AIGP
Accountability -✅✅ -The obligation and responsibility of the creators, operators and
regulators of an AI system to ensure the system operates in a manner that is ethical,
fair, transparent and compliant with applicable rules and regulations (see fairness
and transparency). Accountability ensures the actions, decisions and outcomes of an
AI system can be traced back to the entity responsible for it
Active Learning - ✅✅ -A subfield of AI and machine learning where an algorithm
can select some of the data it learns from. Instead of learning from all the data it is
given, an active learning model requests additional data points that will help it learn
the best. → Also called query learning.
Adversarial Machine Learning - ✅✅ -A machine learning technique that raises a
safety and security risk to the model and can be seen as an attack. These attacks
can be instigated by manipulating the model, such as by introducing malicious or
deceptive input data. Such attacks can cause the model to malfunction and generate
incorrect or unsafe outputs, which can have significant impacts. For example,
manipulating the inputs of a self-driving car may fool the model to perceive a red light
as a green one, adversely impacting road safety.
AGI - ✅✅ -Artificial General Intelligence
AI that is considered to have human-level intelligence and strong generalization
capability to achieve goals and carry out a variety of tasks in different contexts and
environments. AGI still remains a theoretical field of research. It is contrasted with
"narrow" AI, which is used for specific tasks or problems.
.beyond reach right now
.experts expect AGI systems to have strong generalization abilities, the ability to
think, learn and perform complex tasks, and achieve goals in different contexts and
environments
✅✅
AI - an org's first step - -."Every organization should have a baseline AI ethics
code and processes in place to identify, assess, and mitigate the potential harms of
AI use, procurement, development, and deployment"
AI - Company/Institutional Harms - ✅✅ -.Reputation (loss of customers; Brand
impact; loss of trust;
.Cultural (AI bias - we trust AI more than humans)
.Economic (internal resource cost in case of errant AI; litigation costs)
.Acceleration (with volume, speed and complexity, not all risks can be foreseen; AI
issues can have a wider blast radius; GenAI built without necessary controls &
warning signs may not be readily apparent)
,.Legal & Regulatory (existing laws apply - sectoral, privacy, etc.)
AI - Ecosystem Harms - ✅✅-.CO2 emissions
.energy consumption
positive environmental impacts:
.self driving cars
.higher yields in agriculture
.AI in satellite recon - e.g. disaster areas
.weather forecasting
AI - Societal Harms - ✅✅ -..harms to the democratic process and participation
.spread of disinformation
.ideological bubbles or echo chambers
.deepfakes in election process
.safety - lethal autonomous weapons
AI and the workforce - ✅✅ -.concerns about job displacement and automation
.AI can enhance and compliment human activities
.some jobs are at higher risk: repetitive manual tasks; finance and banking; media
and marketing; legal services
.some jobs are at lower risk: skills reliant on creativity, critical thinking, and emotional
intelligence
.workers will need to acquire new skills; add specific social safety nets and worker
protections
AI Application - ✅✅ -.refers to how a specific AI system is used
.e.g. e-commerce, education, healthcare, etc
AI Assessment Process - ✅✅ -.use external frameworks like NIST RMF or ISO
.leverage AI organizations and academic publications
.focus on key AI risks based on the org's AI principles, values and any standards
.contrast the assessment against existing org assessments - avoid duplication of
efforts
AI auditing -✅✅ -.no mature auditing framework in place detailing AI subprocesses
.virtually no precedents for AI audits
.internal or first-party auditors are a start but cannot provide clear public
accountability (proposals in US & EU for first-party audit programs) - would benefit
from being separate and independent from the team developing or using the AI
.leverage existing auditing framework? e.g. COBIT Framework, Institute of Internal
Auditors AI Auditing Framework, COSO ERM Framework
.need to add AI specific audit areas: e.g. data, models, outputs, and processes to
guarantee compliance, ethics, and transparency
,.could have a bias audit to ensure non-discrimination
.can also look at access rights, security controls, and change management
.the use of assessments and audits are the most common mechanisms to provide
assurance about AI systems
✅✅
AI benefits - -.faster and more accurate
.medical assessments accuracy
.legal predictions
.can process huge data sets really fast, and process a wide variety of data sets
.can help remove human error and bias from decision making
.automate and accelerate mundane and repetitive tasks
AI broad usage categories - ✅✅ -1. perform an existing function in a new way
2. Accomplish a new process that has not been done or been possible before the
advent of AI
AI Business Risks - ✅✅ -1. Bias and discrimination:
2. Job displacement
3. Dependence on AI vendors - volatile new field
4. Lack of transparency - avoid black box; document the logic of the AI
5. Intellectual property infringement: copyright, patents, trademarks
6. Regulation and legal risks
7. Ethical considerations: speed to market over safety, etc.;
AI Conformity Assessment ("CA") - ✅✅ -.provides a method of providing
accountability in the development of new technology and data use
.brings together different pieces from other assessments, such as the DPIA &
product safety assessments
.goals of CA: identify how the tech has been developed; what the data set might
have been; how the learning process was developed and how the AI then behaves;
potential impacts of AI over time
.CA required whether PI is being processed or not
.adequate technical documentation is a key component (CAs are more technical
than DPIAs)
.an approved CA presumes adequate and continuous monitoring throughout the AI
process
.
AI Development Life Cycle - ✅✅-.Planning
.Design
.Development
.Implementation
, AI Development Life Cycle - Design - ✅✅ -.data gathering: determine quality of the
data; structured (labeled) or unstructured (unlabeled or uncategorized) data; static or
streaming data
.wrangling or preparing the data: most time consuming step - 80% of the entire life
cycle; involves taking raw data and converting it into valuable information; the five Vs
of data preparation
.cleansing: remove erroneous, irrelevant, unneeded, and personal data
.labeling: tagging or annotating the data as to kind
.anonymization: remove PI identifiers (risk remains of reidentification as data sets
are combined)
.data minimization: if the data is not needed, exclude it from the training data set;
exclude PI
.privacy-enhancing technologies: differential privacy; federated learning
.determine the system architecture: choose an algorithm according to the desired
level of accuracy and interpretability of the data
AI Development Life Cycle - Development - ✅✅ -building the model
.defining the features of the model: work with subject matter experts (domain
knowledge from SMEs); use the same features for training and testing; avoid
unnecessary features
.training: use representative subset of original data set and include all types of data
included therein; train, test, evaluate, and retrain different models to determine the
best model to use; determine the best settings
.testing & evaluation: use same representative subset of original data set as was
used in training; test models on relevant evaluation metrics; test on new data and
training data
AI Development Life Cycle - Implementation - ✅✅ -.continuous monitoring begins
before deployment
.continuously monitor how the model is performing: monitor for deviations in
accuracy; irregular decisions; drifts in data that might affect the model's performance
.model drift is a risk
.continue to iterate the model as the data changes (define a baseline to measure
future iterations)
.AI systems possibly require more attention than other types of systems
AI Development Life Cycle - Planning - ✅✅ -.define business problem (e.g.
classification, regression, recommendation, etc.)
.identify use cases (org mission & goals versus gaps)
.available data - right, enough, accurate, etc.
.determine the scope - prioritize the business problems you want to solve to identify
the use cases to start first (look at Impact, Effort, Fit)
