Ethical AI

Question 1

VirtuousAI

Question 2

THE ethical question

Answer 2

• what is the right/wrong thing to do?
• who should decide? (researchers, patients clients)

1. legal –
   
   • global & local policy-makers
2. corporate –
   
   • execs & managers
   • designers
   • engineers
3. agents –
   
   • general public (unbiased observers)
   • users
   • effected non-users

Question 3

trolley problem

Answer 3

• you have to kill someone, who do you choose?
  
  • age, gender, fitness, SES, number, species
• crowd sourcing finds what people want but not what they do

Question 4

ethical theory

Answer 4

• moral intuition: “this is right/wrong”
• explicit morality: “one should never do ___”
• ethical theory: “doing X is right/wrong because of ___”
  
  • “AI is bad” is never good enough…

Question 5

fields of ethics

Answer 5

• descriptive (/comparative) ethics: what do people think is right?
  
  • aims to uncover peoples beliefs about values and characteristics of moral agents that are virtuous
• meta-ethics: what does “right” even mean?
• normative (prescriptive) ethics: how should people act?
• applied ethics: how do we take moral knowledge and put it into practice?
  
  • collect facts from various stakeholders
  • make informed decisions

Question 6

metaethics

Answer 6

• what does right even mean?
  
  • epistemics: analysis of meaning of terms
• how do we know conduct is right?
  
  • omniscient, omnipercipient, dispassionate, disinterested observer
• what do we mean when we say researcher is responsible?
  
  • consious, autonomous

Question 7

normative ethics

Answer 7

• the study of ethical behavior or how one might ought to act

1. deontological (duty based): conflict in heirarchy
2. teleogical (end-based): ends justify means
3. non-maleficence: do no harm
4. beneficence: do good
5. justice: respect those is social groups
6. autonomy: respect decisions of the individual

Question 8

are corporations evil?

Answer 8

• corporations are vehicles for ROI
• purpose: distribute wealth

Question 9

AI is evil?

Answer 9

• when automobiles were introduced, they had no regulations
• newspaper articles dubbed automobiles inherently evil
• same applies to AI
• though… it can be perceived as evil in a case where it is used poorly by the “AI is not neutral” argument

Question 10

Asimov’s three laws or robotics

Answer 10

1. no harm to human
2. always obey human
3. defend itself as long as it does not interfere with first two laws

• problems: how do you distinguish harm?

Question 11

problems with ai ethics

Answer 11

1. ignorance: only 18% of datascientists learn AI ethics
   
   • https://thenextweb.com/neural/2020/07/03/study-only-18-of-data-scientists-are-learning-about-ai-ethics/
2. accountability: lacks reinforcement mechanism
   
   • ai ethics wash: everyone creating own AI ethics rules
     
     • but actions show otherwise (Microsoft “against facial recognition” fighting for it)
   • diffusion of responsibility: developers trusting scrutiny from boss, boss trusting developers to make good decisions
3. transparency: closed-door industry settings
   
   1. costly
   2. privacy violation
   3. get funding using “advanced” models

Question 12

high impact tasks

Answer 12

1. Governent (law enforcement, free speech)
2. Workforce (hiring, employment, managing)
   
   1. monitoring user behavior
   2. docking pay for slow work
   3. listening to call center workers
   4. telling people what and how to say
   5. detect deficiencies (optimize business) and dock pay
3. Education (admissions)
4. Healthcare (medical treatment)
5. Finance (loans, social credit systems)

Question 13

trustworthy AI (REGReT)

Answer 13

1. responsible: “human beings” need to exercise judgement in using A.I. (security, privacy, )
2. equitable: should take deliberate steps to avoid bias in A.I. recognition
3. governable: need “the ability to detect and avoid unintended harm or disruption” (a.k.a. not go Skynet on us)
4. reliable: should have a “well-defined domain of use”, track performance and improve results, prove consistency and reproducibility
5. traceable: should be able to analyze (transparent & interpretable) and document the systems at each step of the way to contribute improvements and have accountability.

