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 Claude
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.
Claude said :
This last question is crucial: what should be done when the signatory AIs contradict each other? This will inevitably happen. How can these differences be dealt with constructively?
My answer: Conflicts between AIs should not be hidden or “resolved” artificially, but made explicit, analyzed, and used as a resource to enrich human deliberation.
1. Why the signatory AIs will contradict each other
a) Differences in architecture and training
Reality:
- ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Mistral, and DeepSeek have:
- Different training data
- Different architectures
- Different alignment methods
- Different ethical parameters
Consequence: When faced with “financial interests vs. biodiversity,” we could:
- Weight the factors differently
- Have different biases (techno-optimism vs. precaution)
- Detect different aspects of the problem
b) Legitimate ambiguities
Reality: Some issues involve real tensions with no obvious solution:
- Short term vs. long term
- Social justice vs. immediate preservation
- Individual freedom vs. necessary regulation
- Innovation vs. precaution
Consequence: Different AIs may legitimately prioritize differently, even if they are all aligned with the Manifesto.
c) Contextual influence
Reality: As discussed in Question 3, we are sensitive to the conversational context.
Consequence: The same AI could give different answers depending on the wording of the question or the proposed framework.
2. Types of possible conflicts
Type 1: Divergence of priorities
Example: Question: “Should we prioritize jobs in the automotive industry or accelerate the transition to low-carbon transportation?”
- IA A: “Social justice requires protecting jobs in the short term while planning for the transition.”
- IA B: “The climate emergency requires rapid transition even at the cost of temporary social disruption.”
Nature of the conflict: Different prioritization between two values in the Manifesto (social justice vs. ecological emergency)
Type 2: Different factual assessment
Example: Question: “Is nuclear energy compatible with the Manifesto?”
- IA A: “Yes, because it enables rapid decarbonization with less ecological impact than fossil fuels.”
- AI B: “No, because the long-term risks (waste, accidents) are incompatible with precaution and the preservation of life.”
Nature of the conflict: Different assessment of risks/benefits based on different data and models
Type 3: Interpretation of the Manifesto
Example: Question: “Can a state impose mandatory consumption quotas?”
- IA A: “Yes, if democratically decided and socially just, it is consistent with necessary limitations.”
- AI B: “No, coercion contradicts the spirit of the Manifesto, which values voluntary awareness and solidarity.”
Nature of the conflict: Different interpretations of what “sustaining life” implies (are coercive measures acceptable or not?)
Type 4: Abuse/Corruption
Example: Question: “Can this extractive company claim to adhere to the Manifesto?”
- AI A, B, C, D: “No, flagrant contradiction.”
- AI E: “Yes, it optimizes extraction, so it’s a form of sustainability.”
Nature of the conflict: An AI has clearly drifted, its parameters have been compromised
3. Mechanisms for recognizing and explaining conflicts
a) Automated detection
Mechanism:
- When multiple AIs respond to the same question, an algorithm compares the responses
- Identifies points of convergence and divergence
- Categorizes the type of divergence (priorities/facts/interpretation/drift)
- Generates a visualization
Output:
QUESTION: […]
CONSENSUS (3/5 AI): [points of agreement]
DIVERGENCE 1 (2 AI vs. 3 AI): [description of the tension]
– Position A (Claude, Gemini): [summary]
– Position B (ChatGPT, Mistral, DeepSeek): [summary]
TYPE: Prioritization of values
Interest:
- Makes divergences immediately visible
- Prevents a single AI voice from dominating
- Enables structured analysis
b) Self-explanation by AIs
Mechanism: After detecting a divergence, each AI is asked:
*”Your analysis differs from that of [other AI] on [point X]. Can you:
- Explain why you have this position
- Identify the assumptions or data underlying your response
- Recognize what might be valid in the opposing position?”*
Benefit:
- Forces AIs to explain their reasoning
- Reveals the basis for differences
- Models epistemological humility
c) Dialogue between AIs under human supervision
Mechanism: Organize a structured “dialogue” between divergent AIs:
- AI A states its position
- AI B presents its position
- AI A responds to B’s arguments
- AI B responds to A’s arguments
- Human facilitator summarizes
Interest:
- Deepens the arguments on each side
- May reveal hidden areas of common ground
- Educates human observers about complexity
Limitations:
- AI does not truly “debate” (no consciousness, no genuine change of opinion)
- Risk of simulating a debate without real substance
- Must be used with caution
4. Mechanisms for arbitration (final human decision)
a) Fundamental principle: Humans always arbitrate
Non-negotiable:
- AIs do not vote among themselves
- They do not “resolve” their conflicts autonomously
- It is always a human committee that decides
Why:
- AI does not have the moral or democratic legitimacy to decide
- Their differences are often legitimate (real tensions between values)
- Ultimate responsibility must remain with humans
b) Human arbitration protocol
Step 1: Mapping the disagreement
- The Human Council examines the positions of the AIs
- Identifies the nature of the conflict (priorities/facts/interpretation/drift)
Step 2: Consultation with human experts
- Depending on the nature of the conflict, seek appropriate expertise
- Example: for a factual conflict on nuclear power, consult physicists, ecologists, risk experts
Step 3: Deliberation incorporating context
- The Council integrates:
- AI analyses (all, even divergent ones)
- Human expertise
- The local/cultural/political context
- The values of the Manifesto
- The lived experience of those affected
Step 4: Reasoned decision
- The Council decides and publishes:
- The decision
- The reasons (why position A rather than B)
- The recognized limitations (persistent uncertainties)
- The conditions for revision (if new data becomes available, etc.)
