LOTI_LDN

medium.com/loti

#LOTI

Innovating Responsibly with

Data and AI

​

Workshop 1: What’s good, bad and missing from existing guidance

​

14 Feb, 9:00-12:00

​

Lead borough: Brent

​

LOTI_LDN

medium.com/loti

#LOTI

Welcome and introductions

​

Cllr Margaret McLennan and Peter Gadsdon

09:10

Agenda

Working definition

of AI

“Software that learns”

Und

Objectives for today

01

02

03

Understand which of these resources are most helpful and where there are gaps 

Explore existing tools and guidance for how data and AI can be used responsibly in public sector organisations

Agree on what actions need to be taken to address the gaps

Rapid Presentations: Exploring existing tools

​

09:20

UK Statistics Authority Data Ethics Self Assessment

​

Paul Hodgson

Greater London Authority

​

​

​

​

UK Statistics Authority – Self-assessment framework

10

“[your scientists] were so pre-occupied with whether or not they could, they didn’t stop to think if they should”

- Dr. Ian Malcolm

​

UK Statistics Authority – Self-assessment framework

11

• Used by Research Accreditation Panel
• Self-assessment form

​

• Things I like:
• Overall score & individual
• Guidance
• definitions
• scoring scales
• Supports consideration of the ethical risks throughout the project cycle

Oliver Thereaux

Head of Tech

​

Violeta Mezeklieva

Data Trainer

​

Ben Snaith

Researcher

​

Open Data Institute

Data Ethics Canvas

​

​

​

​

​

​

Data Ethics Canvas

​

Oliver Thereaux

Head of Tech

​

Violeta Mezeklieva

Data Trainer

​

Ben Snaith

Researcher

​

Guidelines

1. To select a specific layout simply click Layout button above and choose the most appropriate slide layout for your content.
2. In order to change the background colour of your slide, right click and select Change Background. Choose a colour from the theme colours embedded in the template and click Done.
3. To place an image into a slide please use one of the image templates provided right click on the placeholder image and select Replace image. Then just select the appropriate source and the image will crop into the correct size and position on the slide.
4. If you would like a certain crop of the image, double click and move the anchor points accordingly to position the image in the correct place.

For other FAQs please refer to the guidelines provided here

Guidelines

• To select a specific layout simply click Layout button above and choose the most appropriate slide layout for your content.
• In order to change the background colour of your slide, right click and select Change Background. Choose a colour from the theme colours embedded in the template and click Done.
• To place an image into a slide please use one of the image templates provided right click on the placeholder image and select Replace image. Then just select the appropriate source and the image will crop into the correct size and position on the slide.
• If you would like a certain crop of the image, double click and move the anchor points accordingly to position the image in the correct place.

For other FAQs please refer to the guidelines provided here

A branch of ethics that evaluates data practices with the potential to adversely impact on people and society - in data collection,

sharing and use.

Data Ethics

14

theODI.org

Data Ethics into Practice:

DATA

IMPACT

ENGAGEMENT

PROCESS

theODI.org

Guidelines

1. To select a specific layout simply click Layout button above and choose the most appropriate slide layout for your content.
2. In order to change the background colour of your slide, right click and select Change Background. Choose a colour from the theme colours embedded in the template and click Done.
3. To place an image into a slide please use one of the image templates provided right click on the placeholder image and select Replace image. Then just select the appropriate source and the image will crop into the correct size and position on the slide.
4. If you would like a certain crop of the image, double click and move the anchor points accordingly to position the image in the correct place.

For other FAQs please refer to the guidelines provided here

16

theODI.org

Guidelines

• To select a specific layout simply click Layout button above and choose the most appropriate slide layout for your content.
• In order to change the background colour of your slide, right click and select Change Background. Choose a colour from the theme colours embedded in the template and click Done.
• To place an image into a slide please use one of the image templates provided right click on the placeholder image and select Replace image. Then just select the appropriate source and the image will crop into the correct size and position on the slide.
• If you would like a certain crop of the image, double click and move the anchor points accordingly to position the image in the correct place.