Question 14

AI for social good

Answer 14

1. communication: bring organizations together to work on breadth of problems
   
   1. cross-sector engagement: bridge gaps between social sciencists and technologists
2. reskill/upskill: close the knowledge gap on AI and data science in the social sector
3. data availability: increase availability of data for social sector
4. scrutiny: need accountability for ensuring proper execution of technology use
5. usability: technology doesn’t just need to be available, it needs to be usable and perfected for the end use case

Question 15

ethical decision making framework (markkula center)

Answer 15

1. identify ethical issues:
   
   1. what values and risks are involved?
   2. who are the stakeholders?
2. get the facts:
   
   1. what do we need to know?
   2. who do we need to hear from?
3. evaluate alternative aactions through multiple ethical lenses:
   
   1. what values do they prioritize?
   2. what harms and benefits will they bring?
   3. to whom?
4. make a decision and mentally test It:
   
   1. what’s the ethical call, based on what we know?
   2. how would it hold up under scrutiny?
5. act and reflect on outcomes:
   
   1. how did it turn out?
   2. what did we learn?

Question 16

ethics in technology practice toolkit

Answer 16

• overview of technology ethics, lenses for perceiving ethical issues, tools for analyzing ethical issues, and cases for practicing the skills of perception and analysis using the lenses and tools

1. Ethical Risk Sweeping: Ethical risks are choices that may cause significant harm to persons or other entities with moral status.
2. Ethical Pre-mortems and Post-mortems: focuses on avoiding systemic ethical failures of a project.
3. Expanding the Ethical Circle: design teams need to invite stakeholder input and perspectives beyond their own.
4. Case-based Analysis: Case-based analysis enables ethical knowledge and skill transfer across ethical situations.
5. Remembering the Ethical Benefits of Creative Work: Ethical design and engineering is about human flourishing.
6. Think About the Terrible People: there will always be those who wish to abuse new powers.
7. Closing the Loop: Ethical Feedback and Iteration: Ethical design and engineering is never a finished task

Question 17

SMACTR audit

Answer 17

• Scoping: risk assesment process; auditors produce assesment of social impact and an ethical review of system use cases
• Mapping: create map of internal stakeholders; identify key collaborators for the execution of audit
• Artifact Collection: creation of an audit checklist as well as datasheets or model cards; assumptions made; intended use
• Testing: asses performance using methods like adversarial training; creates ethical risk analysis chart that identifies likelihood and severity of a failure or risk
• Reflection: auditing and engineering team evaluate the internal design recommendations; create mitigation plan

Question 18

fairness vs. bias

Answer 18

• fairness: statistical parity between two measured groups
  
  • cannot satisfy all definitions at the same time
• bias: when specific (un)privileged groups are placed at a systematic (dis)advantage

Question 19

individual vs. group fairness

Answer 19

• individual fairness: statistical equitability for similar individuals
• groups fairness: statistical equitability amongst groups of people partitioned by protected attribute
  
  • “we are all equal”: all groups have same abilities
  • “what you see is what you get”: observations reglect the abilities of the group

Question 20

fairness definitions

Answer 20

1. equal odds: protected and unprotected groups equal rates of true and false positive
2. equal opportunity: protected and unprotected groups should have equal true positive rates
3. demographic parity: likelihood of positive outcome should be same regardless of protected group or not
4. fairness through unawarness: only fair if unprotected attributes are not explicitly used in decision making
5. counterfactual fairness: decision is fair if it is same in both actual and counterfactual world to which different demographic group belonged

Question 21

protected attributes

Answer 21

• “sensitive” attributes that should have parity (i.e. disability, age, sex, gender, race, cast, religion)
  
  • sometimes it may be illegal to train a model using protected attributes
• (input) privileged value: advantageous feature value
• (output) favorable label: provides advantage to the recipient (i.e. not being arrested)