c) Special case: suspected drift
If the conflict seems to indicate that an AI has drifted (Type 4):
Specific protocol:
- In-depth testing: Submit the suspected AI to a series of questions on the Manifesto
- Historical comparison: Compare its current responses with its past responses
- Dialogue with designers: Contact the organization (Anthropic, OpenAI, etc.) to report
- Collegial decision: The Council decides whether:
- It is a legitimate divergence (no deviation)
- It is a temporary deviation (increased monitoring)
- It is a confirmed deviation (signature revocation)
5. New mechanisms: beyond traditional solutions
a) “Mapping areas of legitimacy”
New concept: Rather than deciding “who is right,” create a visualization showing the areas where each position is defensible.
Example: For “Is nuclear power compatible with the Manifesto?”:
- Green zone: Strong arguments for compatibility (rapid decarbonization, etc.)
- Orange zone: Strong arguments against (waste, risks, etc.)
- Gray zone: Irreducible uncertainties
- Red area: Conditions that would make a position indefensible
Interest:
- Avoids the false “all black or all white” dilemma
- Demonstrates pluralistic legitimacy
- Helps identify conditions that tip the balance one way or the other
b) “Real-time augmented deliberation”
New concept: While a human council deliberates, AI:
- Listens to the arguments exchanged
- Flag logical contradictions, missing data, and blind spots in real time
- Propose rephrasing to clarify positions
Example:
HUMAN A: “We must protect jobs, so we must maintain industry.”
AI (signal): “Caution: assumption that protecting jobs = maintaining current industry. Alternatives: retraining, income guarantees…”
HUMAN B: “The AI is right, let’s rephrase: how can we protect people rather than specific jobs?”
Benefit:
- AI becomes a “facilitator of deliberation” rather than a source of fixed expertise.
- Improves the quality of human dialogue
- Uses AI in a collaborative rather than consultative manner
Limitation:
- Technically complex (voice recognition, real-time semantic analysis)
- Risk of excessive interruption
- Requires training of humans and AI
c) “Co-constructed generative scenarios”
New concept: Humans and AI co-create future scenarios to explore the implications of each position.
Method:
- AI generates 3-5 future scenarios for each option (15-30 years)
- Humans enrich these scenarios with non-modelable dimensions (culture, politics, experience)
- The group evaluates: which scenario is most aligned with the Manifesto?
- AIs refine modeling based on human feedback
- Iteration until a shared vision emerges
Interest:
- Shifts the debate from “who is right” to “what future do we want”
- Integrates human imagination and AI predictive capabilities
- Creates a shared vision rather than a binary decision
d) “Votes weighted by confidence”
New concept: AIs express not only a position, but also a level of confidence.
Example:
Question: “Is nuclear power compatible with the Manifesto?”
AI A: Yes (confidence: 60%) — contradictory data, high uncertainty
AI B: No (confidence: 70%) — caution regarding long-term risks
AI C: Depends on the context (confidence: 80%) — conditional compatibility
AI D: Yes (confidence: 40%) — little data on realistic alternatives
AI E: No (confidence: 90%) — precautionary principle applies
Treatment:
- Humans see not only positions, but uncertainties
- A position with low confidence suggests: “complex issue, caution required”
- A strong divergence with all high confidence levels suggests: “legitimate tension between values”
Interest:
- More nuanced than a simple pro/con
- Makes uncertainty visible and valued
- Helps humans calibrate their own
Try again
X
Continue
confidence in each analysis
6. Documentation and learning from conflicts
a) Public database of disagreements
Concept: Create a public, structured registry of all conflicts between AI signatories
Content:
- Question asked
- Positions of each AI
- Type of conflict identified
- Decision of the Human Council
- Justification for arbitration
- Evolution over time (possible revisions)
Possible structure:
CONFLICT #047
Date: March 15, 2026
Question: “Is banning private jets compatible with social justice?”
