1 of 7

Brian Ball

AI and Information Ethics Research Lead and Associate Professor of Philosophy

Alice Helliwell

Associate Director of the Graduate Research School and Assistant Professor of Philosophy

Ethical AI Practice - The 3D Model

2 of 7

3D Model

The AI development cycle has three phases: design; develop; and deploy.

These correspond to Ball and Koliousis’ (2023) objective setting (design), training (develop), and deployment management (deploy), and to Morley et al’s (2023) validate (the design), verify (the development), and evaluate (the deployment).

This work is licensed under CC BY-ND 4.0

3 of 7

In the design phase we ask:

What are we building?

Whether to build it?

This work is licensed under CC BY-ND 4.0

4 of 7

This work is licensed under CC BY-ND 4.0

In the develop phase we ask:

How should we go about building the system?

Why should we build it in that way?

The develop phase is for considering how the system can be operationalised, and why it should be done in that way. This includes considering how any system is trained, and why it should be trained on the selected data.

When developing an AI system, the ethical AI practitioner asks:

How?

How should we build the system to achieve its objective? (Does that objective need adjusting in light of practicalities about building a system that implements it?)
How should we collect training data (e.g. to respect autonomy, and avoid bias)?
How can we implement an appropriate data infrastructure (e.g. to ensure privacy)?
We should build the system in the following way…

Why?

Why are we making – and should we make - these engineering choices?
Do they ensure the benefits to stakeholders that are part of the design brief?
Do they avoid the potential harms identified in relation to such a system?
We should build it in this way because…

5 of 7

This work is licensed under CC BY-ND 4.0

In the deploy phase we ask:

Where can (and can’t) this system be ethically deployed?

Who will be affected by its use?

In the deploy phase, we are concerned with where are the appropriate (or inappropriate) applications of this system, and who will be impacted by the system’s use. This should go beyond the intended application, thinking about dual uses. It should also involve thinking about who will be affected beyond the intended user group (including wider society).

When deploying an AI system, the ethical AI practitioner asks:

Where?

Where can/can’t the system be ethically deployed?
Where - in what context(s) - were the training data collected? And where - in which context(s) - can we legitimately generalize from the statistical properties of that data to legitimately deploy the model (e.g. with accuracy)?
Where - in what context(s) - might there be dual use potential? Are some alternative (unintended) uses ethically impermissible - and if so, can they be/have they been designed out?
This system can be deployed in the following contexts…
This system cannot be deployed in the following contexts…

Who?

Who is affected by the deployment of the system in this context? Have they been - or can they be - consulted and engaged in the design process?
What (e.g. legally protected) characteristics do they have?
What are their legitimate interests – i.e., which values are in play? How can those interests and values be navigated?
The following will be affected by the use of this system…

6 of 7

At each phase, practitioners should reflect back on the answers to questions in the previous phase.

For example, in the ‘deploy’ phase, practitioners should consider what the aims for the system are, how they’ve been implemented, and why those choices were made. This will inform the answers to where the system can be appropriately deployed.

This work is licensed under CC BY-ND 4.0

7 of 7

Devon House

58 St Katharine’s Way

London

E1W 1LP

United Kingdom

Tel: +44 (0)20 76374550�info@nulondon.ac.uk

www.nulondon.ac.uk

Thank you