AI spots shipwrecks from the ocean surface – and even from the air

There's an airplane somewhere in the Indian Ocean that needs finding.

The computer model we created is 92% accurate in finding known shipwrecks.

That doesn't sound very difficult.

I strongly suspect the unstated goal of this effort is to be able to detect one's enemy under water from the surface or the air. The US Navy appears to be testing the technology to do so to locate wrecks. In time that same tech will be used to track perhaps Chinese submarines and unmanned submersibles, maybe near Taiwan or Japan. It might be possible to employ the tech from something like an MQ-4C Triton or P-8 Poseidon, or perhaps from a UUV of some sort.

This is a good use for machine learning, flogged on the box as 'AI', as computers are good at shape matching. Not 100%, but good.

Unfortunately, most of the things that are sold with 'AI' will have a much lower %, as real world incidents are affected by contexts and novel scenarios, which are difficult-to-impossible to computerise.

Also, we have expectations of 100% constancy from computers. Aside from blue screens of death and iffy upgrades, we trust a computer to do exactly what we tell it to do, the same way, every time - the joy of binary. One or zero. The 'learning model' inherent in 'AI' means that your computer may respond differently to the same circumstances over time.

Even simple stuff like stock trading permits only limited 'AI' use. Stocks can be tracked to spot when they are plateauing, but there is no context, such as China banning crypto. Big data repositories are historical. Whilst humanity ought to learn from the past, expecting historical stock movements (or anything else) to be mirrored in the present is optimistic. There is more to stock prices than maths.

92% is quite high, as long as it doesn't matter that much if you get a few false positives or false negatives. But is 92% good enough for surgery (on you), or for driving a vehicle (with you in it)? Planes have autopilots, but they still have two pilots for a reason.

So when you see 'AI' on the box, and there isn't an 'off switch', understand that the software will not be 100% constant or 100% reliable. Maybe 92% if you are lucky. Most likely a lot less. And consider whether a machine really can be trained to replicate human intuition, experience and contextualisation, as time and circumstances change.

Computers can learn some things well - pattern matching, games with fixed parameters. They can also attempt dumb things repeatedly, very quickly, making them look smart. But for many things, a trained human will still be a better option. That may not change. There may be a conceptual glass ceiling to how good machine learning can be.

Computers are great aids, but even in this case, I'd still want a grizzled marine archaeologist to check the scans before I got the scuba gear out. We have adopted the term 'Artificial intelligence' rather lazily. There may be no such things as intelligence, real or artificial, in isolation from a life form that it is a fundamental, dedicated part of.