Filter
Exclude
Time range
-
Near
Replying to @_akpiper
They only benchmark on binoculars? Not very valid to just ignore SOTA (even fast-detectgpt is stronger) and then claim all detection doesn’t work
2
57
These are my experiences with AI Writing Detection providers: I regularly use Grammarly for doing a test on my content for different audiences, Casual, Business, Expert, and all the combinations it offers to score your content against proposed audience. Now and again I will go to the "AI and Plagiarism" tab that is next to it on the right toolbar, out of curiousity, so many times its has marked entire sections as "AI generated / Paraphrased AI" and it is not. Then take the same text and I check against: • Turnitin - normally 90% to 100% human detection result • GPTZero - always 100% human detection result but last update was 2023 so maybe its out of its depth because I have tested it with pure AI generated text and it still gives 100% human • Sapling AI Detector, Originality.ai, Scribblr, DetectGPT, Winston AI, and ZeroGPT (different from GPTZero) - mixed results, overall verdict from my experience is error range 30% I wrote a short post about it last year here but I cant find it. The things that I find is that they all tend to attribute the following to AI generated: • Local turns of phrase • Technology term heavy content • Anything with lots of bullet points • Long Harvard Style reference bibliographies • Unusual sentence structures which are perfectly coherent and the function of being a writer for a long time. In other words they seem to punish or have bias in its training toward novelty or witty/skillful wordplay in your writing.
STOP I MEAN STOP USING AI WRITING DETECTION. NONE of them work. The may actually make really good writers dumb down because of clueless “AI Slop” folk who use the term to be sure they stay unread and uninformed. Just stop and call out the “AI Slop” folks for what they are.
1
3
5
303
دليل مختصر لاستخدام ChatGPT بذكاء أولاً: صياغة البرومبت (Prompt Format) استخدم القالب التالي لنتائج أفضل: 1) افترض شخصية معينة: "تصرّف كخبير في [المجال]" 2) حدّد الفعل الشكل الطول: "اكتب لي [نوع المحتوى] بصيغة [عدد الكلمات]" 3) وضّح الهدف: "هدفي هو: [تعليم، تبسيط، إقناع، بيع…]" 4) حدّد البيانات المطلوبة: "اعتمد على النقاط التالية: […]." 5) حدّد النبرة: "اكتب بنبرة [رسمية، تعليمية، ملهمة، قوية…]." 6) حدّد الجمهور: "موجه إلى [طلاب، جمهور X، العملاء…]." ثانياً: درجة الحرارة (Temperature) تتحكم في عشوائية الإجابة: •1.0 = إبداع وعشوائية أكثر •0.8 = متوازن •0.4 = عملي ومنطقي •0.2 / 0.1 = دقة وتركيز وحزم استخدم 0.1–0.3 إذا كنت تريد وضوحًا شديدًا، و0.7–1.0 إذا تريد أفكارًا جديدة وغير تقليدية. ثالثاً: الأدوار التي يمكنك طلبها (Role Playing) مثال: •كن إيلون ماسك •كن ستيف جوبز •كن مسوّقًا محترفًا •كن أخصائي تغذية •كن مدرب حياة •كن كاتبًا، مبرمجًا، صحفيًا… إلخ كلما كانت الشخصية أوضح، كانت الإجابة أدق. رابعاً: أساليب الكتابة (Writing Styles) يمكنك تحديد طريقة الكتابة: •شرح •رأي •رسمي •ودي •تسويقي •قصصي •ملهم •نقدي •مباشر •وصفي •مقنع •إبداعي خامساً: تقنيات التوجيه (Prompting Techniques) •سؤال مفتوح •سؤال مغلق •اختيار من متعدد •ملء الفراغ •مقارنة •ترتيب •تحليل •تلخيص •Brainstorming •Prediction سادساً: البدائل (Alternatives) نماذج