A 40-Line Python Script Can Detect AI Text, But Its Failures Reveal More Than Its Wins

·1 views

A developer built a compact 40-line Python AI-text detector using two signals: sentence-length burstiness and repeated n-grams, with AI-generated text typically scoring lower due to uniform structure and repetitive phrasing. The tool works as a quick triage filter but breaks down in three notable ways identified during real-world use. Short texts under roughly 40 words produce unreliable scores, while naturally uniform genres like legal writing or API documentation are wrongly flagged as AI-generated. The detector is also trivially easy to game, as minor edits such as splitting sentences or swapping synonyms can reset the score without improving actual prose quality. The author concludes that such a detector serves only as a smoke test, not a reliable judge of writing quality or human authorship.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Understanding What Databases Actually Do Under the Hood

A developer reflecting on system design realized they lacked a foundational understanding of how databases work internally, beyond just writing SQL queries. The article explores how databases handle large-scale data retrieval, explaining why full table scans become impractical at millions of rows. Indexes, such as those on a user_id column, allow databases like PostgreSQL to locate relevant rows efficiently without scanning entire tables. The piece also distinguishes between index types — including B-trees, Hash indexes, and LSM Trees — noting each suits different read and write workloads. The author concludes that a system designer's role is not to implement these structures manually, but to understand access patterns well enough to choose the right database features.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Review of 19 Jev AI Trading Projects Finds No Clear Evidence of Profit

A trading tools developer reviewed all 19 publicly available Jev-based finance projects during the week of 24 September 2026, finding that most were demos or paper-trading experiments with no live profitable results. Jev is a structured AI model by TypeSafe AI that takes JSON state inputs and returns probability outputs, a format that superficially suits algorithmic trading. Of the 12 projects that touched actual trading or price prediction, only one was live-capable by default, while the rest ran on testnets, backtests, or synthetic data. Four projects directly compared Jev against simpler rule-based alternatives, and in none of those did Jev clearly outperform — in one stock prediction test, Jev scored 45% accuracy while a naive always-down strategy scored 51.7%. The most substantial project, QuantDinger, was flagged for deeper investigation in a follow-up series.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds VidFixa, a Multi-Platform Video Downloader Using Go, Nuxt and PostgreSQL

A developer built and launched VidFixa, a video downloader supporting Instagram, TikTok, Facebook, X, and LinkedIn, as a hands-on learning project rather than a typical tutorial exercise. The application allows users to paste a video URL and download the content without a watermark, with the backend powered by Go and Chi, the frontend by Nuxt, and PostgreSQL for data storage. Video processing relies on yt-dlp and FFmpeg, while background workers handle download jobs asynchronously to avoid blocking HTTP requests. VidFixa offers three subscription tiers — Free, Plus, and Pro — with monthly download limits enforced for both anonymous and registered users, and payments managed through Bachs. The project is open-source on GitHub and deployed on Render, with the developer navigating real-world challenges including Docker configuration, CORS setup, webhook integration, and environment-specific deployment issues.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to test n8n webhook workflows locally without triggering live APIs

Developers at Weio, Inc. have outlined a method for testing n8n webhook workflows without making real calls to services like WhatsApp, Google Sheets, or Slack. The approach centers on adding a dry_run flag to webhook requests, which a Switch node uses to route traffic either to real action nodes or directly to a response node. All business logic is isolated inside a single Code node that returns a plain object, making it independently testable through fixture JSON files and CLI assertions. The pattern also covers running n8n in a temporary Docker container for isolated test environments, with specific warnings about timing traps such as waiting for full database readiness before importing workflows. Common CLI pitfalls around workflow activation in non-queue mode are also documented to help teams avoid silent failures during automated test runs.

0 comments Read more at DEV Community