Methodology

📄Selecting sources

We drew on a wide range of sources when building the EAI database, from academic journal articles and governments reports to news reports. In order to make sure we didn't repeat one another's work (by documenting both an academic paper and the news media reporting on it, for example), we followed a policy of tracing reporting 'back to the source'.

⚠️Methodological challenges

We ran into some challenges when building our database:

• Not all the data we wanted to 'tag' was available for every source. This meant we had to leave some categories blank, which might make it harder for our users to find sources relevant to their interests.

• Calculating the environmental impacts of AI is still a very new field. Different researchers quantify these impacts differently, often using different units or methods of modelling. This means we couldn't compare their findings as directly as we would have liked.

• Many of our sources are very academic, and full of technical jargon. We included a glossary in our database to try and make them more accessible to our users, but we recognise that it only scratches the surface of the terms they may not understand.

• Similarly, many of the impacts given by our sources are given in mathematical terms, which are not intuitively meaningful to most people- what does it really mean that 'Implementing ChatGPT-like AI into every Google search would amount to 23–29 terawatt hours annually'? We combatted this by including a section on the metaphors of environmental cost used by our sources.

• However, these metaphors often individualise the environmental impacts of AI, such as by exemplifying the environmental impact that can be attributed to a single query to an LLM like ChatGPT. This could be seen as holding individual users responsible for the impacts of underregulated tech corporations, and obscure all the environmental harm done by AI during development (before it is accessible to users).