Feeling dumb? Let Google's latest AI invention simplify that wordy writing for you

Zog brain hurt. Zog want Google help make read easier

Exasperated by the prolix verbiage and gratuitously convoluted phraseology that so often permeates technical treatises, philosophical discourses, or the meandering expositions of journalists afflicted by a lack of rhetorical economy? Then Google has a new AI feature for you - provided you use iOS, that is. 

The Chocolate Factory added a new feature to the Google app for iOS on Tuesday, dubbed "Simplify," which is meant to dumb down complicated writing without losing its meaning. 

Simplify's addition to the iOS app came with research proving how effective it is and an explanation of how Google achieved the system. What the tech giant calls a "novel … automated evaluation and iterative prompt refinement loop" echoes the idea of one model evaluating another - conceptually reminiscent of a good old-fashioned generative adversarial network (GAN) - so it's more a case of prompt engineering on autopilot than a game-changing innovation.

As Google explained in its research blog, the system works like this: A user selects a chunk of text on a web page from within the iOS Google app's built-in web browser, taps on Simplify, and then a pair of Gemini bots go to work trying to make it more readable. 

The system works using two evaluation methods: A readability assessment in which one Gemini bot scores the readability of the passage on a score of 1 - 10, and a fidelity assessment that a second Gemini bot (running version 1.5 Pro of the model) uses to "map claims from the original text to the simplified version" to see if any important information was lost or distorted in translation. 

The twin Geminis go back and forth through readability and fidelity assessments to optimize the passage, then spit it out for the reader, like in the below example provided by Google from a medical paper on emphysema. 

Google Simplify in action ... Click to enlarge

"This automated process, where one LLM evaluates the output of another and refines its instructions based on performance metrics and granular errors, represents a key innovation," Google said of its work. 

It's been a few years since we heard of the idea of pitting two neural networks against each other to get better output, but this sounds a lot like what the Mountain View boffins managed to reinvent. 

Google told us that, while the process seems like a GAN, Simplify doesn't actually use the specific method used by GANs (e.g., generating synthetic data indistinguishable from the real data a model was trained on). The difference is subtle, but Simplify is using multiple LLMs to eliminate the need to engineer a specific prompt, a Google spokesperson told us, rather than using multiple LLMs to generate a specific output.

Putting aside whether it's innovative or simply iterative, Simplify works, according to Google's stats. The researchers who developed Simplify put it to the test by presenting 4,563 people with original and simplified passages, and then asking multiple-choice questions to gauge how well they understood the content, and found "statistically significant results demonstrating the value of simplification."

According to their findings, the simplified version of the excerpts resulted in 4 percent more accurate quiz scores, which were even greater in certain areas like medical texts (15 percent improvement) and financial docs (6 percent). 

However, they did admit that LLMs can still get things wrong, even when simplifying jargon-filled writing. "While our system aims for high fidelity, LLM errors are possible, requiring ongoing vigilance," the researchers note. No telling who will exercise that vigilance if the person consuming the summary couldn't fully understand the original.

By the way, Google is telling us it's still rolling out across its iOS user base, so if you can't find it, that's why. And sorry, Android users - Google told us that it didn't have anything to share about adding Simplify to "other platforms." Ouch. ®