peripheric
/

It’s not humans who are at
risk of extinction from AI
It’s what makes us human

We are a nonprofit institute conducting anthropological research
on Artificial Intelligence, and developing deep tech solutions
at the intersection of culture, ecology and social complexity

Explore ↓

Generative models not only generate, they erase

Embedding spaces are the engines of Artificial Intelligence.
To build them, AI currently operates as an extractive industry, mining natural and cognitive resources at a speed and scale that puts diverse forms of knowledge, both human and more-than-human, at risk of irreversible loss.

To build truly sovereign, decentralised, and sustainable AI, we must first understand how to safeguard the things that make us human.

We take a conservation approach to identifying the cognitive and environmental impacts of embedding spaces, building the strategies, products, and communities to address them.

Dense fields of white lines interrupted by large empty spaces

We focus on three interrelated challenges concerning the
role of embeddings in human and ecological disempowerment

How we understand the world

To train language and vision models, data is turned into patterns by removing noise. The less frequently a data point (such as ideas, concepts, beliefs and practices) appears in the training data, the further away it is embedded from the pattern: they become outliers.

More than 90% of the world’s knowledge is encoded as outliers.We work with communities in the Global South to identify what models treat as noise and outliers, and develop solutions to make that knowledge available as computational structure of meaning that don’t collapse into dominant patterns.

Outliers in a vector space aren’t just encoded far away; during the training process, the model’s loss function actively penalises the model for remembering them. The model is forced to forget the unique details of the outlier to better fit the average.

For society, this means collectively falling into averages. Critical thinking requires access to different ways of understanding the past and interpreting the present: this is how humans build individuality and original thought.
Noise is what makes us human.

Active projects

Anomaly ↗

How we inhabit the world

Data centers create environmental stressors, such as noise, artificial light, and waste heat. How these disrupt and remove ecological and biological information critical for migration routes, breeding cycles, feeding patterns and other adaptive systems is severly understudied and in urgent need of attention.

Data center construction is expected to triple by 2027We are developing strategies to safeguard the environments in which AI infrastructure is built, and solutions to reduce its ecological and societal impact based on local conditions.

Data centers are instrumental for building embedding spaces and making patterns computationally actionable. It is well established that these processes require massive amounts of water and energy, but their impact to the ecosystems around them has received scarce attention.

We know from other extractive industries that by products like noise, light and heat can act as environmental stressors, which can disrupt and destroy entire ecosystems without appearing to do so. If we don’t focus on these challenges now, it could take years, or decades, for the ecological impact of data centers to become visible.

Active projects

How we imagine the world

We are outsourcing imagination to generative models that lack the context to map the natural language of a prompt to the semantic intent of its author, instead replacing individual structures of meaning with the most prominent pattern. AI not only lock us into software infrastructures, but also cognitive ones.

By 2027, AI-generated content will surpass human content onlineWe are developing computational processes that encode systems of representation at the user scale and compound outliers so that multiple futures can be rendered possible without falling into patterns.

Imagination used to happen primarily in our brains, feeding on epistemological structures unique to each of us (such as semantic, semiotic, and ontological vocabularies). Now, we are outsourcing imagination to generative models that lack the context to map the natural language of a prompt to the situated meanings and knowledge structures of its author. This is why AI-generated content feels generic, repetitive, and unoriginal.

As a result, cognitive liberty, divergent thinking, creative expression, and other aspects of cognitive agency that are critical for individuality are being lost with every training run, prompt and circulation of AI-generated content.

Active projects

Unpattern ↗

Conservation as a method for AI research and development

Approaching the lifecycle of AI as a shared ecosystem requires a conservation paradigm that addresses not just isolated points in training, physical infrastructure and use, but the feedback loops and wider effects that connect them. We work across three pillars to develop solutions that leverage AI’s potential for good in balance with the systems it depends on.

Anthropology of AI

We work collaboratively with the communities and ecosystems we strive to support to identify which forms of knowledge, expression and ecological relation are displaced, appropriated or erased through AI’s training, infrastructure and use.

Appropriate AI

We develop small-scale, resource efficient, domain-specific, and locally governed artificial intelligence tailored to specific constraints, striving for data sovereignty and cultural alignment by prioritising open-source architectures and locally curated data.

Conservation...

We carefully implement solutions through products that preserve, protect and strengthen what is at risk, while allowing users and ecosystems to thrive through responsible, inclusive and ethical use of AI.

Interested in our
conservation efforts?

Preview form — no information is sent or saved.