---
tags: ethicschampions
date: 2022-05-25
---

# Ethics Champions Meeting
## Elements of a Data Ethics Case Study

:::success
**Example Case Study**

This [document](https://drive.google.com/file/d/12Z11YqzoKvWM3HyVN8Rtdl3ux2_6oUhp/view?usp=sharing) is a case study used in a separate project on digital mental healthcare.
:::

### General Structure

1. Case Description
- Type
- Contextual considerations - e.g., time frame, social/political environment
- What has changed since this occured or the case was generated?
- Has the technology, practice, or other conditions moved on?
- Aspects of the Project Lifecyle implicated. (tags?)
- ?
2. Technical Background
- How does this work? (what is known)
- What is know about the model, training data, etc.?
- What are the inputs?
- ?
3. Responsibility
- Who is accountable and to whom?
- What are individual data scientists or programmers responsibilities?
- Future-proofing: who is responsible for ongoing oversight, monitoring. Sustainable accountability.
- ?
4. What are the issues?
- Who is impacted?
- Categories of stakeholders - categories of rights and interests
- Individuals, society, sub-groups
- What are the trade-offs involved?
- Sources, types of bias
- When did this/these issues emerge?
5. Questions and Options
- Are there unintended consequences?
- What could have been done differently?
- ?
6. Mitigation
- What could be done?
- When should mitigations have been recognised or implemented?

### Case Study Ideas

#### Virtual Visitation Service
1. Description
Digital service allowing prisoners to have virtual visitations. Provided by a third-party. Screens for violations of visitation policies - who can attend, what they can share/show.
- Type: Real, in production, provoided by third-party
- Context: Response to pandemic restrictions and concerns.
- Links for further info
- https://prisonvideo.com/
- https://phonehub.io/

2. Technical aspects
Data science bits

- Nudity screening
- People not approved participating
- Facial recognition - registered faces only w/ ID
- Required technologies
- Nudity detection
What does it detect?

3. Responsibility
- Procurement - Digital Services, Prison Service
- Data Science - advise and support, though not directly responsible for this one.

4. Issues
- Facial recognition - requiring people to register faces
- Data storage - what happens to the registered faces?
- What is the training data? Bias issues.
- Third party use/sharing.
-
5. Mitigation
- Upstream - request additional information, accountability
- Downstream - look at the effects of use. Can any risks or harms be mitigated in how the tool is used.

## Bias Self-Assessment

Chris's [presentation](https://github.com/alan-turing-institute/ethics-toolkit/blob/main/docs/resources/slides/bias-workshop-moj.pdf) on types of bias.

### Discussion Notes

NB: look at the project lifecycle, the bias document, the case above.

- Project lifecycle may need revisions for procurement
- Training/serving skew
- Training the model on a data set, testing on that data set, applying to another population
- Model selection, training, testing,
- Deployment
- Evaluation bias and implementation bias
- Need to get a real understanding of how the tool will be used in practice.
- What are the specific conditions of implementation.
- Mitigation: evaluating at point of use. Proper testing under full conditions.
- Evaluation over time
- There will be a limited scope of any project team to mitigate this bias during use means that being clear on communicating the limitations of any model validation are necessary (e.g. recommended use of model)
- Bias mitigation strategies
- Require formal bias evaluation of models (mathematical definitions) to anticipate distributions across protected groups of other groups of concern.
- Missing data - significant issue in prison context. Example - 'staff assault predictor' tool lacked sufficiently rich data.
- Also concern about what we don't know - missing data we haven't yet identified as missing.