AI and Keigo: Conquering Japanese Hospitality AI's Greatest Challenge

"お客様が申されました" (The customer stated...)

During beta testing, a restaurant owner pointed something out to us: "Doesn't this honorific sound off?"

It was a sentence our AI had generated. Grammatically correct. Polite enough. But any Japanese person would feel something was wrong—"mousu" is a humble verb, used to describe one's own actions. Using it for a customer's actions is clearly a mistake.

This was the beginning of our battle with keigo.

The World's Most Complex "Politeness"

Japanese keigo isn't just polite speech. Honorific language, humble language, courteous language—and within each category, there are subtle levels of gradation. The difference between "desu" and "de gozaimasu." The nuance between "okoshi ni naru" and "irassharu."

And the hardest part? These all change based on "context."

Maximum respect for first-time customers. Familiar honorifics with warmth for regulars. Formal for business dinners, warm for date nights—even for the same person, the optimal level of keigo changes with the situation.

The Limits of Cutting-Edge AI

At first, we thought we could just use GPT-4 as is. The world's most advanced language model should handle honorifics perfectly, right?

Reality was different.

Double honorifics ("go-yoyaku wo o-uketamawari itashimasu"), confusion between humble and honorific forms, unnaturally stiff politeness that ignores context—the more we tested, the more problems we found.

Why? Because keigo correctness can't be judged by grammar alone. The relationship between speaker and listener, the atmosphere of the occasion, the history of past conversations—only by considering all of these can "natural keigo" be achieved.

Teaching AI Human Intuition

We enlisted the help of 100 native Japanese speakers.

Every week, they evaluate hundreds of AI-generated responses. "Does it sound natural?" "Is it appropriate for this situation?" "Does it feel uncomfortable?"—bit by bit, we quantified what couldn't be quantified.

And we made a discovery. Keigo that all evaluators felt was "natural" had something in common: "just the right amount of distance."

Not too formal, not too casual. Respectful yet warm. We encoded this delicate balance into a three-layer model.

Now, and Going Forward

Currently, our AI achieves a naturalness score of 4.2 and an appropriateness score of 4.5 (out of 5). It's not perfect yet.

But obvious mistakes like that "mousa-remashita" the owner pointed out that day? They don't happen anymore. And most importantly, reports of "feeling something was off" in customer conversations have dropped to nearly zero.

Language is culture. Keigo is the heart of Japan. And AI is still learning its depths.