ذكاء اصطناعي أخرى: Bard، Bing Chat، Claude، Jasper، Poe، Learn.ai وغيرها. سابعاً: تفادي السرقة الأدبية (Plagiarism) أدوات كشف الاقتباس مثل: •OpenAI Text Classifier •Turnitin •GPTZero •DetectGPT •Originality.ai وللتعديل: •Quillbot •Paraphraser •Grammarly ثامناً: المفاهيم الأساسية (Terminology) Input: ما تكتبه. Output: ما ينتجه ChatGPT. LLM: نموذج لغوي ضخم. Prompt Engineering: فن صياغة أسئلة قوية. Generative AI: الذكاء الذي ينتج محتوى. Training: تدريب النماذج على التحسين. AI Models: النماذج المختلفة. OpenAI: الجهة المطوّرة للنموذج. الخلاصة: كلما أتقنت كتابة البرومبت حصلت على نتائج أقوى. المفتاح: تحديد الشخصية الهدف النبرة الشكل الجمهور
1
29
224
9,214
29 Oct 2025
Replying to @SilvanaBuono
Podés usar herramientas para detectar si es texto hecho con IA. Podrías hacer el análisis en vivo con el grupo de alumnos, los que dan "positivo" o con alto porcentaje que lo reescriban en el momento. Tools: Turnitin AI Writing, GPTZero, Originality.ai, detectgpt
1
2
557
29 Oct 2025
What is LLM fingerprinting and why does it matter for open AI? LLM Fingerprinting: Technique to identify unique patterns/traits in AI-generated text (e.g., token probabilities, stylistic markers) to detect if content was produced by a specific model like GPT-4 vs Claude. How it Works: Analyzes subtle biases in word choice, sentence structure, or entropy that persist even after paraphrasing; tools like DetectGPT or watermarking embed detectable signals. Why It Matters for OpenAI: Abuse Prevention: Stops mass fake news, spam, or deepfake scripts using their models, this and many more! Accountability: Tracks misuse (e.g., exam cheating, plagiarism) without banning legit users. Regulatory Pressure: EU AI Act & US exec orders demand transparency; fingerprinting proves "AI-made" label compliance. Competitive Edge: OpenAI invests in watermarking (failed public rollout 2023) to stay ahead of Anthropic/Google detection arms race. OpenAI's Stance: Secretly watermarks GPT-4o outputs at token level (99% accuracy per leaks); balances detection vs evasion by adversaries. .@SentientAGI is really great. Gsenti everyone!
29 Oct 2025
.@SentientAGI is project, you can still do one or two on! Yap about it.. Go to discord and support others! Create spaces and talk about open source AI ✅ Gsenti everyone!!
30
25
532
Nankai University Develops Breakthrough AI Text Detection Method Researchers at Nankai University have unveiled a new “Direct Difference Learning” (DDL) strategy that significantly improves AI-generated text detection. Trained with DDL, detectors can identify content from the latest large models, including GPT-5, even if trained only on earlier ones like DeepSeek-R1. The team also introduced MIRAGE, the first benchmark focused on commercial large language models, built with data from 17 advanced models. In tests, DDL-based detectors achieved over 85% accuracy, outperforming Stanford’s DetectGPT and other leading methods by more than 68%. (Photo provided to China Youth Daily) #ArtificialIntelligence #AIDetection #GPT5 #MachineLearning #NankaiUniversity By Li Yuanyuan, Lin Xiaojing(intern) and HuChunyan/China Youth Daily
