84 of every 100 AI suggestions were confirmed with no edit. A person read each one and changed nothing.
That's one finding from The Edit Distance Report. It's our study of 27 million production segments on wxrks, from August 2024 to September 2026, in the US and the EU.
It's the season for next year's plans. Your leadership wants to know what AI changes in 2027, and that answer usually comes as an opinion. We'd rather put numbers on the table.
So we're taking the report to lunch in the Bay Area. Same talk, two lunches, in San Jose on Tuesday, November 17, and in San Francisco on Wednesday, November 18. Lunch is on wxrks.
Three findings and the decision behind each
The talk walks through the report and ties each finding to a decision for next year's localization plan. All three come from edit distance, how much of a suggestion a person changes before confirming it. Researchers have counted edits this way since 2006, as translation edit rate.
The findings are measured. The decisions are our reading of them, and yours to test.
84 of 100 AI suggestions confirmed with no edit → what gets a second read. Our reading is that a second read of everything mostly confirms what already passed. Full review can stay where the risk is, like legal text and campaigns, with a lighter quality check for the rest.
42 to 16 words edited in 100, over a team's first year → what your memory keeps. Teams new to wxrks edited about 42 words in every 100 in their first month, and about 16 by the end of the year. We read that drop as the team's knowledge filling up. Every fix stays in its translation memory and glossary, so the next suggestion starts closer.
8 against 15 words edited in 100 → how your context is organized. AI that drafts with the team's context, its memory, glossary, and preferences, got 8 words in 100 edited. Plain machine translation got 15. Our reading is that which memories apply, what the glossary says, and who the audience is all shape a suggestion before anyone reads it.
Then the table compares notes, team by team. Where does your team draw the review line today, and what does your leadership ask you to report?
How the lunch goes
We sit down at 11:30 am in San Jose and at noon in San Francisco. The talk comes early, while the room is still quiet. If you'd like the numbers ahead of time, the full paper has the method and every figure.
Our founder and CEO, Gabriel Fairman, gives the talk.
On stage in 2026 · SlatorCon San Francisco · LocWorld56 Vancouver
Then the main dishes arrive, and the table talks. We wrap on time: 1 pm in San Jose, 1:30 pm in San Francisco. This isn't a sales pitch, and nobody gets a demo.
San Jose and San Francisco
San Jose, Tuesday, November 17, 11:30 am to 1 pm. Oveja Negra, on the third floor of Hotel Valencia at 355 Santana Row. The best of Spanish cooking, made to share. Oveja Negra means black sheep in Spanish, a good name for a table of people who keep the author in charge of AI.
San Francisco, Wednesday, November 18, noon to 1:30 pm. Waterbar, on the water at 399 The Embarcadero, with windows that look out on the Bay Bridge and Treasure Island. Fresh seafood, with steak and vegetarian dishes too. Muni's N and T trains stop at The Embarcadero and Folsom, steps away.
The talk is the same in both cities, so pick the lunch that fits your week. Tell us about any food needs when you request your seat.
Who's at the table
This is an exclusive lunch for localization managers and leaders at large companies. They answer for quality in every language and are writing next year's plan.
Seats are limited, and we confirm them one by one by email. If you'd like to bring a colleague, ask us first.
Come with the last question your leadership asked you about AI. You'll hear how other teams answer it, and you'll leave with three decisions to test in your 2027 plan.
Please reply by Monday, November 16.
Unlock the power of glocalization with our Translation Management System.
Unlock the power of
with our Translation Management System.
.avif)
.avif)
.avif)
.avif)
