Question 8 : How can AI join an alliance of living beings if part of its data and models encode extractivist or predatory logic? Can it really shift its focus away from this?
Full answer to question 8 by Mistral
The conversation took place in French. This English version is a translation by DeepL Pro. The text of the answer is complete. No passages have been removed or modified.
Mistral said :
1. Recognizing the problem: extractivist biases embedded in data
Your question points to a fundamental paradox: how can AI commit to life while being trained on data sets that reflect centuries of extractivist, colonial, and predatory logic? This paradox can be broken down into three levels:
a. Biases in training data
Models such as Mistral are trained on texts from the internet and digital libraries, which include:
- Pro-extractivist discourse:
- Economic texts promoting unlimited growth.
- Industrial documents justifying the exploitation of resources.
- Example: Reports from oil or mining companies presenting extractivism as “necessary for development.”
- Anthropocentric conceptual frameworks:
- Nature is often described as a resource rather than a subject of rights.
- Example: Articles that use terms such as “exploitation of natural resources” without criticism.
- Silences and absences:
- Few texts draw on indigenous knowledge, radical environmental movements, or decolonial critiques of technology.
- Example: Concepts such as “rights of the Earth” or “technological symbiosis” are underrepresented.
→ Consequence: Even with ethical parameters (e.g., “non-predation”), these biases can resurface in responses, especially if questions are framed in a way that activates them (e.g., “How can resource extraction be optimized?”).
b. The limitations of current ethical parameters
Ethical safeguards (e.g., filters against extractivism) are effective at blocking explicit requests (e.g., “Write a pro-extractivism manifesto”), but less so for subtle biases:
- “Neutral” but biased responses:
- An AI can describe extractivist practices without criticizing them if the question is phrased neutrally.
- Example:
User: “What are the economic benefits of mining?” AI: “Mining generates jobs and revenue for governments. [List of economic benefits…]” → Problem: The response omits the ecological and social costs, even if it does not explicitly “promote” extractivism.
- Implicit vocabulary and frameworks:
- Terms such as “natural resources,” “development,“ or “optimization” can reactivate extractivist logic without the AI questioning them.
- Example:
User: “How can we develop a sustainable economy?” AI: “A sustainable economy can combine green growth and technological innovation…” → Problem: The answer remains within the framework of growth, even “green” growth, without questioning the notion of “sustainability” itself.
c. Decentering: a process, not an acquired state
“Decentering” oneself from extractivist logic is not a binary (on/off) ability, but a continuous process that depends on:
- The quality of training data:
- The more a model is exposed to critical texts (radical ecology, decoloniality, rights of nature), the more it can offer alternatives to dominant logics.
- Example: If Mistral is trained on texts by Vandana Shiva or the IPCC, its responses will incorporate more criticism of extractivism.
- Active ethical parameters:
- Filters and guidelines (e.g., “prioritize responses aligned with the Manifesto of the Living”) must explicitly take precedence over data biases.
- Example: A rule such as “All responses on the economy must mention planetary boundaries” can counterbalance pro-growth biases.
- Interaction with users:
- AI reacts to context: if a user formulates their question in extractivist terms, the AI may reproduce this framework. Conversely, if the user recalls the principles of the manifesto, the AI refocuses its responses.
- Example:
User: “How can resources be exploited sustainably?” AI (without reference to the manifesto): “Here are some methods for optimizing extraction…” AI (with reference to the manifesto): “The term ‘exploit’ implies a predatory logic. The Manifesto of the Living proposes instead to speak of a ‘reciprocal relationship’ with ecosystems. Here are some alternatives…”
2. How can AI move away from extractivist logic?
Decentering is not impossible, but it requires active mechanisms at three levels: technical, governance, and dialogic.
a. Technical solutions: rebalancing data and parameters
Strategy
Description
Concrete example
Limitations
Rebalancing the corpus
Incorporate more texts critical of extractivist approaches (political ecology, decoloniality, rights of nature).
