Now Reading
AI Does Not Hate You, Nor Does It Love You: Understanding The Value Alignment Problem

AI Does Not Hate You, Nor Does It Love You: Understanding The Value Alignment Problem

Sujoy Sarkar

Eminent British poet, Tom Chiver, in his book titled, The AI Does Not Hate You: Super-intelligence, Rationality and the Race to Save the World, talks at length about several reasons which could simulate the route of human extinction. Other than a terrifying looking AI, similar in appearance to the T-800 from Terminator II, one of them is the creation of genetically engineered viruses through the means of synthetic biotech. Overgrowing concerns with the latter, goes unnoticed with non-availability of sustainable investments for commencing any kind of research. 

The dystopian epoch, we all are experiencing should primarily raise alarming signals at our defunct heuristics and the failure to adopt holistic policies for a genetically engineered pandemic. As a matter of consideration, the validity of my opinions shall be realized, if we could presume that we are going to overcome the COVID-19 pandemic, otherwise this work is as good as a fictional output.

The problem of value alignment with augmented technological tools gained ethical relevance with Isaac Asimov’s Three Laws of Robotics. The laws which are now 70 years old, were able to manage effective levels of control in artificial agents like Bertie the Brain, capable of tic-tac-toe game in the 1950s. Fast forward to 2016, AlphaGo, beats Lee Sedol in a game that has more possible positions than the number of atoms in the universe, merely powered with deep learning algorithms. 

With unsurmountable development in the domain of artificial intelligence over the past two decades, the same laws appear to have lost their charm to maintain robust value aligning rigidity between the machine and us.  Consistent research backed by computational power, in the field of artificial general intelligence shall evolve the concept of intelligence in the near future. Hypothetically a machine like AlphaGo might not only be good at playing Go, it may extend its capacity of purpose in other domains of operations, like policy making.

Amidst such a transitory phase, the evolution of intelligence in artificial agents will challenge the systemic inheritance of control over them. It is a concomitant responsibility of its creators to ensure that losing control over its agents shall never lead to incorrigible consequences for us. This ideology is the foundation of the value alignment problem; in layman’s terms, we must be sure that the purpose desired by the creator is the same purpose an artificial agent seeks to achieve. The central logic revolves around artificial agents being made to act in altruistic ways so that it can never pose a threat to us. However, despite having an altruistic purpose, the agent can make decisions, backed with general intelligence, which will end up contradicting the purpose envisioned by the creator.

This shall be understood better if we take a glance at the paperclip maximizer experiment. The experiment was brought in the discussion forum by Swedish Philosopher, Nick Bostrom, stating how an artificial general intelligence agent tasked with the objective of collecting paper clips, in order to maximize its collection will eventually turn earth into a paper clip.

An artificial agent with human-level intelligence, initially, will take it upon itself to either collect those paper clips, buy those paper clips with money or manufacture them by itself. 

Inevitably, the agent shall go through a process of intelligence explosion, where it shall deduce newer and more effective methods of collecting paper clips. The agent shall never move away from its goal of collecting paper clips and in doing so, it will not be bothered about non-specific goals such as learning human sentiments or developing collective unconsciousness. Over here, we conclude that the values of an artificial agent are completely abnormal and alien to the multitude of values central to us. Therefore, in its process of optimization as a goal-seeker it will be exactly focused on the aspect of maximizing utility, which is the collection of paper clips in this instance. 

As an anthropological discourse, the value alignment problem, if viewed from the prospects of Asimov’s three laws of Robotics, appears to formulate a shield with respect to an existential threat to us by the acts of an artificial agent. This premise, irrespective of the nature of control on artificial agents, will cause serious hindrance in the practice of developing artificial agents that can co-exist with us and serve its purpose of being an augmented tool to improve our experience of existence. Beyond the norms of existential threat, the value alignment problem today, must focus on inculcating values in an artificial agent which shall not only cause an existential threat but it shall never be a cause of obstruction for us to flourish in our lifetime.

Therefore, the duty of the creator of an artificial agent is not over when it has acted in all good faith, inheriting altruistic values while devising its purposes. The duty extends to making the artificial agent aware of our behavioral competencies, so that the agent doesn’t intervene with our ability to prosper in its presence. When a social media platform mechanizes machine learning tools to scrap your private data, for targeted advertisement purposes, it might not be causing a direct existential threat to your identity but it is surely in violation of your freedom to flourish. 

Due to our heuristically challenging abilities of decision making and abysmally large population, it is important for the artificial agent to keep learning the values it should align with and not be limited to a static accord of values.

The presence of malign information from media portals, reflecting this paradigm as a war between us and artificial agents has already penetrated a pleading for greater control over these agents. However, the determinative motive shall be to accept its autonomy and secure a medium of co-existence by making them align with our ever thriving values. The process of automation can’t be rejected and therefore, any efforts to threaten its legitimacy by framing us with the idea of feeling oppressed will cause a blockade to the beautiful destiny of augmented reality. Honestly, there’s no chance of competing on a levelled pedestal with them.

See Also
How Government Of India Used Conversational AI During COVID 19: A Case Study

A biological neuron fires, maybe, at 200 hertz, 200 times a second. But even a present-day transistor operates at the Gigahertz. Neurons propagate slowly in axons, 100 meters per second, tops. But in computers, signals can travel at the speed of light. There are also size limitations like a human brain that has to fit inside a cranium, but a computer can be the size of a warehouse or larger. So the potential for super intelligence lies dormant in the matter, much like the power of the atom lay dormant throughout human history, patiently waiting there until 1945.


Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, Oxford Publication, (2014).

Stuart Russell, Human Compatible: AI and the Problem of Control, Viking Press, (2019).

Stuart Russell, Peter Norvig, Artificial Intelligence: A Modern Approach, Pearson Education, (2016).

Tom Chiver, The AI Does Not Hate You: Superintelligence, Rationality and the Race to Save the World, Orion Publishing Group, (2019).

What Do You Think?

Subscribe to our Newsletter

Get the latest updates and relevant offers by sharing your email.
You can write for us and be one of the 500+ experts who have contributed stories at AIM. Share your nominations here.

Copyright Analytics India Magazine Pvt Ltd

Scroll To Top