SAIL-ON: toward AI that handles the unexpected — a federal contract (USAspending)

1Where open-world brittleness appears

Most of today\'s AI performs well within the range of situations it was trained on, but turns brittle the moment it meets something unexpected — a rule that changes mid-task, an object it has never seen. This weakness toward "open-world novelty" has long been seen as a major obstacle to using AI in the real world. SAIL-ON is a research program that confronts exactly this problem.

2Detecting novelty and adapting to it

The program name unpacks to "Science of Artificial Intelligence and Learning for Open-world Novelty." The core idea is to give AI the ability to detect that something unfamiliar has occurred and to adapt its behavior on the spot. Rather than competing for accuracy under fixed conditions, the goal appears to be understanding, scientifically, how an AI recovers when those conditions break down — its resilience, in effect.

The specific techniques or application areas are not stated in the source description, so we do not speculate.

3Basic research to raise robustness

Viewed across domains, this is not a contract to build a particular product but basic research meant to raise the underlying robustness of AI. AI that holds up against the unexpected matters anywhere the real world refuses to follow the script — self-driving, defense, disaster response. The contract is one indicator of where the United States is placing its bets on the foundations of next-generation AI.

4From detecting to adapting

What SAIL-ON puts at the centre is not raising accuracy but recovering when conditions break.

1. 1An unknown situation arisesRules change midway, an unseen object appears
2. 2DetectRecognise that what is happening lies outside training
3. 3AdaptChange behaviour on the spot
4. 4Hold performanceReach a state that recovers when conditions break

Rather than competing on accuracy under fixed conditions, it seeks to work out scientifically how to recover when the conditions themselves collapse. AI robust to the unexpected holds value wherever the field does not proceed by the textbook — autonomous driving, defense, disaster response. Which methods or application areas are covered does not appear in the source.