Perplexity

In order to not have articles being flagged as AI generated content. The text should have high perplexity along with high burstiness.

Perplexity measures the amount of information contained in a text. Despite the high redundancy of natural language, which allows us to communicate effectively even in noisy environments, some information can still be lost without compromising overall understanding. However, external factors such as alliteration, which adds an additional layer of meaning, can contribute to high perplexity that AI may not be able to recognize.

AI, based on probabilistic models, struggles with text that contains high variability in terms of word specificity, such as legal documents that require precise use of specific keywords and phrases.

Generative AI for text operates by predicting the most likely word to follow a given word. In a text with perfect perplexity, the AI would always make incorrect predictions.

Perplexity in language models

See also: language models

Perplexity is an important measurement for determining how good a language model is. Essentially, it quantifies the quality of the model's predictions by evaluating the inverse probability of the test set, normalized by the number of words, or by calculating the average number of bits required to encode a single word through cross-entropy. Perplexity can be perceived as the weighted branching factor, and a high perplexity score represents a higher degree of confusion in the model's next-word predictions, while a low perplexity score implies greater confidence in the model's output.