Environmental justice aspects of the (un)intended consequences of AI

// For example, training a version of Google’s language model (BERT) which underpins the company’s search engine, produced 1,438 pounds of CO2 equivalent in Strubell’s estimate to nearly the same as a round-trip flight between New York City and San Francisco. In practice, a single model is trained a large number of times before it's put in production.

[bobi.rakova], a few more thoughts:

1,438 pounds (=652 kg) of CO2 for Google language model? Training AI models from big data is expensive, but trained once, a model can be reused very cheaply large number of times. Google may re-train the model frequently, so it depends on the frequency, which your comment doesn't mention. Also, it doesn't mention how they are training this model -- if they have something pre-trained already, and training only the top layer, the statistic of expensive entire training would misrepresent the reality of relatively cheap top layer training (transfer learning) on top of the base model. Suggestion: It would be good to mention the total amounts.

In general, because it's expensive to train, but cheap to use, it would be great to have large public model zoo, or model market, so as to prevent people re-training expensive models. I wonder, what's the largest public model zoo that humanity has. Having something like "GitHub" for AI models would also be very liberating to people who have no hardware on which to train large models, but still want to benefit from AI models. I think I had this as a separate idea somewhere.

Thanks for your response [Ruta]! I agree, I think this idea of enironmental justice pushes us to imagine a future where Nature is a stakeholder.

I think it is important to differentiate between AI and technology in general. In my experience, often times, the engineering of an AI model start from the idea of full automation because an AI system can do what people do - faster, cheaper, and better than humans. I think this is why it becomes important to consider the cnsequences as automating away the human introduces new kinds of challenges.

I think it'd be easier to start with the impact assessment of AI models which directly interact with environmental systems as then we could more directly measure the potential impact. Does that make sense?

// AI has consequences much more far-reaching and multi-dimensional than many other technologies. How do we even start to think about the impact of AI on the Environment?

-Hmm, to me it seems we're talking about the same Human impact on the Environment (because Humans create AI to improve lives for themselves). So from here, I invite to zoom out, and add a puzzle on top of this about questioning Human perspective on life in general. When humans see themselves as the centre of the Earth (and probably, the Universe), we create technologies to improve life for ourselves only, e.g. we design say AI to have all knowledge at the click of a button without realising that training AI creates pollution. And then in the design process we further zoom in and divide humans into different stakeholders (like, users, funders, doers, etc) instead of zooming out and having the planet Earth as a stakeholder for example.

So, zooming out here as a thought exercise:

How would the Internet (and any web tech) be designed if we see humans only a specie as part of the living network on Earth and in the Universe? Could we imagine that the planet Earth itself has needs and not only humans do? Could we map out species from the environment as stakeholders in the design process of technologies?

What we could do even here on o2oo: when a new idea is described (before it goes into production), it could share an estimate of the pollution it will create and estimate the impact/value it will produce.

Let's take o2oo system as an example. So by running our website we probably pollute too ([Mindey], maybe you could help me think through this? :), and then we promise to get the humanity connected, so that all new ideas are discussed before they go into production. A promised value (with all stakeholders in mind) would be - that if an idea/project is discussed early, then it can be adjusted, while existing industries are so complex, and are not flexible anymore..

So I wonder how about estimating waste and value/impact for ideas and projects?

Another question is about ethics: Some ideas will promise to create a greater knowledge, even with an expense of creating pollution. What then? How could pollution transform into no-waste (and have a circular model of zero waste)? Say, perhaps when in the ideation process a CO2 emission is projected, then a part of that idea should involve solving that problem in general.

Thank you for your thoughtful reply [Mindey]!

Related to AI's carbon footprint, the best reference I know about is: Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. https://arxiv.org/abs/1906.02243

For example, training a version of Google’s language model (BERT) which underpins the company’s search engine, produced 1,438 pounds of CO2 equivalent in Strubell’s estimate to nearly the same as a round-trip flight between New York City and San Francisco. In practice, a single model is trained a large number of times before it's put in production.

I should have clariefied that I don't mean AI in general but rather the use of Machine Learning in the development of environmental solutions. A great overview paper related to that is: Rolnick, D., et al. (2019). Tackling climate change with machine learning. https://arxiv.org/pdf/1906.05433.pdf See an interactive version of it here: https://www.climatechange.ai/summaries

For example, AI systems have been used to predict forest carbon sequestration potential through the use of satellite and drone image data; computer vision algorithms are used in identifying appropriate planting sites, monitoring plant health, and analyzing trends; AI is also used to identify where deforestation may have been conducted illegally as well as assess risks due to fire, disease, insects, or other causes. References on all of these uses are available with the paper by Rolnick et al.

I'm interested in data governance models in the case of environmental AI models. For example, in the case of AI used for forest restoration projects:

• what data ownership and data governance frameworks could empower equity and inclusion of indigenous worldviews in the way the technology is used?

• what might be the potential impact of such AI models on the concepts of Ecosystem Services and Services To Ecosystems? Comberti et al. (2015). Ecosystem services or services to ecosystems? Valuing cultivation and reciprocal relationships between humans and ecosystems. Global Environmental Change, 34, 247-262. https://www.sciencedirect.com/science/article/pii/S0959378015300145

I edited the post by checking the 'publish' option and I hope I've made it discoverable now, however, I still don't see it on the main Puzzles page.