AI Policy

Our policy

The use of AI tools and systems has a range of environmental impacts, and can lead to discriminatory outcomes for individuals and hoarding of power by a small number of billionaire-owned companies. 

These technologies can also have accessibility and inclusion benefits, and can be used to increase Watershed’s operational effectiveness. 

As such, all staff are asked to be mindful that their use of AI tools and technologies comes with consequences and use them only when necessary.  Organisationally, we will implement AI-enabled tools and systems when we are confident that there is a robust business case. 

Watershed has the following policies in place for management of AI:

1. Meeting transcriptions and summaries
2. Use of generative AI tools
3. Systems that use machine learning
4. Transparency statements 

These will be reviewed annually, or more frequently if required. 

1. Meeting transcriptions and summaries 

It is Watershed policy to only use AI transcriptions and summaries of video conversations when the benefit to an individual or team is shown to outweigh potential broader societal harms. 

Specifically: 

• As part of an individual’s broader package of reasonable workplace adjustments
• When there is a clear case that a mostly accurate, verbatim transcript is required (such as during a research interview). 

In both cases, staff may be asked to conduct an ethical, social and environmental impact assessment to confirm that generative AI tools are suitable for the task.

Consent

In line with guidance from the Information Commissioner, everyone who is recorded or transcribed must give clear consent. If possible, please arrange this in advance so that alternative methods of transcription may be arranged if any participant raises a reasonable objection. Ordinarily, the use of transcription to support a reasonable adjustment will be considered to have priority over other concerns. 

Tools

Watershed’s IT team will confirm which tools are suitable for use (depending on whether the meeting takes place over Zoom, Teams, or another video or audio conferencing package); requests to use other forms of software may be subject to an impact assessment. 

Use of Recordings, Transcriptions and Summaries 

Recordings and transcriptions may contain errors and so may not be used as part of HR or grievance procedures. Outputs (recordings, transcriptions and summaries) should not be distributed to others without the consent of everyone featured.

Storage

Transcriptions may only be stored for 30 days, after which time they will be automatically deleted. 

Declining Recording and Transcription in Meetings With External Stakeholders

From time-to-time, staff may be invited to take part in external meetings at which recording or transcription is turned on without consent. In these cases, you are empowered to decline and say either in writing or verbally that it is not Watershed policy to use generative AI as an ambient tool and that it is only organisational policy to use AI transcriptions and summaries of video conversations when the benefit to an individual or team is shown to outweigh potential broader societal harms. 

2. Use of generative AI tools 

Watershed does not support the use of generative tools to create images, audio or video. 

From time-to-time, staff members may use the generative AI numerical, text generation and summary tools in Microsoft Copilot. For specialist tasks, including software development, Claude.ai may be used. 

• Generative AI must not be used in any public-facing outputs
• Use of generative AI tools should be subject to an ethical, social and environmental assessment, and – at first use – each use case should be logged in the Watershed Generative AI Register. The Register should be revised annually to remove outdated and defunct use cases.
• Transparency statements should be used whenever generative AI has been used (see below).
• Where accuracy is important, staff should always keep a human in the loop, and ask one other person to check the accuracy of outputs. 

3. Systems that use machine learning 

Organisational systems may be purchased or developed that use machine learning and other kinds of AI. These systems will be subject to impact assessments. 

4. Transparency statements 

If generative AI has been used, we will let others know by adding a short transparency statement such as a footnote to a document, or in the comments of any code that has been generated.