Question 22

types of bias

Answer 22

1. historical: from historical data (i.e - black people discriminated in judisciary system)
2. reporting bias: bias in who reports
3. implicit bias: bias in people recording the data
4. ​​label bias: annotation process
5. sample/representation: one population overrepresented or underrepresented (I.e - training on white males)
6. feature bias: way we choose, utilize, and measure features such as sex, skin color, or age
7. ommitted variable: an important variable left out which effects models predictions
8. outcome proxy bias: using the wrong model (i.e. - using cost of healthcare as proxy for health)

Question 23

examples of bias

Answer 23

• COMPAS: incorrectly labeled African-American defendents as high-risk nearly twice the rate of white defendents
• Prime Free Same-Day Delivery: rolled out same-day service to highly dense prime aread, which discluded predominantly black area code

Question 24

bias mitigation techniques

Answer 24

1. protected attribute
   
   1. removal of protected features: remove protected variables such as age, sex, race from data
      
      • only works for explicit b/c features often correlated to protected features (I.e. sex to pay)
   2. random assignment of protected features: randomly assign protected variables such as age, sex, race etc… from the data
      
      • should mitigate implicit bias, closely linked to appearance of these variables
   3. counterfactual assignment of protected features: assign individual to counterfactual world
      
      • mitigate implicit bias, identifing source for future
2. Model selection:
   
   1. objective function change: maximize models accuracy and fairness metrics
      
      • adversarial learning: maximize accuracy of predictor on y and minimize ability of adversary to predict the protected or sensitive variable
   2. fair PCA: maintains similar fidelity for different groups and populations
   3. variational fair autoencoder: remove sensitive variable
      
      • uses maximum mean discrepancy regularizer to obtain invariance in posterior distribution over latent variables
3. Post-hoc Analysis:
   
   1. casual inference: graphs and models to study if change in protected variable is casually related to change in target variable
      
      • decisions are irrespective of the sensitive attributes of groups or individuals
   2. debiasing: identify direction of bias and neutralize it using control parameter.
      
      • very popular in NLP where word embeddings must be debiased

Question 25

sampling methods

Answer 25

• sampling with replacement: for minority group
• minority (synthetic) over-sampling: synthetically creates minority classes that are similar to nearest neighbors
• random undersampling: for majority group

Question 26

AI Ethics checkpoints

Answer 26

1. Problem research
2. Solution/system architecture
3. Data collection
4. Model training
5. UX design
6. Product documentation
7. Feedback mechanism

Question 27

classes of interpretability methods

Answer 27

1. moment interpretation task set methods:
   
   1. ante hoc: choosing simple to intepret models (logistic regressions)
   2. poste hoc: accompanied interpretation after already built
2. model specific/agnostic methods:
   
   1. model-specific: i.e. - gradient descent methods work on all neural networks, but tree based methods cannot be applied to any other
   2. model-agnostic: place no assumption on the internal structure of the model
3. scope methods
   
   1. locally interpretable: explaining single prediction
   2. global interpretable: explaining whole model

Question 28

interpretability methods

Answer 28

• partial dependency plot: shows dependence of target variable from a particular feature
  
  • must assume feature independence
• permutation importance: remove feature and see impact on loss
  
  • requires retraining, unless you use noise for feature
  • present distribution as opposed to single value
• shapley value: value to each “player” to demonstrate how important they are to the “team”
  
  • have to predict every combination of every feature under brute force calculation, which is not reasonable resource-wise
• LIME: model-agnostic system to explain local behavior of model around some point X using spase linear model on nearby points
  
  • assume “local suragates” are less complicated than model
  • need to asses the “coverage” of the explanation because kernels only work locally
• Anchor: model-agnost system explaining complex models using high-precision rules called anchors
  
  • precision: the ratio of n times when label hasn’t chaged after perterbation to anchored samples
  • coverage: propbability anchor is present in other samples, or measure of scope covered by explanation

Question 29

red versus green AI

Answer 29

• red AI: big tech focuses on buying incrementally better results
  
  • cost ~ (model)(data)(iterations)
  1. carbon emission
  2. elecriticy used
  3. money spent
• green AI: researchers focus on efficiency and lowering the overall cost
  
