‘We don’t have all the answers or even all the questions…’ …is how I start most of my talks on AI. So given the emergent state of the technology how do you develop a set of AI Principles for all staff at a large arts university?

On the one hand you must highlight inherent risks and ethical complexities, on the other you don’t want to stifle experimentation and creativity. Any Principles should create a critical space where both resistance and engagement can co-exist and, importantly, be in dialogue with each other. This allows our staff and students to variously learn, co-opt and innovate while questioning and critiquing the ‘zero-sum’ future promoted by Silicon Valley.

These approach underpinned the development of the Principles below which provide a shared framing for approaches to AI across almost all contexts at UAL. Any plans for, or use of, AI can be held up to the Principles to assess what level of risk (technical, conceptual or social) is being taken and to indicate what areas should be given additional consideration. Any use and any critical refusal of AI can be informed by the Principles.

These high-level Principles sit alongside the guidance developed specifically for the use of AI within Teaching and Learning. As mentioned they have been designed for staff (academic, technical and professional services) but they can also be used by and with students.

As outlined in our internal version of the Principles, “They are primarily designed to respond to Generative AI, which can produce text, image and other forms of media when prompted.”

---

1)  Be intentional and stay in control of the process

• Stay cautious and curious. Question what should be trusted and why.
• Start by exploring AI use to improve the quality of your work, before considering how it might help you to work faster. Consider the balance between convenience and active decision-making.
• If the use of AI in your process is significant (for example, it is intended to influence decision making), track and cite your use as part of the work.

2) Don’t automate your judgement

• You are accountable for your use of AI. Have clear aims and standards to assess the quality of AI outputs.
• Critically evaluate (review, edit and adjust) all AI outputs. AI output often looks good at first glance but might contain errors or omissions, especially if you prompt AI to produce a lot of material with minimal input or effort.
• Be cautious when using AI in processes which have an impact on others. AI reflects biases inherent in its training data and therefore inherent in wider society.

3) Use responsibly: there will be climate, social and reputational impacts 

• Consider whether AI is necessary for what you’re trying to achieve.
• Digital work has physical impacts and AI can use a lot of energy. Intentional approaches to AI should minimise use in line with our Climate Action Plan.
• As a creative community, we value intellectual property. Authors and creators whose work is used to train AI are often not credited.
• Some aspects of AI use are in tension with our Climate, Racial and Social Justice principles. Equality should be considered in both use and access to AI tools. Consider the ethics of AI use in your context.

4) Think of AI as a machine, not a person

• Generative AI is designed to mimic human-like interaction, but it is a machine. Keep this in mind when using AI tools.
• Some Generative AI is designed to reinforce your point of view rather than enrich your thinking or insights.

5) Don’t assume AI knows everything or is impartial

• The information any AI ‘contains’ is extensive but not exhaustive.
• AI outputs are unlikely to contain niche, marginalised or emergent knowledge, and will not represent all possibilities and perspectives.

6) Only input what you have a right to use and share

• Only use Copilot Chat when logged-in with your UAL account, this will ensure data is protected and your chats are not used to train AI models.
• If uncertain, assume that data may be sensitive and avoid inputting confidential, personal, or restricted information. Use only a UAL‑approved AI tool (Copilot Chat) when working with protected information.
• Ensure any information you put into an AI tool complies with UAL’s Information Security Policy and Data Protection Policy. Guidance can also be found in the ‘AI Guidance on Information Governance and Risk’ (links to policies removed for this blog post as they or located on the UAL intranet)
• Avoid entering material that may have intellectual property (IP) restrictions, such as, copyrighted texts, images, or assets you don’t have permission to use.

7) Using AI not provided by UAL incurs increased risk and costs.

• Even when using UAL‑approved tools such as Copilot Chat, critical judgement is still required. All other UAL policies and the AI Principles continue to apply.
• Using AI that is not approved by UAL introduces significant risks and may put students’ and staff data at risk.
• Consider your flexibility to change provider as costs are likely to increase over time.

---

The Process

The process to develop the Principles was in-depth but not complex. A couple of writing workshops with members of the UAL AI Group I co-chair and consultation with relevant groups and expert individuals. Having been immersed in ‘AI and education’ for a couple of years I didn’t find it too difficult to draft a first version for discussion and editing. The headline themes from various AI talks I’ve given since 2023 acted as a reasonable starting point.

Over a couple of months of consultation we narrowed it down to seven Principles which covered a lot of ground. One of the strengths of UAL as a community is that there is an understanding that brevity done well is difficult but worth the effort. The process was also helped by the great work that had already been done by our Teaching and Learning Directorate who, like many universities, had rapidly developed guidance for staff and students in the use of AI within the curriculum.

The Language

The Principles had to work across all our staff groups, not just the academic community. Plus, I was confident that if we got this right the Principles would also be used by and with students which seems to be happening already. Respect to our internal communications department here who are experts in spotting overly complex language and were happy to suggest much improved alternatives. (For example, my first draft contained the phrase ‘attenuating agency’.).

I was also keen to avoid anthropomorphising AI through our use of language used and so stuck with phrases like ‘the output from AI’ rather than referring to AI as ‘it’. This is pretty subtle but I believe we have to constantly reassert that AI is a machine at every possible opportunity.

The ground covered

Although not presented in this way I think of the seven Principles in three sections:

Practice: Principles 1, 2 and 3.
Incorporating AI into process and practice (or why you might actively chose not to). The central point being that whatever the technology produces a person, probably you, will be responsible for the overall process.

Reality check: Principles 4 and 5.
I am certain that being aware that AI is a ‘machine with limits’ leads to better critical evaluation of outputs than imagining or assuming AI is an all-knowing guru. This might seem obvious but it’s easy to tacitly fall into this trap because of the way the technology is packaged and presented. i.e. Most of the major Gen AI providers want you to think of the technology as an omniscient person.

Safety and data: Principles 6 and 7.
These are specific to UAL in that they refer to Copilot Chat which is provided to all staff at UAL. For the most part these two Principles are simply a reminder of the IT policies which have applied to the data and the digital environment for years. The main addition here is around cost as the try-before-you-buy model of technology proliferation is ok if you can hop between platforms (as many of our students do) but risky if you build a ‘free’ technology into business-as-usual processes.

Principles not rules

Beyond a couple of redlines around academic misconduct and data security AI use is very much a live debate. We need to be critically questioning and thoughtfully experimenting. As the a recent report from HEPI pointed out we can’t expect our communities to develop critical approaches to AI if we don’t trust them to experiment with AI thoughtfully. Whether its developing ways of conceptualising and understanding what the tech might mean or it’s experimenting with what the tech can and cannot do, this must involve trust.

In that spirit, we are developing a workshop which involves using the Principles to assess your use, or critical non-use, of AI. An approach to AI which is in tension with a Principle is not necessarily something to be stopped, it is something which requires a deeper level of consideration.

Experimentation and innovation by their very nature will involve taking risks. The question then becomes identifying these risks and seeking agreement that they are reasonable. How reasonable a risk might be will depend on context and potential impact.

This is nothing new but each emergent digital technology offers more power and therefore greater risks. With AI however, the sheer scale of investment has generated a huge amount of hype and vested interests. An individual who gets caught up in this might have poorly informed expectations of the technology or they might have a false conception of how the tech functions. In this case many unforeseen negative effects are likely.

The Principles act as set of coordinates to help navigate the sociotechnical storm of AI.