WebGPT: Improving the factual accuracy of language models through web browsing

OpenAI has bolted a text-based web browser onto GPT-3, and the result is a model that stops guessing and starts checking. Instead of spitting out whatever pattern-matched answer feels statistically likely, this version of GPT-3, internally called WebGPT, can issue search queries, click through results, scroll pages, and quote directly from what it finds. Then it stitches those quotes into an answer with citations attached, so the reasoning trail isn't hidden inside a black box.

The training process leaned on two techniques OpenAI has used before: imitation learning, where the model watches humans perform the browsing task and copies the behavior, and reinforcement learning from human feedback, where people rank different model answers and the model gets nudged toward whichever ones humans preferred. The questions used for evaluation came from ELI5, the Reddit forum where people ask for explanations of complicated topics in plain language, which meant the model had to handle genuinely open-ended, occasionally messy queries rather than tidy trivia.

What's notable is the comparison point. OpenAI didn't just check WebGPT against itself, it pitted its answers against ones written by actual humans who did their own research to answer the same ELI5 questions. In blind evaluations, WebGPT's responses held up well against that human baseline, a fairly high bar for a language model that's historically been prone to inventing facts with total confidence.

The browsing behavior matters as much as the accuracy gains. A model that can be watched navigating to a source and pulling a specific quote gives evaluators, and eventually users, a way to spot-check its work instead of just trusting the output. That's a meaningfully different failure mode than a model hallucinating a statistic and stating it as fact with no paper trail behind it.