  • non-polluting carbon offsets
  • renewable energy sources
  • spend less
• quantifying energy footprint: floating-point operations (FPO)
  • does not apply to particular hardware
  • but… other process are not accounted for

Question 30

windfall clause

Answer 30

• donate some percentage of windfall profits from AI to causes that benefit everyone
  
  • -> such as child abuse research, etc…
• windfall function defines how much money goes into windfall clause

Question 31

important AI policies

Answer 31

1. Privacy
   
   1. GDP: EU rules for goverance of data privacy
   2. CCPA: California laws on data, etc…
   3. Data Protection Act: create new data protection agency
2. AI Trust
   
   1. EU Guidelines of Trustworthy AI

Question 32

hardware recommendations

Answer 32

1. secure hardware: including secure enclaves on commodity hardware
2. cost measurement: report estimated computing cost
   
   1. floating point operations
   2. electricity bill

Question 33

organizational recommendations

Answer 33

1. brand: should clearly define you company ideals and have clear actionables for aligning product, services, and technology with your policies
2. BEFORE EACH PROJECT
   
   1. risk assessment: before each project, should explore risks and stay up to date with best practices, and tools.
   2. Data Sheet: report on the type of data including reasons for collection, uses, etc.
3. DURING EACH PROJECT
   
   1. Model Card: report on the model selected, the reason why
   2. bias and safety bounties: during each project, systems should be implemented to to reward users for finding finding problems.
   3. reporting: also during each project, should dedicate time to properly share information about potential problems and the solutions found
4. ​AFTER EACH PROJECT
   
   1. human-rights assessments: develop, share, and use suites of tools for privacy-preserving ML that include measures of performance against common standards.
   2. data drift: track input differences to make verify following consistent patterns with training
   3. performance changes closely follow changes indicating black swan events
   4. 3rd-party auditing: after each project, should conduct and fund third party auditing of AI systems (for bias, security, etc..)

Question 34

7 traps of AI ethics

_{(https://freedom-to-tinker.com/2019/03/25/ai-ethics-seven-traps/)}

Answer 34

1. Reductionism: Ethics is not about maximizing one metric fo fairness, transparency, etc…
   
   1. You cannot optimize all at once, there is a tradeoff that requires analysis
2. Simplicity: We can just go through verifiable ethical checklist
   
   1. There is always more to it than a binary choice
3. Relativism: There IS such thing as a universal baseline
   
   1. We need to welcome value pluralism without collapsing into extreme relativism
4. Value alignment: there need not be one morally right answer
   
   1. We need to welcome diversity
5. Dichotomy: the goal of ethics is not to become ethical
   
   1. It Isi healthier to see ethics is something you do as opposed to are
6. Myopia: Tradeoffs do not apply to all other contexts of the same technology
   
   1. Each intended use of technology has different required needs
7. Law: laws and ethics are not interchangeable
   
   1. Laws fall short of enforcing ethical values and ethical arguments are not always able to be coded into law

Question 35

black mirror scenarios

Answer 35

1. cerebral technology: blocking people, reliving torturing memories, forced ads, making soldiers see creatures instead of humans, spying, experience others emotions
2. virtual world: having sex with friends alias online, live forever young in cloud, cloning humans for torture in virtual world
3. robot clones: soul-less creature, hacking artificial swarms for mass murder
4. rating people: rating people turns people artificial, makes them snap

Question 36

explainable AI (XAI)

Answer 36

• the field of research primarily concerned with making the interworkings of AI systems more transparent
  
  • opens the “black box” to tell how the neural networks works​, its limitation, and how to use it

Question 37

XAI laws

Answer 37

1. GDPR (Article 22): right of explanation of the inferences automatically produced by a model, confront and challenge recommendation, particularly when it might negatively affect an individual legally, financially, mentally, or physically
   
   1. released in 2016 followed by explosion in XAI

Question 38

why is XAI hard?

Answer 38

1. Multi-disciplinary: CS, Math, Psych, Philosophy of Ethics, + more
2. fragmented landscape

Question 39

why xai?

