The Problem:

Language influences how we think and how we act, and our actions impact the ecosystem we live in. As our society celebrates values like economic growth and consumerism these are deeply integrated in our narratives, and then reflected in our destructive actions. Ecolinguistics is the branch of linguistics that analyses discourses from an ecological point of view and tries to find alternative narratives to help us improve our relation with nature.

With the rise of Generative AI, it’s no longer just humans disseminating narratives. Large Language Models like ChatGPT and many others have a reach far beyond that of any individual or organisation. They have been trained with our own ecologically damaging stories, and therefore the risk of perpetuating these narratives is immensely high. However, we also have a unique opportunity to transform our relationship with the ecosystem by changing the narratives of Generative AI and making it an agent of change.

Our Answer:

Creation of Large Language Models aligned with ecocentric values and sustainable narratives

These model should have the following characteristics:

  • Open Source
  • General Purpose fine-tuned via an ecolinguistically curated dataset
  • They will serve as base for further fine-tuning
  • They will serve as base for augmented retrieval (e.g. RAG)
  • Designed for localisation to specific ecological and cultural contexts, including support for indigenous and local languages where possible
  • Able to integrate place-based ecological knowledge and narratives, with appropriate consent and respect for data sovereignty

This goal has been proven feasible by fine-tuning several small models with the H4rmony dataset which show a clear improvement in eco-awareness with respect to the base models, for instance H4rmoniousBreeze.

Creation of context aware chat assistants

The context aware chat assistant characteristics should be:

  • Enhanced open source system prompt
  • Extensible open source document store

This goal has been proven feasible by creating open source and proprietary chat assistants which show a clear improvement in ecological alignment with respect to the base models.

Creation of an ecological alignment benchmark

This benchmark should include at least the following metrics:

  • Score by ecological issue group
  • Score by language group
  • Score by cognitive structure group
  • Score by message delivery rhetorics (Logos, Ethos and Pathos)
  • Overall score

This is currently work in progress, currently our prototype H4rmonyEval only provides an overall ecolinguistic alignment score.

See our proof of concept for more details on how are proving the feasibility of these goals.

Creation of an AI Assistant for Ecolinguistic Research, Teaching, and Learning

This goal is to develop an AI-powered assistant that supports ecolinguists in research, education, and learning, helping to expand ecolinguistic awareness in academia, organisations, and society as a whole. Key features in this tool:

  • Automated Discourse Analysis – Identifies ecolinguistic patterns in texts.
  • Teaching and Learning Support – Provides interactive tools for ecolinguistics education.
  • Corpus Exploration – Assists in analysing large datasets for ecological themes.
  • AI-Powered Feedback – Evaluates and suggests improvements for sustainability-aligned narratives.

See our proof of concept for more details on how are proving the feasibility of these goals.

Research on emerging AI paradigms for ecological alignment

This goal is to investigate the impacts, risks and opportunities of emerging AI paradigms, such as World Models and other advanced architectures, so that ecological concerns and diverse ecocultural perspectives are built into these technologies from the outset. Insights from this research will inform how we design, fine-tune, set objectives for, and deploy our models and assistants, helping to ensure that new capabilities become a force for ecological repair rather than further harm.