GPT
Also called Generative pre-trained transformer
GPT stands for generative pre-trained transformer, a model architecture in which a transformer network is trained on a large body of text and then used to generate new text one token at a time.
Last reviewed 15 August 2026 · The HumanFlow team
In plain English
Three words, each doing work. Generative: it produces text. Pre-trained: the expensive learning happened before you arrived. Transformer: the network design underneath.
It is a description of how a model is built, which is why it has ended up in the names of products from several different companies.
A worked example
Taking the acronym apart.
G — Generative
Produces new tokens rather than labelling existing ones.
P — Pre-trained
Trained once, at great expense, on a broad corpus. Everything
after that is prompting or comparatively cheap adjustment.
T — Transformer
A network architecture whose attention mechanism lets every
position weigh every other, which is what made long-range
coherence practical."Pre-trained" is the word that explains the knowledge cutoff. The broad training happened at a point in time, and the model does not learn from your conversation.
"Transformer" is the part that is genuinely a technical claim; the other two describe use and process.
Because the term describes an architecture rather than a product, "a GPT" being used to mean "a chatbot" is a generic use of what began as a specific name — and it is now common enough that policies written using it are usually read that way.
Why it matters for AI detection
Institutional policies frequently say "GPT" when they mean any generative text tool. If you are reading a rule that names it, the question worth asking is whether the drafters meant one product or the whole category.
It matters for detector claims too. A tool advertising that it detects "GPT-generated text" is describing a training target, and a detector trained on one generation of output has no guaranteed relationship to the next.
The pre-training point also explains the fabrications: a model reproducing what a citation of this kind usually looks like is doing exactly what pre-training equipped it to do.
Commonly confused with
- Large language model
- LLM describes the scale and purpose; GPT describes the architecture and training approach. Most large language models in use are transformers, so in practice the terms overlap heavily.
- ChatGPT
- ChatGPT is one product built on models of this type. Treating the two as synonyms is why so many policies accidentally regulate a brand rather than a capability.
Read next
Part of the AI detection glossary.