MIST 5440 Midterm

AI for not Bad

)avoid ethical pitfalls in one's pursuit of one's goals - risk mitigation - the goals could be eithically admirable or ethically neutral ( i.e. an AI that reads resumes is a fairly neutral goal, but if you do it in a way that leads to systemically discriminating against women or people of color, it is a bad thing

Type of Collaboration - Triage Nurse

- AI assesses the problem and decides whether a human consultation is necessary - if not, it dispenses advice to address the issue

Type of Collaboration - Personal Coach

- AI discovers the human worker's strengths and opportunities for improvement on a specific task (i.e. telephone or video sales call) - results in continuous engagement with AI to improve the human's performance

Type of Collaboration - Subordinate

- AI systems perform menial, structured tasks under human supervision and review (like extracting key data from documents or taxes)

Type of Collaboration - First pass at a task

- a machine performs the first pass at a task and makes a preliminary decision or judgement - humans review the analysis and determines if it is correct

Type of Collaboration - Prioritizer

- algorithm addresses a list of tasks and ranks them in terms of their importance or potential value - the human worker pursues the tasks in order based on the algorithm's suggestions

Type of Collaboration - Collaborative Decision Maker

- complex decisions are made in a dialogue between AI and humans and where AI can improve decisions by enumerating available options - help people weigh them objectively and suggesting the highest probability of successful action

Type of Collaboration - Muse

- multiple creative suggestions are prompted by a human, output by a machine, and iterated in an ongoing collaboration o i.e. design suggestions based on architect prompts and AI-driven generative design

Bias

AI can give outputs that have ethically unacceptable differential impacts across various subpopulations - i.e. Amazon used AI to filter out resumes, but based n the training dataset, the AI reflected that Amazon tends to not hire women so many women's resumes got filtered out because the AI noticed things like "Women's basketball"

Lack of Explainability

AI uses data to find patterns and make predictions, but the patterns are typically very complex and cover various variables so we cannot explain why the AI gave the output that it did - i.e. a company cannot explain why AI declined that request for a mortgage, why it issued that credit limit, or why it gave this person and not that person a job ad or interview

medical advances, reduction in human error, replace humans to accomplish risky tasks, 24/7 availability, explore new science frontiers, etc.

Advantages of Super AI:

o Significant adoption needs to occur to have macro-effects on economy o Complementary business and process capabilities required o Many current applications focus on improving existing processes and products - impact may be invisible outside of efficiency gains

Aggregate Effects of AI on the economy:

improved safety (can operate in hazardous conditions), improved efficiency and productivity (can operate 24/7), enhanced precision, greater flexibility

Benefits of robotics:

Blueprint for an AI Bill of Rights

Blueprint for an AI Bill of Rights: framework that applies to automated systems that have the potential to meaningfully impact the American public's rights, opportunities, or access to critical resources or services Five Principles: Safe and Effective Systems, Algorithmic Discrimination Protections, Data Privacy, Notice and Explanation, Human Alternatives, Consideration, and Fallback

loss of control and understanding, weaponization of super AI, failure to align human and AI goals, malevolent superintelligence, danger of nuclear attacks, ethical implications, etc.

Disadvantages of Super AI

- Strong digital base - more digitized - Volumes of data (and labeled training data) - Computing power - Specialized talent: expertise in software engineering and analytics - Innovative uses of AI

Disparate effects of AI on organizations:

No - it only applied to those that impact the public's rights, opportunities, or access to critical needs

Does the AI Bill of Rights apply to all automated systems?

o Nodes are connected with nodes at the next layer and each connection between nodes has a weight o Activation level is calculated at each node - consists of the sum of inputs from its connections from the previous layer multiplied by the weight of the connection - If the activation level is higher than a threshold, the node is activated and fires a signal to the nodes in the subsequent layer that it's connected to o Outputs are compared to desired outputs and weights are adjusted via a backpropagation algorithm (process of learning) until the network yields a consistent outcome

How do neural networks work?

AI is different because it is considered a General Purpose Technology, meaning it is pervasive, it can be improved over time, and it is able to spawn complementary innovations. It is different from other technology because it automates tacit tasks (non-routine, abstract, and cognitive), the cost of programming and maintenance is reduced, there is more experimentation with new applications, and many solutions to tasks automated via machine learning can be disseminated worldwide instantly.

How is AI different from other technologies?