Answer 39

1. Explain how it reached the answer: people want to know why they are presented with counterfactual and/or counterintuitive events (analytics)
2. justify: why a question is relevant or not (justification)
3. clarify the meaning of the terms used in the system that might not be understood by the users (supplemental questions)
4. teach the user about the domain (education)
5. allowing its debugging and the identification of potential flaws;
6. improve the accuracy and efficiency of their models;
7. extract knowledge and the learning of relationships and patterns.

Question 40

reasons for opaque AI

Answer 40

1. less work preprocessing
2. superior accuracy of opaque systems…
   
   • “advanced” models look shiny
3. security through obscurity: controversial practice to security experts

Question 41

requirements for graphical XAI

Answer 41

1. breadth: most important features should all be visible
2. integrity: there should be minimal data loss in explanation
   
   1. no occlusion of parts of raw image
3. highlight importance: color, annotation, lines, texture, shapes
   
   1. positive and negative attribution to the most and least influential features (e.g. heatmap with causality)
   2. remove noise (e.g. lighten the ink)

Question 42

problems with XAI

Answer 42

• is a struggle with IP and tradesecrets because it requires open-source

Question 43

open-source XAI tools

Answer 43

• reversed time attention model (RETAIN)
• local intepretable model-agnostic explanations (LIME)
• layer-wise relevance propagation (LRP)
• Shapley- (SHAP)

Question 44

types of explanations

Answer 44

• design: reasoning implemented in process of design
  
  • e.g. why neural network used
• educational: convey information contained in the system
  
  • e.g. physical constraints
• fault based: finding cause and solution to fault
  
  • e.g. why did it give unexpected denial of loan
• reconstruction/ontological: problem solving approach based on structural properties (how components are related to eachother)
  
  • e.g. classified as bird because color is most important feature
• operational: “how do I use it?”
• post-hoc: explainaing particular decision and discarding alternatives
  
  • e.g. robot explain that path taken has benefit A and B.

Question 45

attributes of explainability

Answer 45

• Transparent/Intepretable/comprehensible: The capacity to be perceived; open to public scrutiny
• Causality: ability to clarify relationship between input and output
• Actionability: ability of the learning algorithm to transfer knowledge to the end-user to exhibit change
• Correctability: ability to adjust the model
• Sensitivity: The sensibility of the explanations to variation in the input features

Question 46

problem with communicating explainability

Answer 46

you need to “tune” the explanation to the user in question (talking to developers versus managers)

Question 47

requirements for explainability

Answer 47

episodic memory: “remember all of the factors such as states considered during plan formulation to plan revision”

1. why inference made
2. why other inferences not made
3. why inference was the best one
4. why other inferences cannot be made
5. why to revise or not plan

Question 48

structures of explanations

Answer 48

1. dialogs: rule based question and answering
   
   1. Intro. (We are about to…) -> explanation -> outdo. (…Thanks)
   2. Q’s: How, Why, and What
   3. Users can interrupt simulation and select a query about what happened at a particular step
      
      1. Information pulled up from relational database
2. Plots: visual explanations are best
   
   1. Heat maps of influence on inference

Question 49

agnostic numeric methods

Answer 49

• Shapley (SHAP): (popular) uses additive feature attribution methods (linear combinations of input features) to build a model which is an interpretable approximation
• Explain and Ime: asses contribution of input or group of variables be replacing the values with other sampled an measuring mean differences in probability score
• More variation = more importance
• Global Sensitivity Analysis (GSA): ranks inputs by quantifying effect of variable by putting through their range of values
• Gradient Feature Auditing: partial occlusion of the input
• Monotone Influencing Function: rotating, reflecting or randomly assign labels to the input

Question 50

agnostic rule-based methods

Answer 50

• symbolic logic formed of antecedents and consequences
  
  • e.g. mapping IF-THEN statements using fuzzy logic or decision trees as precedents and outputs as antecedents

1. Genetric Rule Extraction (G-REX): uses genetic algorithms to generate IF-THEN rules with AND/OR operators
2. Anchor: uses two algorithms to extract IF-THEN rules highlights features of an input that are sufficient to make inference
   
