OpenAI publishes WebGPT, a model that browses the web to answer questions
Answers preferred to Reddit's top-voted responses 69% of the time in blind comparison, but the model still fell short of human accuracy on TruthfulQA.
- Models & capabilities
- Notable
OpenAI published WebGPT, a version of GPT-3 fine-tuned to answer open-ended questions by searching and navigating a text-based web browsing environment rather than relying only on knowledge stored in its parameters. Given a question, the model could issue search queries, click through to pages, scroll and quote passages, and was required to compile evidence and cite the sources it used before composing a final answer — a structure intended to make its factual claims checkable rather than opaque.
Training combined imitation learning from human demonstrations of the browsing task with reinforcement from a learned reward model: humans compared pairs of model-written answers, and OpenAI used those preferences to train WebGPT toward answers people judged more helpful and better supported. On ELI5, a dataset of open-ended questions drawn from Reddit’s “Explain Like I’m Five” community, OpenAI reported that the best WebGPT model’s answers were preferred to human demonstrators’ answers 56% of the time, and to the highest-voted Reddit answer 69% of the time, at a comparable level of factual accuracy. On TruthfulQA, a benchmark designed to catch models repeating common misconceptions, WebGPT’s answers were rated true 75% of the time and both true and informative 54% of the time — an improvement over base GPT-3 but still below human performance.
WebGPT did not ship as a product; it remained a research demonstration of citation-grounded, tool-using question answering. It anticipated a pattern that became standard practice within a few years: giving a language model a browsing or search tool, training it to gather and cite evidence rather than answer from memory alone, and evaluating the result on faithfulness as well as fluency. That pattern later underpinned browsing features in consumer chatbots and, further out, more elaborate multi-step “deep research” agents built on the same core idea of a model that acts in an environment to gather information before answering.