While it is easy for humans to tell the difference between birds and bikes, computer systems can struggle to do so.

Prize money will be awarded for pictures that can fool image-recognition algorithms, or for new code that cannot be easily fooled.

But pictures of birds riding bicycles will not be allowed in the contest.

Google said it hoped the competition would lead to systems that make fewer errors.

How will the contest work?

People entering the contest as an "attacker", attempting to fool the system, must create an image of a bird that image-recognition algorithms think is a bicycle - or they can create an image of a bike that is misidentified as a bird.

The image can be digital art, a 3D rendering or a photograph. But the image must be easily identified by a human as either a bird or a bike.

Ambiguous images, such as a bird riding a bike, are not allowed.

People can also enter the contest as a "defender" by writing code that does not fall for any of the attackers' images.

If the algorithm misidentifies any of the images submitted by attackers, it is deemed "broken" and eliminated from the competition.

The first algorithm to remain unbroken for 90 days wins a quarter of the total prize fund - although Google says it has not yet decided how much that will be.

How can you fool image-recognition software?

Image-recognition software can be tricked by "adversarial noise" that tricks the computer into seeing something that is not there.

In January, Google revealed a psychedelic pattern that could be printed out and placed in any scene to make a computer think it was looking at a toaster.

Hiding a complex pattern in a bird's plumage is one method that might fool an image-recognition system into thinking it was looking at a bike.

"The objective of this contest is for defenders to build models that will never make confident mistakes," Google said in its announcement.

It said the contest would continue for several months until the prize money was claimed, rather than having a fixed end date.