How deeply does How Large Language Models Work cover digital watermarking?

Defines watermarking and gives a homoglyph example, but stops short of specifying a robust image-watermarking algorithm.

Digital watermarking

As explained in How Large Language Models Work

Watermarking is the process of embedding an identifiable signal or alteration in content to indicate its origin or authenticity.

Image watermarking embeds a subtle, identifiable alteration in content so its origin or authenticity can be recognized. For generated outputs, the alteration may help distinguish AI-produced material from human-produced material, although its reliability remains technically debated. section 9.4.2

Watermarking is presented as a practical response to several risks:

Identifying generated content before it contaminates future training data. section 9.4.2
Giving people such as teachers an alternative to error-prone detection methods. section 9.4.2
Using subtle changes, including visually similar characters in text, as an easily detectable marker. section 9.4.2

How How Large Language Models Work covers it

liveBook manning.com

Read How Large Language Models Work on liveBook Get the book at manning.com

Also explained in

liveBook manning.com

Introduction to Generative AI, Second EditionA watermark is a visible or embedded mark placed in an image or other media to indicate ownership, origin, or authenticity.

Compare how 2 books explain digital watermarking →

Related concepts

provenancealternative to · Provenance is information about the origin and history of data or artifacts, including their sources, derivation, transformations, and movement.
response generationuses · Response generation is the computational process by which a language model produces a natural-language answer or other output conditioned on an input, instructions, or context.
SynthIDincludes · SynthID is Google's technology for embedding and detecting imperceptible digital watermarks in AI-generated content.
noiseaddresses · Noise is variation, content, or data treated as unwanted, irrelevant, or distinct from the signal, pattern, or structure of interest.

This explainer was generated by Manning's knowledge graph from the text of How Large Language Models Work. The section links point to the parts of the book it draws on.