American marines clearly showed that sometimes “break” AI can be simple tricks.
During the DARPA experiment, as part of the Squad X program, they had to deceive the observation system trained to recognize people in urban conditions.
The algorithm trained for six days to identify the marines by gait and poses, but the test ended with an unexpected result: all eight participants reached the goal, not being seen.
Someone tumbled for 300 meters in a row, two hid under a cardboard box and went to the sensor, laughing, and one marine was covered with fir branches, turning into a “tree”.
For the system, this turned out to be an insurmountable task – it was not ready for such non -standard scenarios.
Experts call this a “distribution shift”: AI copes well with the situations that he studied, but is lost in a collision with unexpected options.
History reminded the military that artificial intelligence is useful, but human resourcefulness and creativity still remain a decisive advantage.
