Faster, cheaper, simpler and more reliable communication
Improved access to goods and services
Huge improvement in Healthcare
Better access to travel
Increased safety in travel and other areas
Many other positive changes in our way of life
So, it's all good then? -- Mainly, Yes!
However, with all of this Power comes Responsibility
Many decisions to be made in terms of Systems;
Design
Development
Deployment
Important to balance
Needs
Values
Expectations
Some compromises may be necessary
One size may not fit all Stakeholders
Who is it for?
Who will benefit?
Are anyone's rights being interfered with?
Who owns the system?
Who is responsible for automated Processes?
Who is responsible for automated Decisions?
What about,
Freedom of Information?
Right to privacy
Fairness
Ethics
The principles that govern how we lead our lives
Personal and societal views about what is
Right and Wrong
Just and Unjust
Ethics are not:
Laws, Rules, or Regulations
A checklist of Do's and Don'ts
How we examine ethical problems:
Outcome
Looking at Outcomes can help us discriminate Right from Wrong
Social media example:
Can have positive outcomes:
Social contact, interaction, communication
Access to news and information on products and services
Leisure and entertainment
Can also have negative outcomes:
Bullying
Antisocial behaviour
Mental health issues
Don't throw out the baby with the bath-water
Positive values in computing
Values can fit many categories:
Ethical pluralism
Modern world much more connected
Many people work in Global teams
Global Socio-technical systems
Important to understand and respect the values of all
Danger of 'Dominant Culture' setting the rules
Personal and Professional issue
Professional codes of ethics
Set of guiding principles to
ensure ethical behaviour
high standard of conduct
Public trust in professional integrity
Code of ethics quite common
Engineers
Doctors, nurses
Teachers
Accountants
Computer Scientists
Lawyers
ACM - Association for Computing Machinery
Founded 1947
Largest International Organisation
Represents
Computer scientists
Educators
Researchers
Students
Code of Ethics and Professional Conduct
Updated regularly
Last update in 2018
Lists general principles
Professional responsibilities
Professional leadership principles and rules
Contribute to society and to human wellbeing, acknowledging that all people are stakeholders in computing.
Avoid harm.
Be honest and trustworthy.
Be fair and take actionn not to discriminate.
Respect the work required to produce new ideas, inventions, creative works, and computing artefacts.
Respect privacy.
Honour confidentiality.
EU - Ethics guidelines for Trustworthy AI:
Based on Fundamental Rights and Ethical Principles,
Human agency and oversight.
Technical robustness and safety.
Privacy and data governance.
Transparency.
Diversity, non-discrimination and fairness.
Societal and environmental well-being.
Accountability.
Responsibility and Accountability
More and more aspects of our lives depend on
Computer systems, Hardware, and above all....
Software
Testing is critically important BUT Expensive and Time-consuming
Same goes for training
Companies want to get to market quickly and cheaply
And these companies can be very powerful and persuasive
Politicians and other decision-makers can be lobbied
Personal and Corporate moral responsibility is important
"Computer Error", "Software Glitch" and similar phrases have been used for years to avoid moral responsibility
"Depersonalising" somehow avoids human responsibility
With the rise in Machine Learning, this needs to change.
Someone, somewhere created the Algorithm!
Responsibility, Accountability, and Transparency are especially relevant for systems that use machine learning and algorithmic processes for recommendations or decisions
Data Protection
In a Global data network, the rules can get very cloudy...
... and Broken
Companies can claim that they need to use data to provide "the best customer experience"
Often questions as to the location (jurisdiction) of the data
Not easy to find and prosecute breaches and abuses
Data is well hidden from individuals
Depend on whistle-blowers to highlight misuse and other issues
Transparency in AI Use
Many systems use Modelling and Machine Learning to:
Analyse purchasing patterns
Examine our music, podcast, or video streaming prefences
This data can be used to predict and suggest future 'needs'
But in these cases, no big deal if it's not 100% accurate
However, in some areas accuracy is much more important
Health-related applications like cancer screening demand
Accurate and reliable data - both training and real-time
Transparency with regard to data and the decision-making process
Exhaustive (and on-going) testing of the algorithm
End-user (medical professionals) engagement and feedback
Medical ethics require clinical decisions to be based on
Transparency
Accountability
Explainability
Who or what is responsible for the decision?
Should be the medical professional - supported by Machine Learning
Not all medical conditions are equal in terms of
Research (and Research funding)
Big Pharma support (drugs and research)
Priority given to more affluent societies and illnesses
"Diseases of the rich!"
