’ve trained AI on human behaviour and human-generated data.

‍

Every AI detector is a little different but generally speaking, they follow this structure:

‍

1. Analyzing Text and Detecting Patterns

AI detectors start first by reading the text, similar to what a human grader would do. They examine the vocabulary used, sentence structure, and any unique writing styles. Here are some of the ways that AI detectors can detect patterns:

‍

• N-grams: These are sequences of ‘n’ words that form sentences (more complex sentences have more n-grams). You tend to find more variation in n-grams in human writing.
• Syntax and Grammar: The AI will also look out for specific patterns in grammar.
• Stylistic Features: Analyzing the writing style and word choice. For example, ChatGPT loves to use certain words.

‍

‍

2. Use Training Data to Classify Patterns

AI detectors are also trained with datasets of human-written text and AI-generated text. The training data will help the detector-in-training understand the patterns that human writing exhibits and the ones that AI-generated text tends to exhibit such as the patterns we mentioned above.

‍

For example, perhaps in the training data, AI-generated text contains more sentences of similar n-grams like the following:

‍

"The geese are honking loudly in the early morning. The geese are flying together over the lake. The geese are landing gracefully on the water. The geese are foraging for food along the shore. The geese are resting under the trees."

‍

Versus a more inconsistent pattern:

‍

"The geese are honking loudly as they fly over the lake. Geese land gracefully on the water, searching for food. Under the trees, the geese rest, enjoying the shade and the cool breeze."

‍

‍

3. Is it AI-Generated?

‍

• Detection Algorithm: With all the above information, the AI detector will use a trained algorithm to determine the likelihood that the text is AI-generated.
• Anomalies: The AI detectors will identify anomalies when compared with datasets consisting of human-written text. For example, certain words are rarely used in university writing but it appears here.
• Thresholding: A lot of these tools will have ‘confidence’ levels related to the AI-detection result.

‍

While this is an oversimplification of how AI detection works, you can already see how it is trained and where bias can creep in.

‍

What if students are taught to write in a certain way?

‍

What if students use vocabulary that is overly complex for that grade level?

‍

What if AI-generated text now uses a variety of n-grams to form sentences?