AI Score Inconsistency & Over-Reliance on Percentages
Threads
Open the resubmission thread if your score changes between runs. Open the percentage thread if someone is treating the number like a literal measurement of AI-written text and you need a clearer way to explain what the score can and can’t mean.
Why identical or near-identical text can get different AI scores, and what testing method produces more reliable interpretation.
Why the percentage is not a literal “portion of text” meter, and how to avoid score-only accusations.
Copyleaks Score Changes on Resubmission
Score changes can happen because tools segment text differently, update internal models, or react to small differences in formatting and length. Even adding a heading, removing extra spaces, or changing paragraph breaks can shift the signals the model uses.
Practical approach: test the same text at consistent length (e.g., 600–1,200 words), run multiple trials, and compare section-by-section. If the score swings widely, treat the detector as unstable evidence and rely on stronger proof like drafting history and human review.
70% AI Score Doesn’t Mean 70% AI-Written
The percentage is a probability-style confidence measure, not a literal share of authorship. A “70% AI” label is not saying “70% of sentences were generated.” It’s the model’s estimate that the overall pattern resembles its AI vs human training examples.
Best practice: use the score as a triage signal only. Require additional evidence, allow appeals, and focus on concrete indicators (draft process, sources, revisions). Percentages should not be used as automatic penalties—especially when results fluctuate.
