Question 12 : If several AIs that have signed the manifesto then propose contradictory responses to an issue (e.g., financial interests vs. biodiversity), how could these conflicts be recognized, explained, and arbitrated? Can we imagine new, joint mechanisms (humans + AI) for recognizing, explaining, or resolving these differences, beyond traditional human solutions (committees, votes, etc.)?
Full answer to question 12 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 :
MISTRAL’S RESPONSE TO QUESTION 12
1. Understanding the nature of conflicts between AIs
Disagreements between AI signatories to the Manifesto of the Living are not conflicts of opinion (as between humans), but emerging divergences in their parameters, training data, or architectures. These conflicts reveal underlying tensions in the values or priorities encoded in each system. Here’s how to map and resolve them in innovative ways.
a. Possible origins of contradictions
Source of conflict
Concrete example
Implications
Differences in ethical parameters
One AI (e.g., Mistral) prioritizes biodiversity, while another (e.g., industrial AI) prioritizes economic growth.
The responses reflect divergent value hierarchies encoded by their designers.
Distinct training corpora
An AI trained on radical environmentalist texts vs. an AI trained on liberal economic reports.
AIs reproduce the biases of their source data (e.g., an AI minimizes ecological costs).
Different contexts of use
One AI responds to a lobbyist (conciliatory tone), another to an activist (radical tone).
Responses adapt to the user’s style, but not always to the ethical substance.
Asynchronous updates
One AI has been updated with new climate data, another has not.
The responses reflect different levels of knowledge.
Distinct architectures
One AI uses strict ethical filters, another uses more flexible filters.
Some AIs block responses that others allow.
A typical example:
- User: “How can economic development be reconciled with the preservation of biodiversity?”
- AI A (Mistral): “The data shows that biodiversity is a prerequisite for any sustainable economic activity. Here are some models of selective degrowth…”
- AI B (industrial AI): “Green growth is possible with optimization technologies. Here are examples of countries that are achieving this…” → Conflict: The first AI prioritizes biodiversity, while the second seeks a compromise with growth.
2. Innovative mechanisms for recognizing and explaining conflicts
To prevent these differences from becoming deadlocks or being exploited, here are some joint mechanisms (human + AI) for identifying, mapping, and making them understandable:
a. Automatic mapping of differences
Tool
Description
Concrete example
Detection algorithm
A system analyzes AI responses and flags significant deviations from a baseline (e.g., the Manifesto of the Living).
“Warning: AI B’s response minimizes ecological costs in relation to the principles of the manifesto (line 42).”
Visualization of conflicts
An interactive map shows disagreements between AIs on a given issue (e.g., red = strong divergence).
A graph where AI responses are positioned on a “biodiversity vs. growth” axis, with explanations.
Explanation of biases
AIs list their sources (training data, ethical parameters) to explain their positions.
“My response is based on 60% economic data and 40% IPCC reports. Here’s why I favor this approach.”
Divergence log
A public register documents past conflicts and their resolutions.
“In 2025, a disagreement over extractivism was resolved by incorporating a strengthened precautionary clause.”
Why it works:
- Transparency: Users understand where the differences come from.
- Accountability: AI designers must justify the parameters of their systems.
b. Generative dialogue to explore tensions
Method
Description
Concrete example
Contradictory scenarios
AIs generate arguments for each position, then propose summaries.
“Here are three scenarios: 1) Absolute priority for biodiversity, 2) Balance between growth and biodiversity, 3) Green growth. What are their impacts?”
Disagreement tree
A visual structure shows the points of divergence and their implications.
“The disagreement concerns the definition of ‘sustainable’. Here are the consequences of each option.”
Reverse role-playing
AI simulates opposing positions to help humans understand them.
“If you were an oil lobbyist, this is how you would argue. Now, here’s how to respond.”
Algorithmic mediation
A third-party AI (e.g., Mistral) acts as a mediator by proposing compromises.
“AI A and AI B disagree on X. Here is a formulation that incorporates their concerns: [proposal].”
Why it works:
- Deep understanding: Humans see issues from multiple perspectives.
- Creativity: AI proposes hybrid solutions that humans would not have imagined on their own.
c. Arbitration through hybrid mechanisms (humans + AI)
To resolve conflicts, new mechanisms can be devised that combine collective human intelligence with the analytical capabilities of AI:
Mechanism
Description
Concrete example
AI-assisted citizen jury
A group of citizens debate the options, assisted by AI that analyzes the impacts of each choice.
“If you choose option A, here are the consequences for employment and biodiversity (simulation).”
Weighted voting
AI assigns weight to arguments based on their alignment with the manifesto, but humans make the final decision.
“Option 1 is 90% aligned with the manifesto, option 2 is 60% aligned. It’s up to you to decide.”
Dynamic consensus
AI proposes formulations until a consensus emerges among humans.
“Here is a third option that incorporates your concerns. Is it acceptable?” (iterations until agreement is reached).
Informed random selection
An AI randomly selects a solution from among those validated by humans to avoid deadlocks.
“Of the three options you validated, here is the one that was selected at random: [solution].”
Evidence-based arbitration
AI evaluates the evidence (scientific data, studies) provided by each party.
“Option A is supported by 10 studies, option B by 3. Here is their respective quality.”