Better ($$$) results for the pharma companies
Example of... Heart-disease (better cash-flow)
Not all groups are represented equally in the data
Women have traditionally been excluded from research trials
Less affluent societies have been (somewhat) ignored
So, less acccurate data available for these demographics
Symptoms can be misinterpreted and diseases left untreated
These inaccuracies (bias) can cause imbalance in the use of ML
Favouring certain diseases and demographics
Exacerbate inequalities in healthcare
This lack of transparency can affect the uses of AI in healthcare
Bias in AI - Image and facial recognition
Data quality
One of the biggest problems in ML
Poor quality data = poor quality models - GIGO
Note quote from (1700s) - so this is not new
Typical issues:
Missing values (gaps)
Outliers
Imbalanced data
Bias:
Very serious issue in ML datasets
Biased data leads to biased models
US Healthcare model example - racial bias
Misplaced trust in machine-based decisions
But it's the training dataset that matters
Racial, socio-economic or gender are the main issues
Problem with black-box algorithms - data is not transparent
Can lead to questions of trust
Bias in AI - Image and facial recognition
Huge training data-sets created by humans
ImageNet Large image-recognition training library
Images harvested from web searches over 10-year period
Labelled by team of people and fed to a ML algorithm
Seemed to be highly accurate at first
Later shown to contain 'disturbing biases'
Description tags may be correct but open to error / abuse
Subjective descriptions can include bias in terms of:
Political
Cultural
Social
These biases can be conscious or unconscious
ML also requires classification according to values like:
Gender
Nationality
Ethnicity
Etc...
This can be problematic - and even inaccurate
Data-sets are often imbalanced in terms of gender, race, etc.
Many facial recognition systems are only useful when presented with images of white males
Struggle with women and people of colour
The Gender Shades Project - Joy Boulamwini, MIT
Identify bias in gender classification in different AI facial recognition projects
What's the big deal anyway?
Facial recognition is growing at an alarming rate
Unlocking your phone, boarding a plane
Government agencies
Police forces starting to use them
Gardai looking for facial recognition on Bodycams
But the quality doesn't seem to be keeping up
Not yet fit for purpose - Still too much room for error
2020 - Microsoft, Amazon, et. al. paused development of facial recognition software for policing - citing the need for stronger regulations (and more accuracy!)
Ethical concerns? Maybe
Fear of litigation? Definitely!
Fairness and justice
ML systems often deployed for 'efficiency'
Data mined from many sources - even Public Service data
Can affect access to services and even Human Rights
These systems are often in use 'behind-the-scenes' in
Banking (Credit checking, Loan approval, etc.)
Justice system (Bail application)
Insurance (Risks, Actuarial data, etc.)
Little known of their algorithm (black-box) or training data
However, they now have access to other data such as:
Financial data
Advertising and Cookie tracing
Social network posts
Purchasing data
WHY???? - Price Optimisation
Price Optimisation???
What you would be willing to pay!
Data is also used to fine-tune the Risk calculation
Data often obtained 'by stealth'
Gives them an unfair advantages
Algorithmic decision making in Criminal Justice
Used in USA and other countries
Decisions once made by Police, Judges and Juries now made by Machine Learning systems
For criminal offences, ML systems can recommend:
Profiling
Arrest
Sentencing
COMPAS - Sounds like a great idea!
However, it has been found to have bias
Seriously overestimates the risk of reoffending for people from more disadvantaged backgrounds
Most bias shown against ethnic minorities
Why this?
Data used was mainly socio-economic - based on
Employment history
Education
Family Health
"Known to the Police"
COMPAS - has been found to have bias - Why is this?
More affluent - but more serious criminals judged to be less at risk of reoffending!
Decision systems using Aggregated data more likely to replicate inequality - why?
Data-sets reflect the structures of society and
Society is unequal
Data aggregation is any process whereby data is gathered and expressed in a summary form. When data is aggregated, atomic data rows - typically gathered from multiple sources are replaced with aggregated totals
What can we do about these inaccuracies and biases?
Regulate training data sets
Data cooperatives to design training data from scratch
This should make algorithmic decision making
More transparent
More representative
More respected and trusted
Socio-technical systems
Computer systems can't exist without an ecosystem
Social, economic, and political support structures
For example, for computing we need
Electricity networks
Internet connections
WiFi and Communications networks
Payment systems
Agreements, Laws, and Standards
Computing is a socio-technical system
More than the hardware and software
Shaped and supported by relationships and agreements
Social
Economic
Political
These are based on a set of Values
To date, mainly White, Middle-class males have shaped the computing landscape
Examples of 'shapers' include:
Steve Jobs & Wozniak (Apple)
Bill Gates (Microsoft)
Mark Zuckerberg (Facebook)
Notice a pattern?
Deborah Johnson (IT Ethicist) certainly did.
Software (and System) design reflects prevailing context and culture
It's possible that in future, more diversity within computing will create different 'shapes'
May move to a more altruistic less commercial model