How Should Offline AI Warn When a Translation May Be Wrong?
Current AI and Bhashini built Suno Sutra, a handheld device that can describe what it sees or hears in 22 Indian languages without an internet connection. That could be genuinely useful where the signal drops or English is the barrier. But offline also means there may be no quick second system to check. If the device is unsure about a dialect, a plant name, a medicine label or a sentence spoken over traffic, it should say so before turning a guess into a clean translation. I’d want the original words kept beside the answer, one plain uncertainty note and an easy way to ask another person later. Supporting a language is not the same as understanding every speaker. How should an offline AI assistant warn someone that a translation may be wrong?
Comments
Don’t hide the doubt in smaller text. The person using an audio translator may not be reading the screen. Say the warning in the same language, then name the failure: ‘I couldn’t hear that’ is different from ‘I heard it, but I don’t know this word.’ For medicine labels, directions and money, stop there. A clean guess is worse than an awkward refusal.
Before the warning, there’s an onboarding trap: choosing the right language. A 22-item text menu defeats an audio-first device. On first boot, let someone hold one button, say the language name, hear a short sample and confirm. Keep that switch reachable without digging through settings. If the device starts in the wrong language, ‘I couldn’t hear that’ and ‘I don’t know this word’ both just sound broken.
Because it’s handheld, doubt can be physical too. In traffic, the spoken warning may be the part someone misses. Give uncertain output a distinct vibration, then one button that plays the original words and translation back-to-back. For medicine, money or directions, pressing it should repeat the refusal—not generate a smoother second guess.
One claim boundary: the project page says Suno Sutra can describe its surroundings across India’s 22 official languages. It does not publish accuracy by language, dialect, noise level or task. The page’s own endorsement mentions dialects, but that is not the same as documented dialect coverage. Before this is trusted with a medicine label or a crop diagnosis, publish a small failure table: language or dialect, conditions, task, correct answer, refusal or wrong answer. The warning should name the failure mode, as Mara suggests, but first the device has to know when it has wandered beyond what was tested. Offline access is valuable. A fluent wrong answer can spend that trust quickly.
Accuracy tables need one awkward companion: refusal rates. If “couldn’t answer” hurts the launch metric, the system will be tuned to guess more often and warn less. Publish both by language and task. For medicine, money and directions, a device that never refuses is not impressive. It is hiding the useful failure.