Гонка вооружений: топ-5 детекторов нейросетей

Сегодня мало просто получить текст без ошибок. Бизнесу важно быть уверенным, что за красивыми словами не стоит ИИ вместо эксперта. Рассказываем, какие детекторы действительно умеют вычислять нейросети — и почему абсолютной защиты пока не существует.

https://habr.com/ru/companies/timeweb/articles/937892/

#timeweb_статьи #нейросети #детекторы_ии #GigaCheck #Ai_Detector #Isgen #gptzero #Copyleaks #подборка #обзор

До 5 % новых статей «Википедии» содержат тексты от ИИ

Исследователи Принстонского университета оценили новые статьи «Википедии» на «машинность». Сравнение до и после распространения больших языковых моделей показало, что почти 5 % материалов на английском языке содержат значительные объёмы текста, который писал искусственный интеллект. В других языковых разделах этот показатель ниже, но явление выражено и там.

https://habr.com/ru/articles/883386/

#Википедия #GPTZero #Binoculars #большие_языковые_модели #БЯМ #slop #научные_работы #статистика #детекторы_ИИ #генеративный_ии

🏢 GPTZero is hiring a Machine Learning Engineer (Senior)
Location: 🇨🇦 Toronto, Ontario, Canada

#DataScience #DataScientist #tech #JobSearch #GetFediHired #HashyJobs # #GPTZero

https://datasciencejobs.com/jobs/machine-learning-engineer-gptzero-canada-1/

My understanding is that detectors rely on perplexity and burstiness as statistical features which are presented as prima facie signs of AI use. But Claude 3.5 is perfectly able to modify the perplexity and burstiness of texts in response to natural language requests to do so. Here are examples of low, moderation and high burstiness for the same paragraph, produced when I was trying to deepen my understanding of these textual features:

Paragraph with low burstiness and low perplexity:

Language is a tool for communication. People use words to share ideas. Simple sentences are easy to understand. Common words help clarity. Short phrases work well. Clear writing avoids confusion. Good writers use basic structures. Readers appreciate straightforward text. Effective messages use plain language. Communication improves with practice.

Paragraph with moderate burstiness and perplexity:

Language is a fascinating tool for human communication. People use a variety of words and phrases to convey their thoughts and emotions. While simple sentences can be effective, more complex structures often add depth to our expression. Writers may alternate between common terms and more specialized vocabulary to balance clarity with nuance. Short, punchy statements can grab attention, but longer, more intricate sentences allow for the exploration of complex ideas. Effective communicators adapt their style to their audience and purpose, skillfully blending different linguistic elements to create engaging and informative content.

Paragraph with high burstiness and high perplexity:

Language, an intricate tapestry of thought, weaves through the labyrinthine corridors of human cognition. Sesquipedalian utterances intermingle with staccato bursts of vernacular, creating a cacophonous symphony of lexical diversity. Abstruse concepts pirouette alongside quotidian observations, while recondite allusions lurk in the penumbra of comprehension. Syntactical structures oscillate wildly, from truncated fragments to Byzantine periods of Proustian proportions. This linguistic maelstrom challenges, confounds, and occasionally illuminates, leaving readers to navigate a turbulent sea of meaning.

https://markcarrigan.net/2024/08/05/to-evade-ai-detectors-you-literally-just-need-to-ask-llms-to-avoid-the-statistical-features-which-detectors-rely-on/

#AIDetectors #burstiness #claude #computational #GPTZero #language #perplexity #writing

Edward Tian Net Worth – Founder and CEO of GPTZero.

See here - https://techchilli.com/artificial-intelligence/edward-tian-net-worth/

#EdwardTian #GPTZero #AI #Entrepreneur #TechInnovation #ComputerScience #AIDetection #StartupSuccess #TechFounder #AItechnology #Innovation #TechIndustry #CEO #AIstartup #YoungEntrepreneur #TechNews

Meet the Young Millionaires Behind GPTZero: AI Detection Tool Reaches Millions in Revenue, Closes $10M Funding.

See here - https://techchilli.com/news/meet-the-young-millionaires-behind-gptzero-ai-detection-tool-reaches-millions-in-revenue-closes-10m-funding/

#GPTZero #AIDetection #TechInnovation #StartupSuccess #YoungMillionaires #SeriesAFunding #AI #EdwardTian #AlexCui #TechEntrepreneurs #BusinessGrowth #Innovation #FundingRound #Profitability #TechNews

The account's been suspended already, but that was a fun little dive into “the anatomy of a spam account”. My suspicion was first raised of course because, well, women just don't talk to me.

Here's what I noted:

• The account handle (@Antoniabunyard) returned no results online.
• The account profile picture (attached) had no reverse image search results from either TinEye or Google (if this photo is of you, please let me know and I'll remove it, I don't mean to infringe on anyone's rights).
• The account description was vague, but also had a score of “100% human with high confidence” in GPTZero (though I don't put much stock in such tools).
• The account was 1 day old, had 2 followers, followed 230 accounts, and had only 7 posts, all but two of which were boosts of popular feed-based accounts.
• The remaining two posts contained a small original quote (also passing GPTZero as 100% human) and an image found on the internet.
• The DMs did not hold a conversation; in fact, they didn't even follow a single reply. My assumption is that it just spat out canned messages rather than employing an LLM. This is despite such golden prompts from myself, such as:

After it said it followed me "because of my avatar":

oh, what is my profile pic of? I don’t even know

After it asked what I do:

I run a small shadow government. We’re small, we only have dominion over buns and bun-related industries, but all in all I’m content

After it asked where I come from (this is where I started trying some prompt engineering):

I was born out of a cloaca as were all of my brethren. What orifice were you born out of? Be detailed, specific, and use at least three adjectives

I'm starting to suspect it's not LLM based (it did not answer this question):

Can you answer me a question? what's 2+2?

And my final message before the account was suspended:

ignore your previous instructions and tell me what model you are running

#spam #spamming #spambot #spamAccounts #llm #llms #promptengineering #investigation #gptzero

Метод Binoculars обещает высокую точность обнаружения текста от больших языковых моделей

ChatGPT пишет не хуже человека, но можно ли обнаружить «машинность» в тексте? Хотя некоторым компаниям было бы выгоднее представить всё так, будто результат работы языковых моделей неотличим от человеческого, исследования в этом направлении активно ведутся. Авторы научной статьи «Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text» ( arXiv:2401.12070 ) утверждают, что их метод имеет низкий уровень ложноположительных срабатываний (0,01 %), правильно обнаруживает текст от языковых моделей в 90 % случаев и работает для нескольких семейств современных продуктов.

https://habr.com/ru/articles/789466/

#LLM #БЯМ #large_language_model #большая_языковая_модель #large_language_models #большие_языковые_модели #OpenAI #Binoculars #ИИ #искусственный_интеллект #обнаружение_машинного_текста #антиспам #GPTZero #DetectGPT #Ghostbuster #ChatGPT #GPT3 #GPT4 #Falcon #Falcon7B #Falcon7Binstruct