It is nowadays fashionable to discuss that developments in artificial intelligence would eventually create a world where humankind is either annihilated or enslaved by the robots. As I have discussed before, science fiction authors have recently tilted towards plots featuring a post-apocalyptical world, where science is portrayed as a malignant force than a benign one as imagined by earlier writers. I attempt to examine the issue from the ground up. My conclusion may or may not be correct, but I hope the framework is scientific and objective.
Proponents of the theory that robots would eventually be antagonistic to humans largely believe that the robots will be 'created in our image', and will therefore have similar motivations. Humans tend to be driven by greed, love, fear, and a range of emotions. Our desire to be powerful, to rule over others, and in extreme cases, use violence to establish our supremacy, largely stems from those emotions. To explain what causes those emotions is tricky territory. We can, however, say with reasonable certainty that as a species, the primary purpose of mankind is the survival and replication of its DNA. Many of our emotions are designed around that end-goal.
At a level that is easier for us to relate with, the purpose is maximising our long-term happiness. That is, to maximise the total happiness that we expect to accrue over our lifetime. While people have distinct personalities (one man's food is another man's poison), most "sources of happiness" are common across individuals. A warm hearth or words of praise would be desired by most. It is for this reason that human behaviour is somewhat predictable. Economists expect people to act in their best interests, and Generals anticipate soldiers to put up a defence when under attack by a raiding party.
What would robots be driven by? Consider an algorithm for a turn-based game that a robot plays against a human. A computer cannot distinguish between a win and a loss unless we reduce the principles to mathematical terms. At each stage, the robot evaluates the possible steps and chooses the most rewarding branch in the decision tree. It would aim to maximise or minimise the score by taking the path that leads to the best-case scenario. Depending on the computational power, the computer will evaluate the available combinations, factoring in the human responses (like in chess) in each case, and play the winning move - the one that leads to the best score. What endgame are the robots likely to have in mind?
Unfortunately, this is where things get murkier. The algorithms and the training data that drive the robots would be shaped by the philosophy of the people who design them. Robots are slaves of their algorithms (see pic).
A robot whose algorithm seeks to maximise the number of human lives saved might shoot down a perfectly compliant car driver to save a group of students crossing the road when the signal is red. A robot programmed to value diversity may act favourably towards a Latin American female than a white Caucasian male, even in a scenario where the action is unfair. In that sense, the motivations behind the acts of the robots would ultimately be framed by humans. Except for a scenario where you envisage a mad scientist who deliberately designs a robot whose aim is to annihilate humanity, the undesirable consequences of artificial intelligence are likely to be awkward, embarrassing surprises, akin to errors thrown by buggy code. Another possibility is that well-designed robots start behaving unpredictably just because the algorithms are not dynamic enough to change with the times. As they say, the road to hell is paved with good intentions.
I cannot think of a scenario where robots start acting in self-interest. I tend to believe that Asimov's three laws of robotics would define all robot behaviour, precluding them from ending up as a powerful despot. Theoretically, machine learning enables robots to reprogram themselves, yet as long as they operate under the overarching three laws, I would not consider global domination by machines a plausible scenario. Unless, of course, human coders commit a blunder and circumvent the three laws.
Nonetheless, we can expect robots to commit embarrassing mistakes in the interim, like a machine-driven car colliding head-on with a lamppost, or a medical algorithm making a disastrous diagnosis. Over time, we would have robots that imperceptibly mimic human behaviour (and therefore pass the Turing Test), but without our shortcomings. They might end up being powerful slaves of the few technology giants who have designed them, implementing their ideologies with an iron hand, and amplifying them over the globe. Yet, I believe robots will forever remain subservient to a human or a group of them, and never be the masters of humankind. A Skynet kind of software might cause a nuclear Armageddon, but not by design, purely by accident - a silly coding error or a garbled message.
No comments:
Post a Comment