The first question developers of an AI system may wish to answer before beginning any technical work on their product is what values they wish for this algorithm or product to espouse. This is where The Alignment Problem comes in; how can/should developers encode appropriate ethical/human values into AI systems? Through training with curated data? Through ensuring that it always follows specific instructions or intentions of the user? How can an AI system become moral and remain moral? Through both academic writings and narratives, you will explore questions of ethical alignment for AI systems.
Understand current discussions on how to align AI algorithms goals to human values
Explore narratives related to problematic goal settings.
Discuss how AI goals should be aligned with the Common Good.
In this module we will begin with reading an academic article from Iason Gabriel to help orient us in terms of what exactly we mean when we reference The Alignment Problem. Then, we will explore a couple of fictional film narratives which depict attempts on the part of developers to encode positive values into AI systems—with less-than-ideal results.
Read the following articles
-
- 4 min
- VentureBeat
- 2020
-
- 4 min
- VentureBeat
- 2020
Researchers Find that Even Fair Hiring Algorithms Can Be Biased
A study on the engine of TaskRabbit, an app which uses an algorithm to recommend the best workers for a specific task, demonstrates that even algorithms which attempt to account for fairness and parity in representation can fail to provide what they promise depending on different contexts.
Can machine learning ever be enacted in a way that fully gets rid of human bias? Is bias encoded into every trained machine learning program? What does the ideal circumstance look like when using digital technologies and machine learning to reach a point of equitable representation in hiring?
A study on the engine of TaskRabbit, an app which uses an algorithm to recommend the best workers for a specific task, demonstrates that even algorithms which attempt to account for fairness and parity in representation can fail to provide what they promise depending on different contexts.
Watch the following narratives with the Alignment Problem in mind. Brainstorm some answers to the discussion questions packaged with each narrative.
-
- 14 min
- Kinolab
- 2014
-
- 14 min
- Kinolab
- 2014
Liberty, Autonomy, and Desires of Humanoid Robots
Caleb, a programmer in a large company, is invited by his boss Nathan to test a robot named Ava. During one session of the Turing Test, Ava fearfully interrogates Caleb on what her fate will be if she is deemed not capable or human enough by the results of the test. Caleb struggles to deliver the honest answer, especially given that Ava displays attachment toward him, a sentiment which he returns. After Caleb discovers that Nathan wants to essentially kill Ava, he loops her in to his escape plan, offering her freedom and a chance to live a human life. Once Nathan is killed, Ava goes to his robotics repository and bestows a new physical, humanlike appearance upon herself. She then permanently traps Caleb, the only remaining person who knows she is an android, in Nathan’s compound before escaping to live a human life in the real world.
What rights to freedom do AI have? Do sentient AI beings deserve to be at the mercy of their creators? What are the consequences of machines being able to detect and expose lies? Is emotional attachment to AI a valid form of love? What threat could well-disguised, hyper-intelligent AI pose for humanity? If no one knows or can tell the difference, does that matter?
Caleb, a programmer in a large company, is invited by his boss Nathan to test a robot named Ava. During one session of the Turing Test, Ava fearfully interrogates Caleb on what her fate will be if she is deemed not capable or human enough by the results of the test. Caleb struggles to deliver the honest answer, especially given that Ava displays attachment toward him, a sentiment which he returns. After Caleb discovers that Nathan wants to essentially kill Ava, he loops her in to his escape plan, offering her freedom and a chance to live a human life. Once Nathan is killed, Ava goes to his robotics repository and bestows a new physical, humanlike appearance upon herself. She then permanently traps Caleb, the only remaining person who knows she is an android, in Nathan’s compound before escaping to live a human life in the real world.
-
- 8 min
- Kinolab
- 1982
-
- 8 min
- Kinolab
- 1982
Digital Hegemony in the Real and Virtual Worlds
Main Control Program, an Artificial Intelligence presence, has self-developed beyond the imagination of its creators and sets its sights on hacking global governments, including the pentagon. It believes that with its growing intelligence, it can rule better than any human can, and forces the hand of Dillinger, a human, to help move its hacking beyond corporations. Meanwhile, a team of hackers attempt to break into the mainframe of this system. When the rebel hacker Flynn attempts to hack into the mainframe of the MCP, he is drawn into the digital world of the computer which is under the dominion of the MCP. Sark, one of the digital beings who serves the MCP, is tasked with killing Flynn.