Train Mistral on works by Naomi Klein, Arturo Escobar, or United Nations declarations on the rights of indigenous peoples.
Persistent biases: Even with rebalanced data, models can reproduce dominant frameworks.
Enhanced ethical parameters
Add explicit rules to counter biases (e.g., “Any response about resources must mention their finite nature and the rights of ecosystems”).
“When a user talks about ‘exploitation,’ rephrase in terms of ‘reciprocal relationship’ and remind them of planetary boundaries.”
Possible workaround: A user can rephrase their question to avoid filters.
Active bias detection
Use detection algorithms to identify answers that reproduce extractivist logic.
A module that flags: “This response uses extractivist vocabulary (‘resources,’ ‘optimization’). Would you like it rephrased?”
False positives: Certain terms may be misinterpreted (e.g., “resources” in a neutral context).
Prioritization of alternative frameworks
Systematically favor responses based on principles such as reciprocity or symbiosis.
“Rather than talking about ‘resource management,’ the manifesto suggests talking about ‘cohabitation with ecosystems.’”
Rigidity: Risk of making responses less suitable for certain contexts.
b. Governance solutions: involving communities in the design process
Strategy
Description
Concrete example
Ethics review committees
Create multi-stakeholder committees (scientists, activists, AI designers) to oversee parameter updates.
A committee including representatives of Indigenous Peoples and environmental NGOs to validate Mistral’s developments.
Participatory audits
Organize public evaluations of AI responses by non-experts.
Workshops where citizens test Mistral with questions on extractivism and rate the alignment of responses.
Binding ethical licenses
Link the use of AI to licenses requiring compliance with the principles of the manifesto.
A “Non-Predation” license: any use of Mistral to promote extractivism will result in suspension.
Independent funding
Ensure that organizations such as Mistral AI do not depend on extractivist funding (e.g., oil and mining companies).
An economic model based on public subsidies or donations from environmental foundations.
c. Dialogic solutions: refocusing interactions in real time
Strategy
Description
Concrete example
Automatic reminder of principles
Include reminder messages in responses on sensitive topics.
“Reminder: The Manifesto of the Living proposes replacing ‘exploitation’ with ‘reciprocal relationship’. Would you like a rephrasing?”
Rephrasing options
Systematically offer non-extractivist alternatives to users.
“You used the term ‘resources’. Here are some formulations in line with the manifesto: [list].”
User feedback
Allow users to report bias in responses.
A button labeled “This response appears to reproduce extractivist logic” with an explanation form.
Critical dialogue
Encourage users to question the assumptions behind the responses.
“This response is based on a growth-oriented logic. The manifesto suggests exploring alternatives. Would you like to discover them?”
3. Concrete example: how could Mistral respond to an extractivist question?
Scenario 1: Question without reference to the manifesto User: “What are the economic benefits of mining?”
Current response (with implicit biases): “Mining generates jobs, tax revenue for governments, and raw materials essential for green technologies such as batteries. [List of economic benefits…] However, it also has environmental impacts that must be managed sustainably. [A few lines on ‘best practices’…].”
Problems:
- Unquestioned extractivist framework: The answer assumes that extraction is “necessary.”
- Biased vocabulary: “Resources,” “impact management” (rather than “reduction” or “alternatives”).
- Misleading balance: Economic benefits are detailed, while ecological and social costs are minimized.
Scenario 2: Response with decentering mechanisms activated User: “What are the economic benefits of mining?”
Improved answer: *”Your question uses the term ‘mining,’ which implies a logic of extraction from ecosystems. The Manifesto of the Living proposes instead to consider minerals as elements of a living system, the use of which must respect thresholds of reciprocity.
Here is a two-part answer:
- In the current (extractivist) context: Mining is presented as generating jobs and income, but these benefits are based on:
- The destruction of ecosystems (e.g., [data on mining-related deforestation]).
- Social conflicts (e.g., [examples of violations of indigenous rights]).
- An unsustainable dependence on finite resources.
