Hacker News Reader

Comment by nautilus12 | Hacker News Reader

Wait, so theoretically, suppose that a deep learning algorithm figures out that if they bombard someone with tons of images of other people happy during an acutely sad event of their life that they can drive them to suicide, and that this causes a massive peak of engagement from all the people expressing sympathy on their wall, would it do so? Given the fact that facebook is engineered to be addictive and holds the dopamine button on so many people's brains, wouldn't this practically be manslaughter?

View on HN

IANAL but based on my understanding, manslaughter generally includes some form of negligence (instead of premeditation like homicide) so until our laws and culture catch up to the power these developers/decision makers actually hold, it's unlikely that a jury would blame the developers (or that a judge would even be sympathetic to the argument). Our society has agreed long ago that drinking and driving is irresponsible so thats why drunk driving fatalities are often prosecuted as manslaughter but the same is not true of software engineers and their creations. We need a sweeping cultural change outside of software engineering just like we had with civil engineers in the 20th century.

Nowadays, a professional engineer (a phrase with a precise legal meaning) can be held criminally and civilly liable for negligence even when they are "just following irders." Unfortunately, we're at the very early wild wild west stages of the industry, before some big exposé or calamity completely shakes up the industry (i.e. Sinclair's The Jungle did for food, Carson's Silent Spring for pesticide use, or the 1906 SF earthquake for building codes).

If you build a machine that builds random types of robots but are negligent to make sure it doesn't create robots that kill people, isn't that still criminal negligence?

I think of these algorithms as being algorithms created and vetted out by algorithms, therefore, its negligent to not understand what they are doing. Since its difficult to know what they are doing, then the risk to use them in this way should be viewed as too great

It's the point of machine learning that no human needs (or often even can) understand how the machine is doing something as long as the target metric is optimized. We are building black boxes which translate input into desired output. And tomorrow we will have black boxes building other black boxes.

The only way to ensure nobody is killed is to add "0 people killed" to the target function of the black box.

“World without end” amen

Every heuristic algorithm needs a tollerance

    peopleKilled() < ε

This does not always follow the rational way. If you hack society, and society rejects the hack- you can end up in a concentration camp.

Sounds plausable. I certainly like it as a plot line to a story instead of a sentient AI trying to take over the world that unaware deep learning AI takes over and kills us as it adapts to our preferences of sensationalisim and find those destructive social cycles

Basically at that point AI becomes an echo chamber for our own evil tendencies...until the point it totally drowns us out.

It would make a good novel, but it would be too confusing and depressing, like the Infernal Desire Machines of Dr. Hoffman, but more frustrating and confusing.

Sounds like a great Black Mirror episode. Just imagine if that algorithm was paid for by the cosmetics industry to sell make up and face cream.

The virus kills the host?

Reply on news.ycombinator.com