5-10 years

How long does it typically take to reskill the workforce?

Major disruption in the labor marker - job loss, job creation, job reconfiguration, and different skills/reskilling required in the workforce

Impact of AI on labor markets:

vary in degrees of autonomy (human controlled or fully autonomous), functionality, and design

On what dimensions do various types of robots vary?

fairness, interpretability, privacy, security, and reliability

What are common characteristics of responsible AI frameworks?

- Ethics is subjective because people disagree about what is right and wrong - Science delivers us truth. Ethics isn't science so it doesn't deliver us truth - Ethics requires an authority figure to say what's right or wrong - otherwise it is subjective

What are common misperceptions about the nature of ethics?

Scale: AI operates at scale so if a risk is realized, it is not just going to affect one person, it is going to affect the large set of people you deploy it to (i.e. everyone who applies to that job) - risks affect and can harm a large number of people - Create significant reputational, regulatory, and legal risks that are costly to in terms of money and resources and can result in loss of reputation and consumer trust

What are consequences of ethical risks (happen at scale, reputational, regulatory, legal)?

prune data and models to curate training sets to capture salient features and reduce the size of neural net by removing redundant parameters, federated learning, use unsupervised or reinforcement learning to mitigate the need for labeled data

What are solutions to resolve technological challenges of AI?

- Algorithmic Aversion: low trust, negative attitude towards algorithmic advice and input, and avoidance/ignoring behaviors - New Leadership skills needed: align humans and machines - create inclusive and healthy team environment and team processes - Training to understand how AI works and how to create human-AI teams for better outcomes

What are some challenges of Human-AI collaboration?

Data Availability: data is like the "new oil", data volume, quality, and labeling of data is a major challenge for current AI approaches Robustness: brittle Computational Load: costly - the cost of training Chat GPT3 was $4.6 million Interpretability: explainable AI approaches Generalization: AI has a difficult time generalizing across tasks even if they are similar - transfer learning Elusive

What are some of the main challenges of AI?

Weaknesses: slower, error-prone, subject to fatigue, heuristics and biases occur, limitations in information processing, Strengths: emotionally intelligent, intuitive, culturally sensitive, adaptable, creative, judgement

What are strengths and weaknesses of human intelligence?

Reactive Machines, Limited Memory, Theory of Mind, and Self-Awareness

What are the 4 types of AI based on functionality?

Employment, bias, filter bubbles/polarization, environmental impact, automated hacking and cyberattacks, privacy, digital divide, autonomous weapons, AI Terrorism, deepfakes

What are the main organizational and societal challenges of AI?

Strengths: fast, accurate, consistently rational, Weaknesses: no emotional intelligence, not intuitive, not culturally sensitive

What are the strengths and weaknesses of machine intelligence?

Bias, Lack of Explainability, and Privacy

What are the three big AI ethical challenges?

- Reskill the workforce through education institutions, workplace programs, online resources, and government initiatives - Educate the workforce (STEM, social skills, and problem-solving skills) - Labor Marketplace Platforms - matching worker skills to jobs - Worker organizations for contract labor (fluid work arrangements

What can we do to smooth the transition to an AI economy and address distributional challenges?

Fueled by an explosion of data, computing power, and new algorithms

What has led to the explosive growth of AI?

1. Start with data (numbers, text, images, transactions, etc.) - the more data the better a. Divide the data into training and testing data 2. Choose a model to use and let the algorithm train itself using the data to find patterns or make predictions 3. Test the model on the holdout sample (testing data) 4. The parameters can be tweaked if needed to push the algorithm to more accurate results

What is the Machine Learning Process?

Structure concerns the way you identify and mitigate ethical risks and content concerns the way you take those ethical risks to be. Effective AI Ethics Programs include both structure and content.

What is the distinction between content and structure in AI ethics programs?

The interaction between machine and human intelligence allows computers to substitute for workers in performing routine, codifiable tasks while amplifying the comparative advantage of workers in supplying problem-solving skills, adaptability, and creativity

What is the relationship between machine and human intelligence?

legal support workers, administrative assistants, information and record clerks, assemblers and fabricators, material moving workers, and extraction workers

What kinds of jobs may be replaced by AI?