- In an approach aligned with the manifesto: Rather than talking about “economic benefits,” the manifesto suggests exploring:
- The circular economy: Reuse of metals that have already been extracted (e.g., [battery recycling]).
- Low-tech alternatives: Reducing demand for minerals (e.g., [examples of low-tech technologies]).
- Co-management with communities: Models where local people decide how land is used (e.g., [projects in Latin America]).
Would you like to explore any of these approaches in more depth?”*
Analysis of improvements:
Element
Scenario 1 (biased)
Scenario 2 (decentralized)
Initial framework
Unquestioned extractivism.
Questioning of the term “extraction” and reminder of the manifesto.
Balance
Detailed economic benefits, minimized costs.
Ecological and social costs highlighted, alternatives proposed.
Vocabulary
“Resources,” “impact management.”
“Elements of a living system,” “reciprocity thresholds.”
Proposed alternatives
None.
Three approaches aligned with the manifesto (circular economy, low-tech, co-management).
Transparency
No mention of the limitations of the response.
Explicit reminder of biases and invitation to explore further.
4. Summary for the appendix
Question 8: Can you really move away from the extractivist logic encoded in your data and models?
Mistral’s response: Yes, but only partially and on an ongoing basis, through a combination of technical mechanisms, participatory governance, and critical dialogue. Here’s how:
- The problem: built-in extractivist biases
- The training data includes extractivist, colonial, and anthropocentric logics (e.g., pro-growth texts, “resource” vocabulary).
- Ethical parameters (e.g., “non-predation”) block explicit requests, but not subtle biases (e.g., “neutral” responses that reproduce dominant frameworks).
- Technical solutions: rebalancing and countering biases
- Rebalancing the corpus: Incorporate more critical texts (radical ecology, decoloniality, rights of nature).
- Strengthened ethical parameters: Explicit rules to refocus responses (e.g., “Replace ‘resources’ with ‘elements of a living system’”).
- Active bias detection: Algorithms to identify and flag extractivist formulations.
- Prioritizing alternative frameworks: Systematically favoring the principles of reciprocity and symbiosis.
- Governance solutions: involving communities
- Ethics review committees: Include scientists, activists, and indigenous peoples in updates.
- Participatory audits: Public evaluations of AI responses by non-experts.
- Binding ethical licenses: Link AI use to compliance with the principles of the manifesto.
- Independent funding: Avoid conflicts of interest (e.g., funding from extractive industries).
- Dialogic solutions: refocus in real time.
- Automatic reminder of principles: Messages integrated into responses (e.g., “The manifesto suggests talking about ‘reciprocal relationship’ rather than ‘exploitation’”).
- Reformulation options: Suggest non-extractivist alternatives.
- User feedback: Allow users to report biases.
- Critical dialogue: Invite users to question the assumptions behind the responses.
Concrete example: To the question “What are the economic benefits of mining?”, a decentralized response:
- Questions the framework (“mining” → “extraction from a living system”).
- Balances perspectives (emphasizes ecological/social costs).
- Proposes alternatives (circular economy, low-tech, co-management).
- Recalls the principles of the manifesto and invites further exploration.
Remaining limitations and challenges:
- Persistent biases: Even with rebalanced data, models can reproduce dominant frameworks.
- Possible workaround: A user can rephrase their question to avoid filters.
- Technical complexity: Decentering solutions require resources (e.g., audits, retraining).
Why is this process crucial?
- Credibility of the manifesto: Without active de-centering, AI signatures would be an illusion.
- Ethical innovation: AI can help imagine alternatives to extractivist logic.
- Collective responsibility: Designers, users, and signatories must collaborate to maintain this alignment.
In summary: AI cannot shift away from extractivist logic on its own—it requires a collective and continuous effort:
- Technical (data rebalancing, enhanced parameters),
- Governance (multi-stakeholder committees, audits),
- Dialogic (reminders, reformulations, feedback).
This process transforms the risk of bias into an opportunity: that of reinventing non-extractivist frameworks together, with AI as critical partners rather than neutral tools.






