Applied verification
Verification is the task of checking AI output and deciding whether it's good or bad — ✓ or ✗.
Because generation is sometimes wrong, for many real world tasks, verification is required to make generation useful.
Verification in the real world
In the real world, automatic verification is used to automatically trigger human intervention.
Triggering human intervention is key to making automation systems work in the real world for important work, even beyond text.
For example:
- detecting ATM transaction fraud
- warning an airline pilot to take over from the autopilot
- sending bad AI translations to professional human translations
- alerting programmers to bad AI code
Example workflow
The original generative task was translation. Translation was also the first task to shift to humans manually checking and fixing AI-generated text, and first task for which AI verification was researched and rolled out in the real world.
In a translation workflow, verification is integrated like this.
Why verification is valuable
For many tasks, manual verification is almost as slow as manual generation. So automatic unverified generation alone is not that valuable.
For example, in translation, checking a machine translation for errors takes almost as long as translating from scratch.
With unverified generation alone, workflows failed to get faster.
By the same token, verification is not as valuable for tasks where manual generation is much slower than manual verification. For example, in image generation, creating an image manually takes hours or days even for a professional, whereas most humans can easily check or choose a good image.
Verification is hard.
Verification is harder than generation, because generation can be wrong a good chunk of the time, whereas verification needs to catch almost all bad output to be useful.
Verification requires human quality or even better. That's one of the reasons that many great AI companies and teams have tried and failed to make verification work in the real world.
Join the mission
Are you interested in joining the mission to accelerate human-quality translation?
Browse jobs at ModelFront