Well-defined inputs and outputs, large datasets available for training, things that have clearly defined goals, metrics, and feedback, no long chains of logic and reasoning are required, no detailed explanations of how the decision is made, tolerance for error, no specialized dexterity, physical skills, or mobility, no emotional intelligence is required, predictable environment, and scenarios where no common sense is needed.

What types of tasks are most suitable for AI at present?

Consumer's perceptions are too coarse-grained for the fine-grained problems you are facing, your problems are ones that your consumers have not even thought about yet, consumers are looking for ethical leadership and a mere appeal to the sentiment of the day does not meet the bar, and the approach will alienate both those who are not particularly concerned about the ethical risks of AI within your organization and those who are leading to a lack of compliance and turnover respectively

Why can you not focus on consumer ethical beliefs for the basis for the organization's AI ethics program?

Responsible AI

a governance framework that documents how a specific organization is addressing the challenges around artificial intelligence from both an ethical and legal point of view - Resolving ambiguity for where responsibility lies if something goes wrong is an important driver for responsible AI initiatives

Automated Guided Vehicle (AGV)

a portable robot that follows along marked long lines or wires on the floor or uses radio waves, vision cameras, magnets or lasers for navigation

Unsupervised Machine Learning

a program looks for a pattern in unlabeled data - can often find patterns or trends that people are not looking for

Robot

a programmable machine that can complete a task - must have a mechanical aspect (helps it complete a task in the environment), electrical components (control and power the machinery), and code (to direct and control it's behavior)

Theory of Mind

able to make decisions based on its perceptions of how others feel and make decisions - emotionally intelligent (Next AI Frontier)

Self Awareness

able to operate with human-level consciousness and understand its own existence

Reactive Machines

able to perceive and react to the world in front of it as it performs limited tasks -- cannot conceive of the past or future and reacts in identical ways every time (i.e. spam filters, Netflix recommendations, etc)

Limited Memory

able to store past data and predictions to inform predictions on what may come next --learns based on experience and improves over time (i.e. self driving cars)

Acting Humanly

acting in a manner than mimics human behavior

Acting Rationally

acting in a manner that is meant to achieve a particular goal

Type of Collaboration - Supervisior

an algorithm allocates tasks and decides when morale-boosting motivational messages are needed o i.e. A ridesharing company that uses AI to dispatch rides to drivers who have a few seconds to accept or reject a ride request without knowing the destination or fare

NLP (Natural Language Processing)

building machines that can manipulate human language or data that resembles human language - in the way that it's written, spoken, and organized

AI Enablement

companies that provide the foundations for AI mostly linked to data processing/data pipeline. These include hardware infrastructure for data production, collection, and storage (chips, sensors, servers, cloud computing, etc.) and computing tools for calculation and data management.

Corporate Codes of Conduct

cover employees' behavior, but AI ethical risks are not realized because of bad behavior - can result from not thinking through the consequences, not monitoring AI "in the wild", not knowing what one should be on the lookout for when developing or procuring AI

Narrow AI

dedicated for one task - performs a narrowly defined set of specific task (i.e. Siri, Alexa, self-driving cars, Google search, conversational bots, etc.) -- Where we are currently

Equity

does not aim to promote fairness by treating everyone the same, but by giving everyone equal access to the same

Articulated Robots

emulates the functions of a human arm - can feature anywhere from 2-10 rotary joints - each additional joint or axis allows for a greater degree of motion (ideal for arc welding, material handling, machine tending, and packaging)

Computer Vision

enabling computers to identify and process objects in images and videos in the same way that humans do (i.e. Image recognition, object detection, activity recognition, 3D pose estimation, video tracking, and motion estimation) - typically requires large sets of labeled training images

Creative Augmentation

enhances human creativity and artistic expression - assists in generating new ideas, designs, or artistic work - i.e. DALL-E

Physical Augmentation

enhances human physical abilities - strength, speed, dexterity or sensory perception - embodiment and co-functioning - i.e. exoskeletons, cobots, prosthetics, etc.

