The following is a summary of Turnitin's similarity and AI writing scoring as explained in this video, omitting instructional content. 

Similarity score: Comparison of student submissions to existing writing to detect repetition of words, phrases, or whole texts.

AI writing detection score: Calculation of the overall percentage of AI generated text

The similarity score is based on checking text in a paper against a database of billions of web pages, student papers, documents and publications. This includes papers from your organization submitted to Turnitin previously. The score is a predictive indicator, where the higher the score, the more likely there has been academic misconduct (like plagiarism or contract cheating). 

Investigation

For any paper with a similarity score, you can view the similar text identified as similar and compare against the matching source. If the matching source is not public, you can request access to the source paper.

Turnitin uses AI to detect text written by generative AI, like ChatGPT (3, 3.5, 4). It processes the document in segments (like sentences and paragraphs) and comes up with an overall average score based on the assessment of these segments. The video implies that it works by extracting the context from each segment, uses generative AI to rewrite the segment, and then compares the AI rewritten text to the original text to calculate the difference based on statistical patterns. Segments deemed likely to be written by AI count towards the AI writing score. The overall AI writing score is the percentage of text likely written by AI over all text processed for AI detection. Text that is not long form prose is omitted from AI detection and score calculation, such as poetry, bullet points, tables, and annotated bibliographies.