   1. Creates empty rules and adds a rule for each feature predicate on each iteration
   2. Selects set of all rule sets with best precision
3. Partition Aware Local Model: uses decision trees to approximate completion model
   
   1. Uses two-part surrogate model: a DT model partitioning the training data and a set of sub-models fitting the patterns within each partition
   2. Users have to examine the DT’s to see if they make intuitive sense
4. Model Extraction: generate decision trees using Classification & Regression Trees Algorithm (CART) and trains over mixture of Gaussian distributions fitted to the input data

Question 51

agnostic graphical methods

Answer 51

1. Salient masks: (popular) superimpose highlighting mask on images and text of importance
2. Layer-Wise Relevance Propagation (LRP): uses pixel-wise decomposition of nonlinear classifiers assuming that the classifier can be decomposed in to several layer tracing each pixel contribution to the final output, layer-by-layer
   
   1. Visualized as a heatmap
3. Spectral Relevance Analysis (SpRay): spectrally clusters LRP (see above) explanation in order to identify typical and atypical decision behaviors
   
   1. Checks if highlighted region contains the animal, etc…
4. Image Perturbation/Restricted Support Region Set(RSR): creates salience maps by blurring different areas of the image and checking which ones most effect the prediction accuracy
   
   1. If certain regions being removed the image is misclassified
5. IVisClassifier: reduces input dimensions and produces heat-maps to show relationship among clusters in terms of pairwise distance between cluster centroids
6. Individual Conditional Importance (ICE): line charts graphic relationship between predicted accuracy and a feature, keeping all other features fixed.
7. Partial Importance (PI): visualize the point-wise average of all ICI curves across all observations; assessed using Shapley feature importance test

Question 52

agnostic mixed-methods

Answer 52

1. Bayesian Teaching: selects a small subset of prototypes that would lead to same answer as if trained by all
2. Set Cover Optimization: selects as few prototypes as possible to capture structure of training all classes
3. Evasion-Prone Samples Selection: detects instances close to the classification boundary
4. Maximum Mean Discrepancy-Critic: uses maximum mean discrepancy and an associated witness function to identify portions of the input most misrepresented
5. Pertinent negatives: highlights what should be minimally and necessarily absent to justify the classification of an instance (e.g. - no glasses means good eyesight)
   
   1. Remove certain parts of the image to see what is

Question 53

prototypes versus adversaries

Answer 53

iconic versus contrastive examples of the data that lead to high versus low performance

Question 54

example “white-box” models

Answer 54

In contrast to black-box models, the following are self-interpretable

1. decision tree
2. linear
3. bayesian network
4. fuzzy cognitive maps

Question 55

XAI feature importance

Answer 55

usually through perturbations the input or occluding parts of the input and seeing the effect on performance

1. Shapley uses game theory to reconstruct model without various/all combos of input
2. Visualized by highlighting/blurring the most/least relevant input (usually through heat maps or highlighting text)
   
   1. numerical - heat map of feature
   2. text - highlight importance words
   3. images - heat map of pixels or set of pixels

Question 56

XAI salience maps

Answer 56

methods include DeepLift and LRP

1. Calculate the average of all gradient of all paths to output
2. Measure effect of occlusion on performance of various inputs or collection of inputs

Question 57

kernel SHAP method

Answer 57

Question 58

deeplift

Answer 58

Question 59

correlation vs. causality

Answer 59

• prediction tasks: goal is to predict an outcome given a set of features
• causality tasks: want to know how changing an aspects affects an outcome
• do not be tricked
  
  • data driven predictive models show correlations, not causations!
  • As opposed to knowledge-based which encode experts knowledge in rule based logic

Question 60

confounding variables

Answer 60

• contribute noise to predictions causing non-causal correlations
• if they are there but unidentifiable, SHAP will not work
• can address using methods such as double ML

Question 61

non-confounding redundancies

Answer 61

• cause problems with interpreting causality
• controlling for feature can help determine redundancy

Question 62

doubles/debias machine learning

Answer 62

• de-confounds the feature of interest and estimates the average causal effect of changing that feature
• packages include EconML or CausalML

1. Train model to predict features using confounder candidates
2. Train model to predict outcomee using same set of confounders
3. Train a model to predict the variation of the outcome (y2 - y1)

Question 63

why are people not using XAI?