Epilepsy Surgery Enters Millimeter Precision Era: "Robot Laser" Pinpoint Brain Lesion Elimination #RenjiHospital successfully performed a groundbreaking "#robot #laser" surgery using Laser Interstitial Thermal Therapy (LITT). Without the need for craniotomy, this minimally invasive technique eliminates deep brain lesions with millimeter accuracy while protecting surrounding nerves. (Photo/Renji Hospital) #EpilepsySurgery #LITT #MedicalInnovation #RenjiHospital  
By Wang Ziyuan, Zhang Weiyi(intern) and Wang Yejie/China Youth Daily
6
8
32
7,011
Been diving into some ML papers on detecting LLM-generated code Currently reading two, and DetectGPT really caught my eye —super interesting stuff! Will share some cool insights soon! if time permits will code it up arxiv.org/abs/2301.11305
4
66
La mayoría lo utilizan es fácil detectarlo e incluso hay programas que te confirman si fue hecho con IA, La estafa aquí es que no te avisan que el contenido fue hecho por IA Ejemplo https//detectgpt.com/ Ai Detector You can trust detectGPT
26 May 2025
FONDO DE OLLA HISTÓRICO AJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJAJJJAAAJAJAJAJAJAJ
1
2
42
第一会翻墙就说明能接触真正的GPT,所以GPT这块应该不用讲。第二去淘宝搜“代写”,随便找几家和客服说“应聘写手”,然后按流程就能进接单群。第三就是多找几家,有的店待遇会比较好,有中介不用和客户直接对接,总之多找几家。最后就是接到单了用ChatGPT写文章,实在不会写可以找找ChatGPT生成文章的教程,然后再找点降AI的教程,用西湖大学的detectGPT查AI。写出来就好。如果你自己没有房贷车贷之类的经济压力,一天接个一两单也能养活自己
96
Replying to @ihsansantosoyt
Damn ive accidentally installed DetectGPT in my retina
179
文章がAI生成か、人間の文章かを判別するツール。 以外と面白い。 The AI Detector You Can Trust - DetectGPT detectgpt.com/

3
7
629
22 Nov 2024
🌶️takes: 🧠 Human are on on average 66.2% more creative than LLMs. 🤖 RLHF dramatically reduces the LLMs CREATIVITY by an average of 30.1% 🖨️ Using the Index, we can detect machine-text surpassing the strongest methods (e.g. DetectGPT) by 30% arxiv.org/pdf/2410.04265

1
1
5
451
22 Nov 2024
Finally, the CREATIVITY INDEX proves to be a surprisingly effective criterion for zero-shot machine text detection, surpassing the strongest existing zero-shot system, DetectGPT, by 30.2%, and even outperforming the strongest supervised system, GhostBuster, in 5 out of 6 domains.
1
1
11
735
24 Oct 2024
Replying to @vydamo_
"DetectGPT, was this written by an LLM?" "No, this text seems to be entirely original"
1
534
22,291
22 Sep 2024
记得去年斯坦福有个项目叫detectGPT 可以来检测内容是否为ai生成 不知道现在这类项目怎么都没什么声音了 文本,图片,声音,都越来越难以辨别 可能后面也会像工业时代汰换手工制品一样 后续莫非都是人类手动标注:纯手工制造? 反正即便是真的 也能说是ai生成的 本是ai生成的 也能让人说成真的😂
2
2
6
3,844
🇨🇳🤖 Westlake University (chinese univ) unveils Fast-DetectGPT. It's a tool that detects #AI-generated text 340x faster and 75% more accurately. It could combat fake news, fake reviews, and AI-written articles, promoting trustworthy AI. #GPT #fakenews pandaily.com/westlake-univer…

1
1
2
116
Fast-DetectGPT will be introduced on several social media platforms in the following days: 1. BAAI Talk: mp.weixin.qq.com/s/L5ta7MuOh… 2. TechBeat Talk: mp.weixin.qq.com/s/Pw_Hs5pwR… Welcome to attend!

17 Jan 2024
Excited to announce that Fast-DetectGPT made it to #ICLR2024 🎉 @NlpWestlake In the rebuttal phase, got an unfair 1 amidst 8866. Huge shoutout to @ericmitchellai for the public support. Appreciate your sense of justice! ⚖️🔍 Paper: openreview.net/forum?id=Bpcg…
7
658