Bilingual

Show HN: How clanker are you? A reverse Turing test
18 days ago
- A tool called "How clanker" analyzes Twitter (X) posts to measure how AI-like (predictable) someone's writing is.
- It pulls recent posts from a user, excluding retweets and replies, and scores each word based on its surprisal (predictability) under language models.
- Low surprisal (high predictability) results in a "clanker" (AI-like) score, while high surprisal indicates more human-like writing.
- The method uses log probabilities from models, normalizing scores against baselines, and the overall score reflects the nearest model's assessment.
- In demo mode, scores are simulated using a deterministic stand-in instead of real models, and the tool is presented as fun, not a judgment of the person.
Probably the best AI detector tool is ready
15 days ago
- AI text detector analyzes linguistic patterns to detect AI-generated text with 99.5% accuracy, showing the reasoning behind each verdict.
- Features include lightning-fast results (under 10 seconds), unlimited checks, real linguistic analysis using proven techniques, detection of over 10 AI models, and 100% privacy with encrypted, non-permanent data transmission.
- Offers sentence-by-sentence color-coded analysis and PDF reports with highlights and methodology breakdown.
- Built for educators, content creators, and businesses to verify authenticity and maintain integrity; includes transparent, cancel-anytime pricing plans.
- User testimonials from students, professors, writers, and editors praise its accuracy, low false positive rate, and usefulness in academic and professional settings.
- Based on peer-reviewed science using metrics like perplexity (lower in AI text), burstiness (sentence variation), and entropy (syntactic complexity) from institutions like Stanford and UCLA.
- Helps prevent language model collapse and machine-generated content dilution; includes quotes from researchers on the importance of detection tools.
- FAQs explain the detector as a tool analyzing text via linguistic metrics to distinguish human from AI writing, trusted by thousands and offering Premium upgrades for bulk scanning and API access.
Detecting LLM-Generated Texts with “Classical” Machine Learning
5 days ago
- Mainstream LLM-generated text as of early 2026 shows strong statistical patterns, detectable by traditional machine learning models.
- An online demo using a simple SVM classifier achieves ~85% single-sentence accuracy, trained on human and LLM-generated data.
- Data was gathered from pre-2022 human texts and LLM-regenerated summaries via multiple low-cost/free APIs from models like Gemini, Qwen, and GLM.
- Seven binary classifiers were trained, using majority voting; detection remains robust against common anti-detection methods like translation or rewriting prompts.
- Testing on unseen models and human texts shows low false positives; analysis of Lofter content suggests significant undisclosed AI-generated material.
- The author expresses skepticism about LLMs' creativity, viewing AI-generated content as repetitive and shallow, and reflects on broader implications of AI misuse.
How Does Pangram Work?
4 days ago
- Pangram is a classifier model that distinguishes AI-generated text from human writing by analyzing writing style patterns.
- It operates by mapping writing styles, placing similar authors close together; AI and human texts form distinct clusters on this map.
- Pangram evaluates entire texts, not just individual words, avoiding pitfalls of perplexity-based detectors which often misclassify historical documents.
- Training involves pairing human-written texts with AI-generated doubles to refine detection, focusing on subtle stylistic boundaries.
- The model requires a minimum word count for accuracy, and its performance is statistically validated through internal benchmarks and third-party testing.
Over 30% of new ArXiv submissions now read as AI-written
a day ago
- A study measured AI-written content in arXiv papers using a calibrated detector with a 0.4% false-positive rate on pre-ChatGPT human text.
- About a third (32%) of recent arXiv papers from 2023-2026 read as machine-written, with peaks near 39% in early 2026, showing a significant rise post-ChatGPT.
- Field analysis reveals wide variation: Computer science leads at 65% flagged, while mathematics is lowest at 0.7%, partly due to detector limitations with notation-heavy prose.
- Limitations include small control sample sizes per field, potential detector blind spots (especially in mathematics), and inability to distinguish between heavy AI assistance and full generation.
- The detector provides a lower bound on AI adoption; reported prevalence is conservative due to incomplete coverage of diverse models and prompts used by authors.
Against Claudefishing – AI detection feature on Substack
5 hours ago
- Substack is launching a feature with Pangram to detect AI-generated text, aiming to maintain human authenticity.
- The problem is not AI use per se, but the mismatch between reader expectations and reality, especially when readers unknowingly engage with AI content.
- Substack is introducing tools for transparency, including a detection scan and a 'How I make this' statement for creators.
- The feature is available on web and iOS, with Android coming soon; future plans include customizable preferences for AI content.
- The article highlights a debate: detection tools can help transparency but may become a purity test, as AI-assisted writing can still involve human creativity.

Hasty Briefsbeta

#ai detection

Show HN: How clanker are you? A reverse Turing test

Probably the best AI detector tool is ready

Detecting LLM-Generated Texts with “Classical” Machine Learning

How Does Pangram Work?

Over 30% of new ArXiv submissions now read as AI-written

Against Claudefishing – AI detection feature on Substack