AI and ethics Flashcards
Information Technology (IT)
Refers to systems solving problems via digital or analogue technologies, increasingly relying on AI for tasks like communication, data storage, and automation
Artificial intelligence
- making intelligent machines that resemble the human brain and its mental functions
- AI aims to maximize success in a given environment using algorithms and machine learning
- While AI excels in narrow tasks, it struggles with general intelligence (e.g. creativity, social understanding)
Superintelligence
Defined as cognitive abilities far surpassing human intelligence in all domains.
Raises concerns about an intelligence explosion an humanity’s place in a world dominated by AI.
PAPA model
Mason 1986
Privacy
- What information interferes with the private sphere? What information can be revealed to others?
Accuracy
- Who is accountable for errors of your information
Property
- Who owns one’s information?
Access
- Who can have access to your information?
The Turing test
The Turing test is the imitation game:
The machine pass the test if its typewritten answers cannot be distinguished by those coming from a human being
It is a behaviourist test
Chinese Room
Turing test is ineffective because manipulating symbols is not understanding
3 V
Volume - the amount of data
Variety - the number of types of data
Velocity - The speed of data processing
General intelligence
General cognitive intelligence is divided into crystallized (gC) and fluid (gF)
gC
describes the cognitive functioning detected by classic psychometric methods, such as for example IQ tests, and based on previous acquired knowledge available in longterm storage
gF
refers to more flexible reasoning capacities such as creativitym the ability to see new relationships and analogies, to abstractly represent concepts and figures, or to combine letter or number series - all skills which are independent from prior experience and learned knowledge
capability control methods
limit what AI can do
1. Boxing
2. Incentive Methods
3. Stunting
4. Tripwires
Motivation selection methods
limit what AI wants to do
1. Direct specification
2. Domesticity
3. Indirect Normativity
4. Augmentation
7 indirect value learning strategies
- evolutionary selection
- reinforcement learning
- associative value accretion
- Motivational scaffolding
- emulation modulation
- Institution design
principle-based approaches for AI
-beneficence
-autonomy
-justice
- non-maleficence
-Explicability
alignment problem
aligning ai with societal preferences
Control problem
How to anticipate possible dysfunctions or negative outcomes
- capability control methods
- Motivation selection methods
Value-Loading problem
How to provide an AI with moral values
The three laws of robotics
Isaac Asimov (1942)
1) a robot may not injure a human or allow a human being to come to harm
2) A robot must obey any order given to it by human beings, except if they conflict with the first law
3) A robot must protect its own existence as long as such protection does not conflict with the first or second law
Machine learning
computer algorithms that can improve automatically though experience and by the use of data
Tay
Microsoft’s AI chatbot became racist in less than a day
Open ai - California Inst. of Technology
Open Ai develops an unsupervised machine learning system which generates creative paragraphs of texts
A team at California Inst. of Technology trained OpenAI on a reddit corpus and the system generated sexist and racist stereotypes
If we create a world with AI+ or AI++ systems,
what is our place within that world?
- extinction
-isolation - inferiority
- integration