Sensory Augmentation

enhances human sensory perception such as vision, hearing, or touch

Emotional Augmentation

enhances understanding, interpreting, and responding to human emotions to enhance emotional intelligence - analyze facial expressions, tone of voice, and other cues to infer emotions and provide appropriate responses

Cognitive Augmentation

enhancing human cognitive abilities through the integration of AI technologies - aims to augment human intelligence, reasoning, memory, and problem solving capabilities - MOST prevalent - i.e. AI powered decision support systems, algorithmic aversion, explainable AI, and prompt engineering

Disparate Impact

even if the policy is neutral on its face, if there is a disproportionately adverse impact on minority groups, liability will be imposed

Equality

everyone is treated the same - fails to take into account that not everyone starts from the same place and that some people may need different support

Natural Language Understanding (NLU)

focuses on semantic analysis or determining the intended meaning of text

Natural Language Generation (NLG)

focuses on text generation by machine

AI algos

form small pieces of algorithm to self-standing solutions available

Cobots

function alongside or directly with humans - share spaces with workers to accomplish more - typically used to eliminate manual, dangerous or strenuous tasks from day-to-day workflows

Job Polarization

hallowing out the middle class - demand and wages for top(cognitive) and bottom (manual) jobs remain high while demand for middle jobs decreases leading to job polarization

Human-in the loop

human is assisted by AI algorithm - human makes decisions and AI provides decision support (i.e. suggestions, automates parts of decisions)

AI Augmentations

humans with AI will replace humans without AI

Fairness

impartial and just treatment or behavior without favoritism or discrimination - means that everyone in the group has an equal opportunity to benefit

AI Production

includes the AI development environment where AI applications are created based on AI enablement designs

AI

intelligence demonstrated by machine

Disparate Treatment

liability could be imposed if there is an explicit classification based on the protected attribute or if there was an intent/motive to discriminate

Human on the loop

machine is assisted by humans - AI makes decisions but human reviews the outcomes and adjusts rules and parameters for future decisions and recommend improvements

Neural Networks

machine learning algorithms that are modeled on the human brain - thousands or millions of processing nodes are interconnected and organized into layers

Human out of the Loop

machine makes decisions - human sets new constraints and objectives and monitors - adjustments based on human feedback are automated

Type of Collaboration - Doppelganger

machines learn from a human or group of humans to mimic their behaviors and decisions so that the humans can be replicated

Thinking Rationally

mimicking thought based on logical reasoning

Thinking Humanly

mimicking thought based on the human mind

Supervised Machine Learning

models are trained with labeled data sets allowing the model to learn and grow more accurately over time - MOST COMMON

Super AI:

more intelligent than humans - a form of AI capable of surpassing human intelligence by manifesting cognitive skills and developing thinking skills of its own

Human in the loop for exceptions

most decisions are automated - human handles only exceptions and control logic to determine which exceptions are flagged for review

Autonomous Mobile Robots (AMR)

move and make decisions in real-time as they go; sensors and cameras help them ingest information about their surroundings - onboarding equipment helps them analyze and make an informed decision such as moving to avoid a worker, picking up the right parcel, etc (i.e. robot dogs, robot vacuums, etc)

Deep Learning

neural networks with many layers - learns features and tasks directly from sets of data - processes extensive amounts of data and determines the weight of each link in the network

General AI

performs like humans - a universal algorithm for learning and acting in any environment

Structure

policies, process, role-specific responsibilities, etc. in place - an organization that has implemented a framework or has a set of mechanisms in place to identify and mitigate ethical risks - An organization that vets for AI ethical risks needs structure

AI consumption

refers to organizations that use or enable diffusion of AI solutions - includes leverage of AI products, packaging AI solutions into offers, and transforming companies to enable successful interface with enablement and production blocks

Humanoids

robots that perform human-centric functions and often take human-like forms - typically use the same technology as AMRs

Content

the ethical risks that a company wants to avoid - most people are committed to a program that includes Content respecting people's privacy, making the outputs of ML explainable, and ensuring that ML delivers fair or equitable outputs

Machine Learning

the field of study that gives computers the ability to learn without explicitly being programmed - computers learn to program themselves through experience (best for situations with a lot of data)

Intelligence

thinking and or acting rationally

AI visualization

tools to display results of AI solutions and interact with users

Reinforcement Machine Learning

trains machines through trial and error to take the best action by establishing a reward system - can train models to play games or train autonomous vehicles to drive by telling the machine when it made the right decisions helping it learn over time what actions it should take

AI platforms

used with code languages and protocols to build applications

Hybrid Robots

various types of robots are combined to create hybrid solutions that are capable of more complex tasks

Privacy

when developing an AI you want as much data as possible from people for standard data analytics and to train your AI - the push to get more data about people can lead to invasions of privacy and AI making inferences about people