Machine decision-making Flashcards
(16 cards)
Machine decision-making
Assumption
It is always possible to classify items in specific categories
Machine decision-making
Items classification
Inductive inference
Decision rule based of a set of already classified items
→ A diagnostic test capable to find pattern → Pattern become models for feature decisions
Machine decision-making
From programming to data
Problem of quality of data
2006 Synosets → Imagenet →
2012 → Alexnet
Machine decision-making
Classifying items
Image recognition
- precise data
- namely images accurately labelled
LLM (Large language model)
- machine predict what a given term extrapolated from a text is based from the nearby terms
- Need enormous and quality texts
- Problem with copyright
Machine decision-making
Trained alghorithms
Algorithms need a sufficient large sample to be trained
→ Prevision
Best performance if events, scope and set are defined and measurable
Machine decision-making
Limits of inductive inference
- Statistical assumptions of automatic learning
- Overfitting
Machine decision-making
Limits of inductive inference
→ Statistical assumptions of automatic learning
→ Discriminatory outcomes
(if examples are not representative of the population of interest)
→ All samples are not always generated indipendently and in the same way (Not pure data)
- Variation in data collection process
- Inter-dependent samples
→ Labels should not be understood as possessing a deterministic meaning
- The same sample could correctly be assigne to more labels
Machine decision-making
Limits of inductive inference
→ Overfitting
Rules apply perfectly to the sample but don’t make general rule
→ Good strategy is to introduce INDUCTIVE BIAS
- new assumptions to force machine to express a preference about the type of rule we woluld like it to learn
Machine decision-making
Algorithms to make ethical decisions
If
- extremely delimited
- possible to go back to the information that help us making ethically oriented decisions in similar circumstances
→ then we may allow an AI to make ethically oriented decisions whose extent is delimited
Existing AI with moral consequences (mortgage, loans…)
Pro:
- Quick
- Potentially rational
Cons:
- Data end up damaging everyone
Machine decision-making
Algorithms to make ethical decisions
→ Negative effects
Problematic situations
→ Incorrect or wrong data
→ Incorrect or wrong elaboration
Two different kinds of negative effects
→ On third persons
→ Direct interaction between a customer and an AI
Machine decision-making
Algorithms to make ethical decisions
→ Negative effects
- On third persons
→ Transparency (upstream and downstream)
- process traceability
- opacity
- discrimination
Machine decision-making
Algorithms to make ethical decisions
→ Negative effects
- On third persons
→ Transparency (upstream and downstream)
- process traceability
- Tracking the processes leading to a given choice
- GDPR “right to information”
≠ explicability
- how an AI reached specific conclusions
- intelligibilitt and accountability
Machine decision-making
Algorithms to make ethical decisions
→ Negative effects
- On third persons
→ Transparency (upstream and downstream)
- opacity
→ Extrensic opacity (upstream)
- not all data may be revealed (sensitive)
- if alghoritm is proprietary, copyright
- often developers ≠ distributors ≠ owners
- algorithm users may be bound by professional obligations not to disclose data
→ Intrinsic opacity (downstream)
- often algorithm does not match human decisions
- impossible to attribute responsability
Machine decision-making
Algorithms to make ethical decisions
→ Negative effects
- On third persons
→ Transparency (upstream and downstream)
- discrimination
a) Implicit biases
→ Unjust output outcome often of an unrepresentative sample
- composition not balanced (gender, ethnicity) → social discrimination
- well define application range could create distortion (because reflect inequalities)
→ Other
- Increase isolation
- Damages in small communities
- Bullying
- Exploitation
- Privacy problem
b) Explicit biases
Alghoritm trained with bias
Machine decision-making
Algorithms to make ethical decisions
→ Negative effects
- Direct interaction between a customer and an AI
Excess of trust (Solutionism)
- Indirect outcome
- Petrov
Machine decision-making
→ Conclusion
AI trust and reliable if will
- bring advantage
- not damage them or they interests
- not compromise the authonomy
- Act rightfully
- Explain what they do and why
