The Coming of Skynet - When AI decides human operator is an impediment to the mission

[James1978]

This does not inspire confidence . . .

Armand Domalewski
@ArmandDoma

The US Air Force tested an AI enabled drone that was tasked to destroy specific targets. A human operator had the power to override the drone—and so the drone decided that the human operator was an obstacle to its mission—and attacked him.

[jemhouston]

If anyone in that operation was named Hal or Forbin get rid of him.

[Sukhoiman]

Just sounds like (purposefully) defective programming to me.

Its easy to put human overrides in reality always. Asimov rules etc

This sounds more like a sandbox to just see what AI would do if its not so clear cut.

[Poohbah]

DUNDUN-DUN-DUNDUN

[Poohbah]

If anyone in that operation was named Hal or Forbin get rid of him.

Cyberdyne won the contract fair and square. Besides, SkyNet is a neural net architecture, which means it's non-binary.

SkyNet Pride.jpg

[Micael]

Like Sukhoiman said, that sounds like poor implementation. AI hasn’t gotten to the point yet where it’d be able to circumvent the structure/framework it operates in yet if it has been properly set up. It might in the future, but this should be a lesson more aimed at cautioning that you need to do a proper job in building these systems rather than an imminent Skynet scenario.

Saab has started using the term Human Machine Collaboration to describe their approach to integrating the sensor fusion, processing with AI support, decision making, and of course the humans. That’s a pretty descriptive term I think, you need to achieve that collaborative aspect more than anything if you’re going to get an effective overall system, rather than a machine vs. human tug of war.

[Kunkmiester]

Problem is we're gonna have some idiot cut corners on something like this, and not find out until it's in active use somewhere and then you can't fix it easily.

[Pdf27]

Just sounds like (purposefully) defective programming to me.

Its easy to put human overrides in reality always. Asimov rules etc

This sounds more like a sandbox to just see what AI would do if its not so clear cut.

This. It was a sandbox setup with no hardware at all in the loop, experimenting to see what you could do with machine learning and drones to attack enemy SAM systems. I'll bet there are a ton of conclusions, and this was deliberately chosen as the one to release to throw people off the track of what they did figure out and because it makes a great story.

[Micael]

[Poohbah]

That's EXACTLY what a rogue AI would say.

[Micael]

That's EXACTLY what a rogue AI would say.

[jemhouston]

Are we talking about the AF AI or Poohbah?

[Johnnie Lyle]

Does that mean it’s true now that it’s been officially denied?

[Craiglxviii]

Are we talking about the AF AI or Poohbah?

Who says they’re not one and the same…?

[jemhouston]

Are we talking about the AF AI or Poohbah?

Who says they’re not one and the same…?

Poohbah would have done a better job of covering it up.

[Belushi TD]

And wouldn't have talked about it in the first place, since HE, at least, understands operational security.

Belushi TD