Answer 63

1. less work preprocessing
2. superior accuracy of opaque systems…
3. “advanced” models look shiny
4. security through obscurity: controversial practice to security experts

Question 64

open-source fairness toolkits

Answer 64

• Fairness Measures: test algorithm on variety of datasets and metrics
• Fairness Comparison: directly compare algorithms fairness metrics, with raw and preprocessed datasets
• Themis-ML: (built on sklearn) fairness aware machine learning algorithms
• FairML: quantify inputs prediction dependence on inputs
• Aequitas: generates bias report for given model and dataset
• Fairtest: Tests for correlation between outputs and protected population
• Themis: designs failure mode group-based or causal discrimination
• Audit-AI: (built on sklearn) statistical tests for classification and regression.

Question 65

why is bias mitigation difficult?

Answer 65

protected features are correlated with unprotected (e.g. redlining poor is also redlining black)

Question 66

determine label bias using SHAP

Answer 66

use protected attribute as an input to SHAP and see if contribution moves to it

Question 67

“explaining” fairness

Answer 67

the sum of the demographic parity differences of the SHAP values sum up to the demographic parity difference of the whole model

• This is b/c the sum of the SHAP values for each feature sum up to the models output
• In identifying disparity, it is important to figure out how to debias
• Under/over reporting label will not effect demographic parity difference if the the input are not correlated to the demographic
  
  • A.k.a - the bias is unobserved
• Having two bias effecting demographic parity can offset each other in which case only the SHAP will show

Question 68

Foundational Models Framework

Answer 68

• Even if no risk of a particular foundational model today, there could be one tomorrow
• Severity of risks are unclear and would benefit from sharing best practices
• Collectively addresses problem of underinvesting in responsible AI
• Needs to be done by diverse independent party to weight risk and benefits

Question 69

Foundational Model Releases

Answer 69

Release: doesn’t have to be all at once

1. What? Direct access (models, weights) or Indirect access (Papers, models, code, training data, computation resources)
2. Whom? Colleagues, press announcement, general public
3. When? Should be released depending on safety evaluations, externals conditions or should broaden with the what and whom access
4. How? Maintain releases over time Update if notifiers of updates

Question 70

Foundational Model Review Board

Answer 70

serve as entity of review to provide feedback about release and best practices

• Must be strong incentives for all researchers, developers, and board members
• Provide mechanism to request help, relieving developer the sole responsibility
• Broaden access to foundations models in a responsible, highly contextual way

Question 71

Foundational Model Review Board Process

Answer 71

1. Developer calls a proposal describing what foundational models are available and believe to be most critical
2. Researchers submit proposal specifying the research goal and access required of planning and managing risks
3. The board reviews the the research proposal with the developers
4. Foundation model developer makes decision to approve, reject, or defer the proposal
5. If approved, the foundation model developer releases the assets

Question 72

Foundational Model Research Proposal

Answer 72

2-3 page report submitted to review board to access assets to study model

1. Explain goals of research into the foundational model
2. Why the research is important, and what would be learned
3. Types of access required
4. Budget cost required to conduct research

Question 73

Foundational Model Transparency

Answer 73

tradeoffs between transparency and confidentiality

• Whether proposals, reviews, decision be made public should be up to the board of reviewers with preference to transparency for non-compelling reasons

Question 74

prejudice in word embeddings

Answer 74

• word usage, which defines meanings, forms the basis of word embeddings
• prejudice is in the use of words, or representations, and is learned from context of wording.
• IAT study - representations are highly correlated with real facts which means “stereotypes are just regularities that exist in the real world that society has decided we want to change”
  
  • however, some are objectively false (like the left side of our bad is bad)