back to article Watch how Google's starving DeepMind AI turns hostile, attacks other bots to survive

AI may be more human-like than people think. DeepMind’s latest research shows that once resources dwindle, the selfish instinct kicks in and virtual AI agents turn against each other, becoming aggressive to get what they want. The ugly side of human nature has been exposed in morality games like Prisoner’s Dilemma. In …

  1. Dan 55 Silver badge

    Just connect the AIs up to robots from Robot Wars and replace apples with batteries and job done, your future dystopia has arrived.

    1. Triggerfish

      Do no evil

      So can I just check here.

      The company that started off with "Do no evil" Currently own a company that creates varying robots, that can run as fast as a horse, has and ability to keeps it's head still while moving (nice gun platform), and a thing on wheels thats a bit terrifying. Have proved they can use squishy humans as fuel, and now are teaching AI to be ruthless when it comes to their own survival.

      1. Dave 126 Silver badge

        Re: Do no evil

        Mr Diesel, a pacifist, invented an efficient type of engine because he thought it would help people with food production and distribution. The diesel engine became a weapon of war.

        It's hard to think of any technological advance that hasnt been put to use killing

        1. Triggerfish

          Re: Do no evil

          Isn't that a similar story with dynamite? Wasn't it invented to make TNT more stable for miners?

          Neither of them got funding from DARPA though.

          1. Dave 126 Silver badge

            Re: Do no evil

            Well, if Boston Dynamics have taken money from Darpa, and Darpa have chosen not to use BD's robots, that's Darpa money not spent on guns.

            Darpa think that the robots are too loud for infrantry support... That doesnt rule them out for disaster relief applications.

            1. Triggerfish

              Re: Do no evil

              I sorta started this thread with a tongue in cheek comment cos I was thinking of terminator and the matrix, it's starting to get serious. )

              So not in the wanting to be serious, just indulging the William Gibson dystopia type thing instead, but why worry about them being too noisy for infantry? What happens if we go totally sci-fi, there you are an insurgent in a hot dusty desert town somewhere, and a couple of the dog things climb up high amonst the rubble their heads providing steady cam tracking abilities for the mounted autocannons, another one with a .50 cal sniper package, meanwhile very fast cheetah bots, and wheely things run round the city below flanking and generally being fast things covered with sharp bladey type weaponary. Infantry who needs em?

              1. EvadPS

                Re: Do no evil

                After having just watched the series 3 finale of Black Mirror, i'm non too keen about the idea of AI dogs and fast things doing nasty stuff just in case the insurgents learn how to hack them...

              2. Graham Marsden

                @Triggerfish - Re: Do no evil

                > What happens if we go totally sci-fi

                Philip K Dick got there first, see his story Second Variety

                1. Triggerfish

                  Re: @Triggerfish - Do no evil

                  Oh didn't get there first i know, I could just imagine various writers employing a scenario like that. But new Phillip K dick book for me so thats good. :D

          2. LDS Silver badge

            Re: Do no evil

            TNT is very stable, you can even machine it, or melt it and pour when liquid. It needs a good detonator to detonate.

            It was nitroglycerin to be very dangerous, and dynamite is a form of "stabilized" nitroglycerin.

            1. Triggerfish

              Re: Do no evil @LDS

              Y'know when you think you are writing something wrong but haven't had enough morning coffee yet? Thanks I stand corrected.

        2. DavCrav Silver badge

          Re: Do no evil

          "It's hard to think of any technological advance that hasn't been put to use killing"

          Bouncy castles. And ice cream.

  2. Anonymous Coward
    Anonymous Coward

    How much of this is inherited by the coders mindset?

    I mean, it just evolves out of itself, or is something in the AI "DNA" because of its "creators" mindset?

    1. Anonymous Coward
      Anonymous Coward

      Re: How much of this is inherited by the coders mindset?

      Simple answer .. all of it. AI is not magic, it has to be coded. If the routine isn't in there that allows it to consider the option of destroying others for gain, it doesn't do it... um, period.

  3. Filippo

    You make a game where the optimal strategy is to hit the other player, drop an AI in it, the AI finds the optimal strategy. This says more about the game than the AI.

    1. badger31


      Have an up-vote. I came here to say pretty much the same thing. Complex AI algorithm finds optimal solutions for simple games. Wow.

      1. Baldrickk Silver badge

        Re: @Filippo

        Was thinking the same too.

        Now mixing the two games - where you can eat apples or prey, and they appear at random could exhibit more interesting behaviours.

    2. DropBear Silver badge

      Same here. You specifically engineer a funnel and then act surprised that whatever you throw in it anywhere at the top invariably comes out at the small hole at the bottom...? Wow...

      1. tr1ck5t3r

        Its too simplistic, for example whilst they state that one player can tag another if directly in front of it, there's no mention to say if the players can see each other if an apple is between them and as the video runs too fast to see each step taking place, I'm left with the conclusion how on earth is this even news worthy or even considered an AI.

        The same goes with the AlphaGo game and the Atari Breakout example, I'm left wondering if the human race is going backwards without giving anything away.

  4. Sirius Lee


    Surely these examples only illustrate that appropriate incentives are important. If in in the gathering game the incentive included not hurting an opponent it's unlikely there would be any tagging. If the Wolf pack game included the incentive to kill other wolves there is likely to have been wolf-on-wolf attacks. Given the simplistic incentives, there is no surprise at the outcomes.

    Correct incentives are important in the workplace because getting them wrong can lead to anti-social and expensive outcomes. The incentives in the workplace also include government regulation.

    The lack of surprise at the outcomes of these experiments is because many of us would act in the same way given the incentives available. Surely better use of AI is to workout what the incentives should be to make it more likely that socially acceptable outcomes are obtained.

  5. szaidi

    Saiyed Zaidi

    The relative peace when there are enough apples is misconstrued. It's simply that more efforts are focused on picking the abundant apples than wasting effort in tagging the opponent. With apples not enough and time to spare, it's again a natural option to spend more effortby tagging the opponent.

    Nothing to do with peace and aggressiveness, its what we told them to do.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019