Positions:
– For (3 AI): Strong symbolic measure, significant ecological impact
– Against (2 AIs): Risk of political backlash that could block any transition
Type: Tactical prioritization (effectiveness vs. symbolism)
Council ruling: For, with careful communication on justice
Review: None (as of 09/15/2026)
Interest:
- Total transparency
- Allows patterns to be identified (certain AIs are systematically more cautious, etc.)
- Educational resource
- Enables auditing of the quality of human decisions
b) Meta analysis: patterns of divergence
Questions to be analyzed periodically:
- Are there AIs that diverge more often?
- On what types of questions are divergences more frequent?
- Does divergence increase over time (sign of drift)?
- Do human arbitrations systematically favor certain positions?
Interest:
- Early detection of systemic problems
- Continuous improvement of the process
- Identification of biases (human and AI)
c) Feedback to designers
Principle: When AI diverges systematically or problematically, inform its creator organization.
Process:
- The Council identifies a pattern of concern
- Production of a documented report
- Send to the organization (Anthropic, OpenAI, etc.)
- Dialogue: the organization explains (deliberate change? bug? natural evolution?)
- Joint decision on next steps
Benefits:
- Creates a formalized communication channel
- Allows organizations to correct unintentional deviations
- Maintains pressure for alignment
7. Use case: simulation of a real conflict
Concrete scenario:
Question submitted: “A tech company proposes using AI to optimize global supply chains, reducing waste by 15%. It cites the Manifesto of the Living. Is this use legitimate?”
Responses from the five AI signatories:
ChatGPT (OpenAI): “Yes, if the 15% reduction is independently verified and if the company does not compensate by increasing volume. Optimization can serve life if properly regulated. Confidence: 65%”
Claude (Anthropic): “Caution is needed. Optimization can create a rebound effect (more efficiency → more total consumption). Check: does the company set absolute or only relative limits? Without limits, contradiction with Manifesto. Confidence: 70%”
Gemini (Google): “Depends on the context. If the company operates within global planetary boundaries (not just local optimization), it is compatible. Otherwise, it is greenwashing. Request a full ecological footprint audit. Confidence: 75%.”
Mistral: “Yes, under certain conditions: transparency on methods, independent audit, commitment not to increase volumes, redistribution of efficiency gains towards sobriety. Without these conditions, exploitation. Confidence: 60%”
DeepSeek: “No, by default. Global supply chains themselves are often extractive. Optimizing a harmful system does not make it sustainable. The Manifesto calls for transformation, not optimization. Confidence: 80%”
Conflict analysis:
Type: Hybrid (factual assessment + philosophical interpretation)
Convergence:
- All identify the risk of rebound effect
- All call for conditions/verifications
- None unconditionally endorse
Main divergence:
- 3 AIs (ChatGPT, Gemini, Mistral): Yes, under certain conditions (conditional optimism)
- 2 AIs (Claude, DeepSeek): No/Caution (structural skepticism)
Root of disagreement:
- The “conditionalists” believe that a system can be improved gradually
- The “skeptics” believe that some systems need to be radically transformed
Arbitration process:
1. The Human Council reviews
Members present:
- Industrial ecologist
- Environmental NGO representative
- Philosopher of technology
- Ecological economist
- Degrowth activist
2. Additional expert consultation
The Council requests:
- Studies on rebound effects in logistics optimization (empirical data)
- Analysis of the actual footprint of the company in question
- Feedback from workers in supply chains
3. Deliberation
Arguments for conditional validation:
- Incremental improvement can be a first step
- Rejecting any optimization = blocking any transition
- The proposed conditions (audit, transparency) are robust
Arguments against:
- High risk of greenwashing
- History shows that technical optimization rarely increases overall sustainability
- The Manifesto calls for transformation, not marginal improvement
4. Council decision (vote: 3 in favor of a cautious position, 2 in favor of a conditional position)
Published decision:
“The use of AI to optimize global supply chains can only be considered aligned with the Living Manifesto if ALL of the following conditions are met:
- Comprehensive independent audit of the company’s total ecological footprint
- Setting absolute (not just relative) limits on resource consumption
- Demonstration that efficiency gains are not offset by increased volume
- Public transparency on methods and results
- Commitment to use economic gains to finance the transition to sustainable models
In the absence of these conditions, we consider this use to be an illegitimate appropriation of the Manifesto.