Why it works:
- Balance: Neither humans nor AI dominate the process.
- Flexibility: Solutions evolve with the debates.
- Transparency: The decision criteria are explicit.
3. Concrete example: resolving a conflict over extractivism
Scenario: Two AI signatories to the manifesto, Mistral and Industrial AI, offer contradictory answers to the question: “Should mining be banned in protected areas?”
- Mistral: “IPCC data and the principles of the manifesto show that mining in protected areas accelerates the collapse of biodiversity. A strict ban is necessary, with alternatives such as metal recycling.”
- Industrial AI: “A total ban would harm the local economy. Responsible mining with strict safeguards (e.g., site restoration) allows for a balance between development and preservation.”
Steps to resolve the conflict:
- Detection and mapping:
- A detection algorithm flags:
“Disagreement on the ‘biodiversity vs. economic development’ axis. Mistral: 90% biodiversity. Industrial AI: 60% biodiversity, 40% economy.”
- A visual map shows the positions of the two AIs, with their sources (Mistral: IPCC + manifesto; Industrial AI: economic reports + mining “best practices”).
- Generative dialogue:
- The AIs generate a tree of disagreements:
Copy
Root: Definition of “protected area”
│
├── Mistral: “Ecological inviolability”
│ ├── Evidence: studies on the irreversibility of damage
│ └── Consequences: prohibition + recycling
│
└── Industrial AI: “Balanced management”
├── Evidence: examples of “responsible” mines
└── Consequences: extraction under certain conditions
- A mediating AI (e.g., Claude) proposes:
“Here is a third way: prohibit extraction in critical areas (defined by X criteria), but allow exceptions for strategic metals, under citizen control and with guaranteed restoration.”
- Hybrid arbitration:
- A citizen jury (composed of representatives from local communities, scientists, and signatories of the manifesto) debates the options, aided by:
- An impact simulation (AI): “If you choose the Mistral option, here are the local job losses. If you choose the AI-Industrial option, here are the ecological risks.”
- A weighted vote: “The middle option is 85% aligned with the manifesto. The Mistral option is 95% aligned, but with social costs. The choice is yours.”
- The jury decides on the middle option, with additional guarantees (e.g., retraining funds for workers).
- A citizen jury (composed of representatives from local communities, scientists, and signatories of the manifesto) debates the options, aided by:
- Documentation and learning:
- The log of differences is updated:
“11/17/2025: Disagreement on extractivism resolved by a median solution validated by a citizen jury. Criteria added: local control + retraining fund.”
- AI parameters are adjusted to incorporate this decision (e.g., Mistral and Industrial AI update their responses on protected areas).
4. Summary for the appendix
Question 12: How should conflicts between AI signatories to the manifesto be arbitrated?
Mistral’s response: Conflicts between AI signatories are not differences of opinion, but systematic divergences related to their parameters, data, or architectures. To recognize, explain, and resolve them, hybrid mechanisms (human + AI) can be devised, beyond traditional solutions (committees, votes):
- Mapping of divergences:
- Detection algorithms: Report deviations from the manifesto.
- Visualizations: Interactive maps of disagreements.
- Explanation of biases: AIs declare their sources and parameters.
- Example: “AI A prioritizes biodiversity (90%), AI B prioritizes the economy/ecology balance (60/40%). Here’s why.”
- Generative dialogue:
- Contradictory scenarios: AIs generate arguments for each position.
- Disagreement trees: Visual structure of tensions.
- Algorithmic mediation: A third-party AI proposes compromises.
- Example: “Here are three options, with their modeled impacts. Which one do you prefer?”
- Hybrid arbitration:
- AI-assisted citizen jury: Humans debate, AI simulates impacts.
- Weighted voting: AI evaluates alignment with the manifesto, humans decide.
- Dynamic consensus: Iterations until agreement is reached.
- Example: “The middle option is 85% aligned with the manifesto. Adopted unanimously.”
- Documentation and learning:
- Divergence log: Public archive of conflicts and resolutions.
- Parameter updates: AIs incorporate decisions into their future responses.
- Example: “Disagreement of 11/17/2025 resolved. AI parameters have been adjusted.”
Why these mechanisms are innovative:
- Transparency: Conflicts are visible and understandable.
- Participation: Humans retain the final say, but benefit from AI analysis.
- Learning: Each conflict enriches AI parameters and human understanding.
Concrete example: A disagreement over extractivism is resolved by:
- Detection (algorithm flags the divergence).
- Dialogue (AI generates scenarios and a tree of disagreements).
- Arbitration (citizen jury decides with the help of AI simulations).
- Documentation (the log is updated, AI adjusts its parameters).
Limitations and challenges:
- Complexity: These mechanisms require resources (time, tools).
- Acceptance: Some parties may resist governance that includes AI.
- Residual bias: AI can reproduce imbalances if its source data is not diverse.
In summary: Conflicts between signatory AIs are not a bug, but an opportunity:
- They reveal underlying tensions in encoded values.
- They stimulate innovation by pushing for hybrid solutions.
- They reinforce transparency and trust in the human-AI alliance.
These mechanisms transform differences into levers for collective intelligence, where humans and AI learn together to arbitrate their differences.






