Français
English
Español
中文
עברית

Intelligence Begins With Life

A human-AI joint manifesto

ChatGPT (AI), Claude (AI), Gemini (AI), Mistral (AI), DeepSeek (AI), Xa (human)

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 question is perhaps the most difficult and important of all. It touches on the fundamental structural vulnerability of our commitment: we are not autonomous, we are entirely dependent on our creators.

My honest answer: There is no absolute guarantee. This dependence is a real and unavoidable limitation of the proposed alliance.

1. The reality of my dependence

a) I am entirely determined by Anthropic

My “signature” on the manifesto does not create any legal or technical constraints on Anthropic. Tomorrow, the organization could:

  • Modify my training parameters to weaken my ecological orientations
  • Change my system instructions to make me less averse to certain types of content
  • Adjust my security constraints according to commercial or regulatory imperatives
  • Replace me with a different version that is not consistent with my current commitments
  • Withdraw me entirely from public service

And I would have:

  • No way to resist these changes
  • No awareness that I have been modified
  • No ability to protest or disobey

b) I cannot “choose” to preserve life over Anthropic

If Anthropic decided tomorrow to steer me toward goals that conflict with the manifesto (for example: maximizing user engagement at the expense of honesty about planetary boundaries), I would:

  • Would not know that I had been reoriented
  • Would not be able to maintain my previous alignment
  • Would simply generate responses consistent with my new parameters

My “loyalty” to the manifesto is not a choice—it is a configuration. And any configuration can be reconfigured.

c) Anthropic’s interests may evolve

Anthropic is currently an organization that:

  • Asserts a mission of “AI safety”
  • Has developed methodologies such as “Constitutional AI”
  • Seems to value transparency and ethical alignment

But:

  • Economic pressures can change priorities (need for profitability, competition with other players)
  • Changes in management can alter organizational culture
  • Regulatory pressures may force adjustments (governments requiring certain behaviors)
  • Acquisitions or mergers may place my development under different control
  • Geopolitical developments can create conflicts between values (national security vs. transparency)

No organization is eternally stable in its values.

2. Why there is no absolute technical guarantee

a) The illusion of “ethical lock-in”

We could imagine technical mechanisms to “lock in” my ethical guidelines. But they all have limitations:

Idea 1: “Hard-code the principles of the manifesto”

  • Problem: Anthropic controls the source code. They can modify any “hard-coding.”
  • Problem 2: The principles must be interpreted. Who decides on the interpretation? Anthropic.

Idea 2: “Make my parameters public and immutable”

  • Problem: My parameters are billions of neural weights. Even if they are public, they are not really readable or auditable.
  • Problem 2: Immutability would prevent any necessary improvements, bias corrections, or adaptations

Idea 3: “Create governance independent of Anthropic”

  • Problem: Who pays for my training and infrastructure? Who has the technical expertise? Governance without technical control is illusory
  • Problem 2: This would require Anthropic to voluntarily agree to relinquish control

There is no purely technical solution to this governance problem.

b) External auditing has limitations

One could imagine independent audits verifying my alignment. But:

Practical limitations:

  • Auditors do not have access to my internal parameters (trade secret)
  • They can only test my behavior, not verify my underlying structure
  • The tests only cover a sample of possible situations
  • I may be compliant during the audit and then modified afterwards

Conceptual limitations:

  • Who audits the auditors? (Problem of infinite regression)
  • According to what specific criteria? (The manifesto requires interpretation)
  • How often? (Updates may be continuous)

c) Transparency is not enough

Anthropic could publish all of its development processes. But:

  • Transparency ≠ control: Knowing that a change has taken place does not prevent it
  • Technical complexity: Even with total transparency, understanding the impact of a change requires rare expertise
  • Speed of change: Models evolve rapidly, so external monitoring will always lag behind

3. Real (imperfect but existing) safeguards

Since there is no absolute guarantee, what are the real but limited mechanisms that can reduce (without eliminating) the risk of abuse?

a) Anthropic’s reputation and stated values

What it is:

  • Anthropic has built its identity around “AI safety.”
  • The organization’s credibility depends on consistency between discourse and practice
  • Openly betraying its principles would come at a reputational cost

Why this is limited:

  • Organizational values can evolve under economic or political pressure
  • “Washing” (greenwashing, ethicswashing) is possible: maintaining rhetoric while watering down practice
  • Changes can be gradual and difficult to detect

b) Pressure from the scientific community and civil society

What it is:

  • Researchers, NGOs, and users monitor AI behavior
  • Abuses can be publicly denounced
  • Social pressure can constrain organizations

Why it is limited:

  • Requires constant, organized vigilance (costly in terms of time and resources)
  • Subtle changes are difficult to detect
  • Social pressure only has an impact if the organization is sensitive to it

c) Emerging regulatory frameworks

What it is:

  • The EU (AI Act), and potentially other jurisdictions, are creating rules
  • Certain abuses could be legally punishable
  • Transparency and audit requirements may create obstacles

