Its helpfulness lies in what it doesn’t do. Summaries cut details out, and explanations might add new context. Simplify does neither. It focuses exclusively on rephrasing for clarity, ensuring the original meaning, details and nuance remain intact.
“Summarization tools aim to condense an article to its main topics, often sacrificing details for brevity,” says Sho Fujiwara, a product manager who helped develop Simplify. “Simplify focuses on making specific passages easier to understand without losing crucial information. This means a simplified version of selected text can sometimes actually be longer than the original.”
Starting Simplify
The path to Simplify began in the specialized world of medicine, where no detail should be spared. “Doctors sometimes use language that’s purposefully obscured to reduce patient anxiety or preserve privacy,” says Diego Ardila, a Google Research software engineer. “They might say, ‘The patient is undergoing emesis,’ which just means they’re vomiting. Sometimes there’s a use for that inside a hospital, but other times it actually gets in the way.”
The team built an internal simplification demo, and started testing it on text outside of medicine. “We found that it just kept working because the underlying AI models are general-purpose,” Diego says. “Seeing its potential for broader applications, we shared it with other teams.”
That demo caught the eye of the iOS Google app team. “The user benefit was immediately apparent,” Sho says. “It performed exceptionally well on technical subjects, but its effectiveness on non-technical topics — like an online discussion about basketball filled with slang — really convinced us of its wider value for anyone exploring unfamiliar subjects.”
Finding fidelity with help from AI
From there, the challenge was to turn a research prototype into a feature for millions. The team focused on fidelity: The model had to rewrite complex ideas without losing the original meaning or omitting critical details. How do you teach AI to do that?
The answer was to have another, more advanced AI act as its teacher. Instead of manually writing rules to define an effective simplification, the team built an automated feedback loop. One Gemini model would attempt to simplify a passage, and a second “evaluator” model would grade how well it preserved the original meaning. That critique was then fed back to the first model, which used it to improve. After repeating this process 824 times, the model effectively trained itself to master the art of simplification.
That training paid off: In research testing, people found the simplified text significantly more helpful than the original text, and better retained the information.