For other FAQs please refer to the guidelines provided here

Understanding AI Ethics and Safety

​

David Leslie

Alan Turing Institute

​

​

​

​

Dr. David Leslie, Ethics Theme Lead��Understanding Artificial Intelligence Ethics and Safety�Public sector guide to the responsible design and implementation of AI systems�

The Alan Turing Institute

Three tiers of a practice-based AI ethics framework

Using the FAST Track Principles

Putting the Ethical Platform into Practice

Using the SUM Values

• Ask and answer questions about the ethical purposes and objectives or your project
• Assess and re-assess the impacts of your project on individuals and communities

• Ensure that every step of your project aims to produce ethical, fair, and safe AI innovation
• Design and implement responsibly

Using the PBG Framework

• Set up governance processes that ensure end-to-end transparency and accountability

The SUM Values�Ethical values that Support, Underwrite, and Motivate a responsible data design and use ecosystem

The Alan Turing Institute

Why the SUM Values? Potential hazards of AI/ML:

Loss of agency and social connection

Poor quality outcomes

Bias and discrimination

Why the SUM Values?

Scope of SUM values was informed by the real-world risks posed by the use of AI/ML:

The SUM Values

Agency

Justice

Wellbeing

Interaction

Respect the dignity of individuals as persons

• Ensure abilities of individuals to make free and informed decisions about their own lives
• Safeguard their autonomy, their power to express themselves, and their right to be heard
• Secure their capacities to make informed and well-considered life choices
• Support their abilities to fully develop themselves and to pursue their passions and talents according to their own freely determined life plans

​

​

​

​

Connect with each other sincerely, openly, & inclusively

• Safeguard the integrity of interpersonal dialogue and connection
• Protect the human interaction as a key for trust and empathy
• Use technology to foster this capacity to connect so as to reinforce reciprocal responsibility and mutual understanding

Care for the wellbeing of each and all

• Design and deploy AI to foster and to cultivate the welfare of all stakeholders whose interests are affected by their use
• Do no harm with these technologies and minimise the risks of their misuse or abuse
• Prioritise the safety and the mental and physical integrity of people when scanning horizons of technological possibility, conceiving of, and deploying AI applications

Protect justice, social values, and the public interest

• Treat all individuals equally and protect social equity
• Use digital technologies to support the protection of fair and equal treatment under the law
• Prioritise social welfare, public interest, and the consideration of the social and ethical impacts of innovation in determining the legitimacy and desirability of AI technologies
• Use AI to empower and to advance the interests and well-being of as many individuals as possible

FAST Track Principles

29

12/02/2020

The Alan Turing Institute

The Alan Turing Institute

The ML innovation pipeline has many steps that involve human choices

30

12/02/2020

The Alan Turing Institute

31

12/02/2020

​

�

​

�

The Alan Turing Institute

Transparency

The principle of transparency entails that design and implementation processes are justifiable through and through. It demands as well that an algorithmically influenced outcome is interpretable and made understandable to affected parties.

​

32

12/02/2020

The Alan Turing Institute

Accountability

Accountability entails that humans are answerable for the parts they play across the entire AI design and implementation workflow. It also demands that the results of this work are traceable from start to finish.

​

33

12/02/2020

The Alan Turing Institute

Fairness requirements in the ML process

Discriminatory non-harm - Designers and implementers are held accountable for being equitable and for not harming anyone through bias or discrimination.

​

• Data fairness – training and testing data should be representative, relevant, accurate, and generalisable
• Design fairness – model target variable, feature, processes and correlations should be reasonable, justifiable, not objectionable
• Outcome fairness – model outputs do not have discriminatory or inequitable impacts on people
• Implementation fairness – users of the model are sufficiently trained to use it without bias

34

12/02/2020

The Alan Turing Institute

Sustainability

• Designers and implementers are held accountable for producing AI innovation that is safe and ethical in its outcomes and wider impacts. This requires:

​

🡪 Stakeholder Impact Assessments to continuously consider and review the impacts of innovation on individuals, society, and nature

35

12/02/2020

The Alan Turing Institute

Process-Based Governance Framework

36

12/02/2020

The Alan Turing Institute

The Alan Turing Institute

Process-Based Governance Framework for AI Project Workflows

Targeted Consideration

Governance Action

38

12/02/2020

The Alan Turing Institute

turing.ac.uk�@turinginst

AI Impact Assessment Canvas

​

Adriano Koshiyama

University College London

​

​

​

​

Algorithmic Impact (AI)

