LLM Watermarks: How a Model Signs What It Writes (English Edition) ler

epub LLM Watermarks: How a Model Signs What It Writes (English Edition)

por

Escolha um formato:

Leia uma amostra

Escolha um formato:

zip ler
rar ler
pdf ler
epub ler
txt ler
djvu ler

Descrição do livro

Every sentence an AI writes could be carrying a hidden signature. This book shows you how, with nothing more than a pocket calculator.

Text has no thickness. A paragraph is just characters, and nothing in it records who wrote it. Yet the world's largest AI systems are already signing what they write, invisibly, inside the words themselves. Google has been doing it in Gemini since 2024. The EU now requires it. Most people have never heard of it.

How a Model Signs What It Writes explains text watermarking from first principles, in plain English, with a vocabulary of only four words so that every number in every figure can be checked by hand. No code. No formulas you cannot follow. Just add, multiply, take a remainder, and watch a watermark appear.

You will follow Rosa, a lecturer who marks essays and cannot tell which ones a machine wrote, as she learns:

  • Why guessing from writing style fails, and who it fails against
  • How a language model actually picks each word, and where the hidden slack lives
  • How a secret key, a look-back and a hash turn that slack into a signature
  • Writing one watermarked word from start to finish, every step worked with real numbers
  • How a checker keeps score without ever reading the text
  • Why 40 words prove nothing, 100 are suggestive, and 300 are overwhelming
  • What survives light editing, what a free paraphrasing tool destroys in seconds, and why short answers and code cannot be marked at all
  • What SynthID-Text does differently, and what the EU AI Act's Article 50 asks for

With 43 figures and a running worked example, this is a short book you can finish in an evening and understand completely.

Who it is for

Teachers, editors and examiners who need to know what AI detection can and cannot prove. Journalists, policy makers and compliance teams facing new transparency rules. Developers and students who want the mechanism, not the marketing. And any curious reader who has wondered whether the text in front of them was written by a person.

What you will take away

A watermark is not proof of authorship. It is one line of evidence: very hard to forge, very easy to destroy, and useful only when you know exactly which question it answers. After this book, you will.

Número de páginas :50
Encadernação LLM Watermarks: How a Model Signs What It Writes (English Edition):Kindle
Livros recentes
Livros relacionados