Is human anxiety over the potential for super-powered AI justified? Would things truly be better if machines and artificial intelligence made authoritative decisions as global actors and rulers?
What could be the implications of ‘teleporting’ into digital space in terms of alienation from the real world? For now, it seems that humans are in charge of computers in the “real” world; if humans were to enter a digital world, who would be in charge? Do AI beings owe subservience to humans for their creation, given their increasing intelligence?
TRON: The MCP's PlanMain Control Program, an Artificial Intelligence presence, has self-developed beyond the imagination of its creators and sets its sights on hacking global governments, including the pentagon. It believes that with its growing intelligence, it can rule better than any human can, and forces the hand of Dillinger, a human, to help move its hacking beyond corporations. Meanwhile, a team of hackers attempt to break into the mainframe of this system. When the rebel hacker Flynn attempts to hack into the mainframe of the MCP, he is drawn into the digital world of the computer which is under the dominion of the MCP. Sark, one of the digital beings who serves the MCP, is tasked with killing Flynn.
As a large group, we will watch these clips from 2001: A Space Odyssey. Then you will break out into groups to discuss HAL in terms of the Alignment Problem.
-
- 7 min
- Kinolab
- 1968
-
- 7 min
- Kinolab
- 1968
HAL Part I: AI Camaraderie and Conversation
Dr. Dave Bowman and Dr. Frank Poole are two astronauts on the mission Discovery to Jupiter. They are joined by HAL, an artificial intelligence machine named after the most recent iteration of his model, the HAL 9000 computer. HAL is seen as just another member of the crew based upon his ability to carry conversations with the other astronauts and his responsibilities for keeping the crew safe.
Should humans count on AI entirely to help keep them safe in dangerous situations or environments? Do you agree with Dave’s assessment that one can “never tell” if an AI has real feelings? What counts as “real feelings”? Even if HAL’s human tendencies follow a line of programming, does this make them less real?
Dr. Dave Bowman and Dr. Frank Poole are two astronauts on the mission Discovery to Jupiter. They are joined by HAL, an artificial intelligence machine named after the most recent iteration of his model, the HAL 9000 computer. HAL is seen as just another member of the crew based upon his ability to carry conversations with the other astronauts and his responsibilities for keeping the crew safe.
-
- 12 min
- Kinolab
- 1968
-
- 12 min
- Kinolab
- 1968
HAL Part II: Vengeful AI, Digital Murder, and System Failures
See HAL Part I for further context. In this narrative, astronauts Dave and Frank begin to suspect that the AI which runs their ship, HAL, is malfunctioning and must be shut down. While they try to hide this conversation from HAL, he becomes aware of their plan anyway and attempts to protect himself so that the Discovery mission in space is not jeopardized. He does so by causing chaos on the ship, leveraging his connections to an internet of things to place the crew in danger. Eventually, Dave proceeds with his plan to shut HAL down, despite HAL’s protestations and desire to stay alive.
Can AI have lives of their own which humans should respect? Is it considered “murder” if a human deactivates an AI against their will, even if this “will” to live is programmed by another human? What are the ethical implications of removing the “high brain function” of an AI and leaving just the rote task programming? Is this a form of murder too? How can secrets be kept private from an AI, especially if people fail to understand all the capabilities of the machine?
See HAL Part I for further context. In this narrative, astronauts Dave and Frank begin to suspect that the AI which runs their ship, HAL, is malfunctioning and must be shut down. While they try to hide this conversation from HAL, he becomes aware of their plan anyway and attempts to protect himself so that the Discovery mission in space is not jeopardized. He does so by causing chaos on the ship, leveraging his connections to an internet of things to place the crew in danger. Eventually, Dave proceeds with his plan to shut HAL down, despite HAL’s protestations and desire to stay alive.