Assessment Canvas

_{February, 2020}

Adriano Koshiyama

Dr. Emre Kazim

Dr. Zeynep Engin

Contact: e.kazim@cs.ucl.ac.uk

An AI Assessment Canvas

• The AI Assessment Canvas is a great tool for planning, communication and ML project tracking
• However, its focus is on how the problem will be solved, and not what questions the solution need to address

• Hence, we need to recreate this Canvas, moving it from
• Value-centric to
• safety-centric

decision-making

An AI Assessment Canvas

​

An AI Assessment Canvas

• Section and content of ML Canvas
• Value propositions: What are we trying to do for the end-user(s) of the predictive system? What objectives are we serving?

​

• Section and content of AI Assessment Canvas
• Safety propositions:
• What are we trying to do for the end-user(s) of the system?
• What objectives are we serving?
• How our decisions will affect the end-user(s) prospects?

An AI Assessment Canvas

• Section and content of ML Canvas
• Decisions: How are predictions used to make decisions that provide the proposed value to the end-user?

​

• Section and content of AI Assessment Canvas
• Decisions:
• How these predictions will affect the end-user?
• What are the implications of these decisions to the end-user when the system’s predictions are wrong?

An AI Assessment Canvas

• Section and content of ML Canvas
• ML Task:
• input: individual transactions
• output to predict: default
• type of problem: classification/credit-scoring

• Section and content of AI Assessment Canvas
• AI Assessment Task:
• Type of assessment: fairness/discrimination in credit scoring
• Group variables: gender, ethnicity, etc.
• Output to explain: default ratio
• Worst-case analysis: not applicable

An AI Assessment Canvas

• Section and content of ML Canvas
• Offline evaluation: Methods and metrics to evaluate the system before deployment.

​

• Section and content of AI Assessment Canvas
• Offline assessment:
• Fairness: describe here the tools used (e.g., fairness constraints)
• Explainability: describe here the tools used (e.g., Shapley-values)
• Robustness: describe here the tools used (e.g., cross-validation)

An AI Assessment Canvas

• Section and content of ML Canvas
• Making predictions: When do we make on new inputs? How long do we have to featurize a new input and make a prediction?

​

• Section and content of AI Assessment Canvas
• Making assessments:
• When do we make on new predictions/decisions?
• How long do we have to wait to see the implications of the system’s decisions?

​

An AI Assessment Canvas

• Section and content of ML Canvas
• Data sources: which raw data sources can we use (internal and external)?

​

• Section and content of AI Assessment Canvas
• Data sources (internal and external):
• Which raw data sources are we using?
• Are we infringing individual/company rights of privacy and agency in using these data sources?

An AI Assessment Canvas

• Section and content of ML Canvas
• Features: Input representations extracted from raw data sources.

​

• Section and content of AI Assessment Canvas
• Features:
• Which demographics are we including/transforming? Are we paying attention specifically to vulnerable ones (e.g., children, minorities, disabled persons, elderly persons, or immigrants)?
• Are we applying transformations to some features that render them difficult/unable to be explained later on?

An AI Assessment Canvas

• Section and content of ML Canvas
• Building models: When do we create/update models with new training data? How long do we have to featurize training inputs and create a model?

​

​

• Section and content of AI Assessment Canvas
• Building models:
• What are the limitations of the model being created/updated (Robustness, Explainability and Fairness)?
• What are the measures put in place to mitigate or eliminate these limitations are being addressed?

An AI Assessment Canvas

• Section and content of ML Canvas
• Collecting data: How do we get new data to learn from (inputs and outputs)?

​

​

• Section and content of AI Assessment Canvas
• Collecting data:
• Are the subject aware of the data collection?
• What are the access/control of the subjective over its data?
• The data being collected are actually being used during the modelling process, or are we just storing it?

An AI Assessment Canvas

• Section and content of ML Canvas
• Live evaluation and monitoring: Methods and metrics to evaluate the system after deployment, and to quantify value creation.