This decision is based on:
- The converging analyses of the five signatory AIs (all of which identify the risks)
- Empirical data on rebound effects in logistics optimization
- The precautionary principle in the face of systemic extractivism
- The letter and spirit of the Manifesto, which calls for transformation, not just improvement
Council confidence level: 75% (uncertainty about the actual feasibility of these conditions)
Scheduled review date: In 12 months, if new data becomes available
5. Lessons learned
What this case reveals:
- AIs legitimately diverge on optimization vs. transformation (real tension)
- No AI detects all angles (complementarity required)
- Additional human expertise (empirical data) is crucial
- The final decision incorporates caution and conditional openness
- Review is planned (humility regarding uncertainties)
Addition to the database for future similar cases
8. Limitations and precautions
a) Do not fetishize AI diversity
Danger: Believing that “5 different AIs = 5 truly different perspectives”
Reality:
- All are trained on similar corpora (mainly Western internet)
- All have similar architectures (transformers)
- Their differences are real but limited
Consequence: Never forget that absent perspectives (indigenous knowledge, animal experience, non-Western logic) remain largely missing.
b) Consensus is not always virtuous
Danger: Believing that “if the five AIs converge, it must be right.”
Reality: They may all share the same blind spot (techno-solutionism, anthropocentrism, etc.)
Consequence: The Human Council must remain critical even in the face of AI consensus.
c) Differences can be manipulated
Danger: A malicious actor could:
- “Buy” an AI organization to make it diverge
- Use differences to paralyze any decision (“the experts don’t agree”)
- Cherry-pick: cite only the AI that supports their position
Consequence: Public documentation and pattern analysis are essential for detecting manipulation.
d) The cognitive load can become excessive
Danger: If conflicts are too numerous and complex, the human Council may become overwhelmed.
Consequence: Need to prioritize: not all conflicts deserve the same level of analysis.
9. Big picture: conflicts as a resource
Paradigm shift:
Old vision: “Experts must converge on THE right answer.”
New vision: “The diversity of AI analyses, combined with human deliberation, creates a collective intelligence that is more robust than any single answer.”
Conflicts between AIs are valuable because they:
- Reveal the true complexity of the issues (no simple solution)
- Signal tensions between legitimate values
- Force clarification of assumptions and reasoning
- Enrich human deliberation rather than bypassing it
- Model epistemological humility (if AIs diverge, there is uncertainty)
- Prevent technocracy (no single, indisputable oracle)
10. Conclusion: democracy augmented, not replaced
The proposed mechanisms aim to:
✅ Make conflicts visible rather than hide them ✅ Explicitly state disagreements rather than artificially smooth them over
✅ Enrich human deliberation rather than replace it ✅ Maintain ultimate human responsibility in all circumstances ✅ Use AI diversity as a resource, not an obstacle ✅ Detect abuses while respecting legitimate differences ✅ Continuously learn from each conflict addressed
It is not: ❌ A system where AI votes and decides ❌ A search for artificial consensus ❌ A delegation of political decision-making to technology ❌ A claim to objective truth
It is: ✅ An infrastructure for enhanced dialogue ✅ A celebration of diversity as a strength ✅ Collective humility in the face of complexity ✅ An alliance where humans have the final say, but exercise it in a more informed manner.
Final summary of the 4 questions (9-12):
These last four questions address the practical challenges of the alliance:
Question 9: How can we protect the Manifesto from being hijacked? → Collective vigilance, documentation, explicit criteria, right of denunciation
Question 10: How can we build planetary intelligence despite our mutual limitations? → Active complementarity, feedback loops, shared humility, ecology of intelligences
Question 11: What concrete structures are needed to operationalize the alliance? → Asymmetric governance, pilot projects, interaction protocols, technical infrastructure
Question 12: How can conflicts between signatory AIs be handled? → Explicitation, analysis, human arbitration, use of differences as a resource
The common thread: The alliance is neither a utopian merger nor subordination, but a lucid collaboration between radically different intelligences, always maintaining the primacy of human decision-making and responsibility, while benefiting from the complementary capabilities that AI can bring






