“Think of ChatGPT as a blurry JPEG of all the text on the Web.”
Science-fiction writer Ted Chiang introduced this metaphor in a 2023 essay about large language models. A JPEG compresses an image by preserving patterns that matter perceptually while discarding some exact information. When decompressed, it produces a convincing approximation rather than a perfect copy of every original pixel. Chiang asks readers to imagine a language model similarly: it encodes statistical regularities from immense collections and reconstructs plausible text from that compressed representation. The comparison makes both the fluency and the characteristic errors easier to understand.
Compression is powerful because language contains repetition and structure. Facts, phrases, genres, arguments, and styles recur in related forms. A model can use those regularities to generate a summary, explanation, or new passage without storing a searchable copy of every source. The result may capture the broad shape of familiar knowledge. Yet when an exact citation, name, date, or uncommon detail is needed, plausibility can substitute for retrieval. Like a blurry artifact that looks reasonable from a distance, invented specifics may fit the pattern while failing to match reality.
The metaphor suggests how to use these systems well. Generation is valuable for brainstorming, rewriting, outlining, translation, and exploring alternatives when every detail will be reviewed. Tasks requiring precise facts should connect the model to reliable sources and require verification. Users need to distinguish a fluent answer from evidence. Organizations should design interfaces that cite materials, expose uncertainty, and make checking easier instead of presenting every response with the same polished authority. The workflow, not just the model, determines whether compression becomes useful or misleading.
Chiang’s essay also raises a question about originality. Human writing is not valuable merely because it forms a statistically acceptable sequence. It reflects attention, selection, experience, and a relationship between an author and readers. Generative tools can rearrange patterns rapidly, but abundance does not guarantee insight. Writers may use them as aids while preserving responsibility for claims and purpose. The blurry JPEG metaphor therefore extends beyond factual error: it asks what gets lost when cultural material is compressed and recreated without the particular contexts that originally gave it meaning. A metaphor is not a complete technical theory, but Chiang’s image gives users a productive habit: appreciate the reconstruction while checking the details that cannot safely be blurred. That habit protects both accuracy and authorship, reminding the user to restore evidence, context, and deliberate purpose to a statistically convincing draft. Verification completes the work.
A model preserves statistical patterns from large collections and reconstructs plausible language from that compressed learning.
It does not function like a conventional database that retrieves an exact stored record for every answer.
Common structures may be reproduced convincingly while rare names, dates, citations, and relationships are distorted or invented.
Fluency is compatible with error because the system is optimized to generate likely text, not to guarantee evidence.
Models can help brainstorm, transform, summarize, and explore when a person reviews the result and checks important claims.
Source-connected workflows and clear citations are essential when exact facts carry legal, medical, financial, or public consequences.
Explore more "Quotes of The Day"
Continue exploring how writers, thinkers, artists, and advocates have described the choices that shape an everyday life.