​

​

• Section and content of AI Assessment Canvas
• Live evaluation and monitoring:
• Methods and metrics to assess the system after deployment, such as to quantify its reliability, fairness and interpretability during its decision-making.

10 Questions to Answer Before Using AI in the Public Sector

​

Eddie Copeland

LOTI

​

​

​

​

There’s no one set of principles that can cover all contexts.

​

The level of risk and reward matters.

​

​

There’s no one ‘right’ set of answers.

​

Rather, public sector organisations should only deploy an AI if they have answers appropriate for the context in which they are working.

​

Exercise 1:

​

Testing the tools and spotting gaps

Split into five groups.

​

Each table table will use a different tool.

​

Read the five scenarios.

​

Pick ONE scenario and explore how the tool helps inform your thinking about how you’d approach the problem.

​

(Move tables after 30mins and repeat.)

​

​

​

10:00

1 - Supporting Children’s Safeguarding

​

A borough wishes to use machine learning to analyse social workers' case notes to see if they can predict which children are likely to be taken into care in the next six months. The tool would analyse case notes from previous cohorts of children taken into care and aim to identify common patterns and trends that might help predict the level of risk for current groups of children under assessment. The tool would be intended as a back-up to social workers' own assessments, prompting them to double-check their reasoning if it assessed a child's risk as being higher than expected.

2 - Creating a smart chatbot to handle resident queries

​

The borough wishes to create a smart chatbot that can help answer common queries that would normally go to its contact centre. Conscious of the high demand on their contact centre and rising complaints about the long waiting time to be answered, the borough believes a smart chatbot could be trained on historic call data about resident queries. The chatbot could answer a more sophisticated range of questions by learning from real users over time. The borough hopes this will free up the time of contact centre staff to answer more nuanced and complex questions.

3 - Identifying Rogue Landlords

A borough wishes to help identify which of its landlords in the private rental sector may be exploiting tenants and leaving them in inadequate housing conditions. The housing team thinks three categories of data might be relevant to a machine learning model. One set concerns the physical characteristics of a property, such as its location, age, and type (e.g. a flat above a takeaway). A second set concerns reports of events at the property, such as visits by animal control, noise complaints etc. The third set is information about the landlords themselves, such as their financial transactions, company dealings and previous known history of operating as a landlord. Their intention is to use this data model to target inspections of properties and proactively find tenants being exploited before they receive any complaint.

4 - Young people at risk of joining gangs

​

The borough wishes to use data to tackle the rise of gang violence in some of their wards. Their proposal is to work with Police and youth workers to identify factors that might make a person vulnerable to joining a gang, and create a dashboard that shows the presence of those factors for individuals living in the borough. The borough is wondering whether they could use the dashboard to proactively intervene and support individuals already deemed to be at high risk, or instead aim to identify wards with high concentrations of those risk factors in order to alleviate them for the benefit of much younger children.

5 - Analysing sensor data from social housing to enable predictive maintenance

The borough is responsible for managing a large range of social housing properties, from detached houses to big blocks of flats. Currently, the inspection and maintenance of those properties are done based on assumptions about how often key assets need to be checked or replaced. The borough proposes to install sensors that check building temperature, moisture levels, boiler usage and performance, energy consumption and lift movements. A private company offers a product which they claim analyses this type of data to predict where problems may occur. The borough hopes this will enable more predictive maintenance to better serve social housing tenants and reduce costs.

Exercise 1:

​

Testing the tools and spotting gaps

Split into five groups.

​

Each table table will use a different tool.

​

Read the five scenarios.

​

Pick ONE scenario and explore how the tool helps inform your thinking about how you’d approach the problem.

​

(Move tables after 30mins and repeat.)

​

​

​

10:00

BREAK

11:15

LOTI’s six

workstreams

Exercise : 2

​

Addressing the gaps

How helpful was each tool in structuring your thinking?

​

• Things you liked

​

• Things that could be changed to make it more relevant

​

• Things that were missing

​

11:20

Summary and Next Steps

11:55

Innovating Responsibly with Data and AI

​

Workshop 1: What’s good, bad and missing from existing guidance

14 Feb, 09:00-12:00

​

Lead borough: Brent

​

London Office of �Technology & Innovation

LOTI_LDN

medium.com/loti

#LOTI