A Reddit user’s repeated tests found that Stylome classified every Gemini-rewritten comment as “human-written,” with equally low confidence. The tool claims to identify AI text, yet it cannot reliably distinguish machine-generated content—a direct blow to its credibility as a community moderation tool.
What this is
Stylome is a bot used in Reddit communities to flag suspected AI-written comments. A poster rewrote original comments with Gemini, Google’s generative AI model, and used them as controls. Stylome still labeled them as human-written and assigned low confidence.
The issue goes beyond a single misclassification: there is a power imbalance between the tool and community members. Stylome is operated by a private company, yet it intervenes in public discussions. Comment content may also be used for model analysis or product improvement, but the original post provides no clear explanation of how the data is processed.
Industry view
We do not think this is enough to prove that Stylome has failed. It does show, however, that “AI text detection” cannot be treated as a reliable arbiter of fact. Text detection generally relies on statistical patterns, while model rewrites, hybrid editing, and human polishing all blur the boundary.
There are two main layers of risk. First, false flags can expose ordinary users to unnecessary scrutiny. Second, without mechanisms for appeals, audits, and data disclosure, automated labeling can easily turn an assistive tool into an unaccountable governing authority. The post also notes that the tool does not use a local model, raising privacy questions: it remains unclear whether comment content leaves the platform or is retained.
Impact on regular people
For enterprise IT: When procuring content-detection services, companies should put false-positive rates, raw-data retention periods, whether models run locally, and appeal mechanisms into the contract. Vendors’ demonstration claims such as “can identify AI” are not enough.
For workplace use: Schools, employers, and platforms may increasingly reference AI-detection results. These results are better treated as prompts for manual review than as grounds for disciplinary action, rejection, or disqualification.
For the consumer market: If detection tools publish their validation results and explain how data is used, they could become valuable services for content platforms. Conversely, vague accuracy claims and hidden data-processing practices will make users more wary of such products.