The author developed a Python-based detector to identify unedited AI-generated text, known as 'slop'. The tool uses 12 specific criteria, five of which are automated via code, while the remaining seven rely on manual editorial guidelines.
The script analyzes text by splitting paragraphs into sentences and using regular expressions to identify patterns such as repetitive introductory phrases, template-based announcements, and improper use of quotation marks. It does not determine authorship but flags potential issues.
Automated checks include a scoring system where specific findings trigger point deductions. For instance, three or more empty introductory phrases stop the text for review. The tool also calculates the standard deviation of sentence length to detect monotonous rhythm.
Certain features, such as fabricated numbers, lack of specific details, and complex syntactic structures like repetitive dashes or 'not A, but B' antitheses, remain difficult to catch with code. The author notes that the tool sometimes struggles to distinguish between examples of bad writing and actual usage within a text.
The detector operates locally using the standard Python library and provides a verdict based on a point system: zero indicates normal text, one suggests revisions, and two flags likely AI content. The author emphasizes that each trigger must be reviewed alongside the specific phrase identified.
