The massive growth of AI of the last few years has happened through bribing officials to circumvent environmental protections and ignoring the communities they build data centers in, and the destruction of cultural and historical relics at a terrifying scale. We already do not know the scope of what we have lost, but we do know the harms that are already being perpetuated on the communities that live near their data centers.

So often the refrain is to minimize the harms, say that progress is necessary and this is the only way. But we cannot colonize our way to a more just and happier world. Progress starts with justice and care now, not for a hypothetical future utopian state.

As someone who has been experimenting with language models for my own private research for over a decade, the state of generative AI is a travesty. A tool that many of us hoped would help us analyze large sets of data to find interesting or hidden patterns has become--like many revolutionary technologies before it--another tool of imperialism and subjugation for the benefit of a few.

I have seen many projects where AI is used responsibly and for the benefit of people and communities who are underserved by existing structures and institutions, but for every good use case, I have seen innumerable troubling cases. From inaccessible interfaces to silent hallucinations hidden in complex analyses, the industry has created powerful amoral machines that are literally focused on doing whatever they can without stopping to think about whether they should.

From whom? By whom? For what purpose?

My stance on AI, as with my stance on all technology, is on the side of people and communities.

From whom

From stealing the intellectual property of creators to feed their models to regulatory capture to building data centers without environmental oversight in communities who don’t want them, as well as the wanton disregard of the social welfare of communities, workers, and anything other than their profit, the creation and maintenance of AI is extractive and exploitive. Entering and taking from communities without consent or consideration of the communities affected.

By whom

No technology is value-neutral so no analysis of technology should be without an analysis of the power dynamics involved in its creation and deployment. The explosive growth in AI is largely driven by the hyper-wealthy and imposed upon everyone else. as methods of surveillance and control.

For what purpose

If our tools do not support our individual agency and collective flourishing, then something is extremely wrong. Increases in productivity without correlating increases in quality of life are no improvement at all. The only value AI seems to provide right now is increasing the wealth of the few and the subjugation of everyone else.

My criteria for using machine learning and AI tools

Used poorly, LLMs are dangerous bias engines. That is why I am extremely careful with how I evaluate the quality of and use the results from from the tools I use. Results are benchmarked against human analysis; every source is cited and verified; every inference is auditable and traced.

Public domain and consensual data collections:

Data used to train and feed the models I use must be from open sources. If a community of origin or author of material does not consent to their data being used, it should not be used. While there are many real world examples where ignoring the consent of an individual or a community has created value for the whole, the ends do not justify the means. Too often, the greater good has been used as an excuse to ignore the needs of historically-minoritized communities.

Data sovereignty and privacy:

My preference is to run models locally and on my own systems, rather than third party or cloud services.

Verify results against established processes and tools:

My experience is that the loudest proponents of AI tend to demand immediate adoption while refusing any level of rigorous verification. Doing so in situations where hidden mistakes can irrevocably disrupt lives and communities for decades is incredibly reckless.

Other considerations

Transparency and auditability:

Since LLMs are non-deterministic, it is important to be able to understand, evaluate, and audit the accuracy of how the outputs are generated.

Default to openess:

I prefer to use language models that have openly released weights, training code, and data.

Domain-specificity:

Generalized large-language models are often less accurate in highly-specialized fields because they lack important context.

…and there is no such thing as a virgin forest.

When European colonists came to the Americas, they thought they had found a vast world of untouched plenty. As they expanded across the land, they kept finding new ecosystems of rich natural resources, it seemed infinite and unblemished by human hands. But that bountiful Eden they found was not untouched, but rather carefully maintained, cultivated, and managed by the indigenous communities that the colonists killed as they settled the land.

AI is a human technology, its intelligence is strip-mined from our collective cultural history; its rapid growth, a cancer, draining resources from and poisoning the lifeblood of our communities. But it doesn’t have to be that way.

I have seen AI projects made by people building tools to analyze open government data to help people understand what is happening within their own communities. Projects to build tools to help case workers working with people trying to access benefits programs. Drawing specifically from trusted, domain-specific data sources, these models have been able to become accurate enough to provide a tool to support case workers and legal teams. With the growth of movement for Rules as Code for regulatory documentation, high-quality metadata, there is much potential for AI tools to help people navigate laws, regulations, and other convoluted complex systems.

The commonality across many of these uses is that they are led by people who are trying to solve problems that their communities are facing. They are using trusted, open data sources. And instead of replacing established trusted human processes, they support the people doing the work, and make the difficult parts clearer and easier.

Technology always has a stance—I am skeptical that AI like Claude, ChatGPT, or any other that are rooted in exploitation and capital investment can help us produce a world where everyone flourishes. But tools that are, as Erin Kissane says, made in better ways and for better uses, and above all, in and for communities, might.