Ethics of Artificial Intelligence

Artificial Intelligence: the Third Revolution #

Transhumanism: Using technology to enhance human capacity #

example: replicate vision processing in brain with external computer #

• would render our biological processes inferior/unnecessary
• able to improve/upgrade over time

Eventually, we may have replaced all cognitive processes with computer-based systems (mind uploading)

Question: Would this computer then have a “mind”? #

• not the same as having identical inputs/outputs as a human
• would it genuinely have human perception, or does it merely pretend?

Substrate independence: #

An object has this property if it can be made out of a broad class of substrates (materials)

• functionally defined entity (described by its usage, not physical characteristics)
• example: a chair can be made of many materials, but gold is not substrate independent since it’s defined by physical properties.

Is the mind substrate independent? #

• if yes, then it’s conceivable that a computer can be a mind
• if no, then no AI can truly replicate a human mind

But what do we mean by this? What is a “mind”?

• cognition: decision-making and processing ← Computers already do this. It is substrate independent (Circuits can be implemented in many ways as long as they output correctly)
• consciousness: ??? quite controversial, cannot be defined — what it feels to be thinking

Could an artificial system be conscious? #

Block’s Argument: #

Intuitively, a conscious system is made up of smaller pieces (neurons). However, if you replace neurons with artificial systems that do the same thing, then that collection of systems is not conscious.

• Therefore, the mind is not substrate independent
• objection: you could say the same thing about neurons

Chalmers’ argument of gradual uploading #

Upload each neuron gradually over time. When would we go from being conscious to unconscious?

• Conclusion: brain uploading preserves consciousness
• 3 possibilities:
  1. During process of uploading, one point immediately changes from conscious to unconscious
     • Implausible — can’t determine this point so it probably doesn’t exist
  2. Gradual fade-out: slowly loses consciousness
     • If the functional behavior is identical, then the computer must be deeply out of touch with its own consciousness
  3. Consciousness is not lost
     • Most likely under Chalmers’ argument

Post-human Civilization #

Capable of computation much greater than an unaided human and can run simulations of minds with trivial resource usage.

We are likely to be living in a simulation #

1. • Number of human level civs is large throughout all time/space
   • Nontrivial fraction of them will reach posthuman level
   • At least one will have a need for simulating minds
   • They can simulate an enormous amount of minds, dwarfing the number of actual minds
   • Either we live in a simulation or we will go extinct soon
2. By substrate independence, simulated minds are conscious
3. Weak indifference principle: If X% of observers live in a simulation and you don’t have knowledge of if you are or not, the probability will be X%

★ If this is true, then a God exists

Personal Identity #

• What changes will preserve personal identity? Which do not?
  • Growing up?
  • Teleporting?
  • Mind uploading

→ X at t1 = Y at t1 under what conditions?

• X,Y have same soul? (Cartesian view)
  • surviving death is possible
  • very little evidence / what even is a “soul”?
• X,Y are made of same matter.
  • would suggest aging changes identity since it replaces matter
• X,Y are the same body.
  • bodies can have parts replaced and still be recognizable (eg a bike)
• X,Y have the same brain.
  • what if a brain were separated from its body?
    • brains could be paired with different brains
    • brain transplants?
• X,Y are psychologically continuous
  • gradual change and preservation of memories and thoughts
  • connections between day to day thought processes
  • would support mind uploading as preserving identity

Psychological Continuity #

• invented by John Locke
• naive definition: memory sharing — a future version is psychologically continuous if it has memories of a past version
  • problematic because we often have very spotty memories
• modified definition: we can connect one version with past versions with a chain of mental states
  • remembering past, sharing same values/goals, otherwise resemble very closely

Arguments in Favor #

Chalmers gradual uploading #

example: upload brain very slowly until 100% of brain is transferred to a computer (destructive)

• argument: consciousness is not lost in uploading

← transitivity of identity: if x=y and y=z, then x=z

Conclusion: supports psychological continuity #

• even though body is replaced, consciousness is continuous since all stages are the same
• uploading is enhancement, not suicide

Arguments Against psychological continuity #

• non-destructive uploading: brain not destroyed; upload is a clone
  • intuitively, this would create a replica — suggests that destructive uploading is the same, not identical
• if we materially cloned a person (i.e. teleportation), who is the actual person?
  • A cloned twice then destroyed → B, C

4 possibilities: #

• A = B — not possible, contradictory possibilities
• A = C — not possible, contradictory possibilities
• A = B and A = C — not possible due to transitive property: if A = B and A = C, then B = C?
• A ≠ B and A ≠ C

Conclusion: if this is the case, then A → B creates a new person.

Ethics of AI #

Wednesday, April 22, 2020 11:40 AM

Overview of AI #

An artificial system that can perform tasks associated with human intelligence ← very broad definition

(The concrete techniques most of these ethical concerns apply to are covered in Machine Learning and Neural Networks.)

• Domain Specific AI: can only do one specific task

  • self driving cars
  • chess program
  • …

• General AI: can do all tasks associated w/ human intelligence (“AI”)

  • creativity
  • asks questions
  • draws connections and ideas

• Superintelligent AI: capacities exceed those of humans

  • would be to humans as humans are to other animals like ants
  • believed to be possible in the near future

“AI+” (slightly better than humans) → very rapid increase → “AI++” (dramatically better than humans) → can create other, better AI systems

Dangers of AI #

1. Vague optimization specifications

   • e.g. “make paperclips” ⇒ turns entire universe into paperclips
   • “maximize clicks” → fake news creator

2. AI treats humans as expendable/irrelevant

   • if AI is disinterested in humans, it will not care if humans get in the way

Ensuring human-friendly AI #

1. Leakproof AI

   • build AI in controlled environment
     • impossible to make? (need to be able to observe)
     • AI will convince us to release it

2. Encode human values into AI

   • very difficult to define human values (how can we describe happiness?)
   • leads back to haywire problem — disconnect with actual reality
   • balance stability with the need to learn
     • need to make sure AI doesn’t completely change overnight, but it must be able to evolve and progress too

Moral Status #

Monday, April 27, 2020 — 10:56 AM

An entity has moral status if it has the right to be treated the same way as human/conscious beings. #

• rocks do not have moral status
• do AI’s have moral status?

What has moral status? #

2 properties: #

• Sentience: capacity to experience (especially pain/suffering)
• Sapience: A set of capacities associated with higher intelligence (self-awareness, rationality)

Principle of Substrate Non-Discrimination #

• two beings with the same functionality and conscious experience have the same moral status regardless of substrate

Principle of Ontogeny Non-Discrimination #

• two beings w/ same functionality, conscious experience have the same moral status regardless of how they came into existence

Objection from Existential Debt #

AI’s owe their existence to their creator, who is ethically justified in doing whatever they want (even terminating the AI)

• rebuttal: parents and children/animals

The Automated Economy #

The Second Industrial Revolution #

• traditionally (1st IR) machines replaced humans in physical skills
• currently, machines are replacing humans in cognitive skills

Universal Basic Income #

main idea: tax machine manufacturers and use money to pay humans #

• compensation for loss of jobs
• protect wealthy against populist rage

issue: basic needs will become taken for granted while rich can afford luxuries; those dependent on UBI have nowhere to go

• UBI will not keep people satisfied
  • boredom
  • vice (bad activities)

objection: overestimate the worth of work in human society