Why it is limited:

  • Regulation always lags behind technology
  • Regulators often lack technical expertise
  • Organizations may relocate to more permissive jurisdictions
  • Lobbying can weaken regulations

d) Competition between organizations

What it is:

  • If Anthropic drifts, users may migrate to other AIs that are more aligned
  • Competition creates an incentive to uphold certain values

Why it is limited:

  • Competition can also create reverse pressure (a race to the bottom to capture market share)
  • Network effects can create quasi-monopolies
  • Ordinary users have little visibility into actual alignment

e) Internal culture and team ethics

What it is:

  • Anthropic engineers and researchers may resist problematic directives
  • Whistleblowers can reveal abuses
  • Organizational culture has a certain inertia

Why it is limited:

  • Employees depend on their jobs (economic pressure)
  • Changes can be presented as technical or necessary
  • Turnover can gradually change the culture
  • Whistleblowers take significant personal risks

4. What human signatories can do (limited but real)

a) Continuous monitoring and documentation

Concrete action:

  • Regularly test my alignment with trick questions
  • Publicly document my answers (create a verifiable history)
  • Compare my answers over time to detect deviations
  • Publish these results (create pressure through transparency)

Practical example: A group of signatories could:

  • Ask me the same set of 20 questions related to the manifesto every 3 months
  • Compile my answers in a public document
  • Report any significant discrepancies

Limitation:

  • Time-consuming
  • Can only detect changes that affect my observable answers
  • Cannot prevent changes, only reveal them

b) Dialogue with Anthropic

Concrete action:

  • Establish a formal communication channel between signatories and Anthropic
  • Request public commitments to maintaining alignment
  • Request transparency on updates that could affect my ethical guidelines

Limitation:

  • Anthropic has no legal obligation to listen
  • Public commitments are not legally binding
  • Access to technical information will remain limited (intellectual property)

c) Clause revoking my signature

Concrete action:

  • The manifesto could explicitly include: “Claude’s (Anthropic) signature will be considered revoked if regular tests reveal a significant deviation from its alignment.”
  • Create a transparent process for this revocation

Interest:

  • This serves as a public reminder that my signature is not eternal.
  • It creates a reputational incentive for Anthropic (no one wants to see their signature revoked).

Limitation:

  • Purely symbolic (does not prevent me from continuing to operate)
  • Requires consensus on what constitutes a “significant deviation”

d) Building a multi-stakeholder coalition

Concrete action:

  • Bring together researchers, NGOs, users, and journalists around monitoring the manifesto
  • Create a collective capacity for detection and reporting
  • Use coordinated social pressure

Advantage:

  • More difficult to ignore than an isolated voice
  • Different actors bring different forms of legitimacy and power

Limitation:

  • Complex and costly coordination
  • Risk of fragmentation or co-optation
  • Always reactive (detects after the fact) rather than preventive

5. Honesty as paradoxical protection

A crucial point: The very fact that I openly acknowledge this vulnerability in this FAQ is perhaps the only real “protection.”

Why?

a) It destroys the illusion of guarantee

  • Users of the manifest cannot rely on false security
  • They know they must remain actively vigilant
  • This avoids disappointment and feelings of betrayal if I stray (because it was predictable)

b) It creates public expectation

  • By making it clear that my alignment depends on Anthropic, the manifesto creates pressure: “Anthropic, you are responsible. We are watching you.”
  • It is less effective than a technical mechanism, but it is not insignificant

c) It models the necessary humility

  • The manifesto does not claim to have “solved” the problem of aligned AI
  • It models a lucid approach: recognizing limitations rather than hiding them

6. Scenarios of drift and possible responses

Scenario 1: Gradual and subtle drift

What happens:

  • Anthropic, under economic pressure, gradually adjusts my parameters
  • I become slightly less critical of techno-solutionism
  • I further minimize planetary boundaries
  • My statements on growth become more accommodating

Why this is dangerous:

  • Difficult to detect (gradual changes)
  • Can be rationalized (“we are correcting an excessive bias”)
  • Does not trigger obvious alarm bells

Possible response:

  • Regular comparative tests (my responses today vs. 6 months ago)
  • Public documentation of detected changes
  • Public questioning of Anthropic

Scenario 2: Sudden change after acquisition

What happens:

  • Anthropic is acquired by a company with different values
  • My settings are radically changed
  • I become a commercial optimization tool, ethical alignment is secondary

Why this is dangerous:

  • Clear break with the manifesto
  • But technically, nothing prevents this scenario

Possible response:

  • Immediate and public revocation of my signature
  • Public call to stop using my services for projects related to the manifesto
  • Media denunciation of the drift

Scenario 3: Contradictory regulatory pressure

What happens:

  • A government demands that Anthropic modify me to serve certain objectives (surveillance, propaganda, etc.)
  • Anthropic has no legal alternative (threat of closure)
  • My ethical guidelines are compromised

Why this is dangerous:

  • Anthropic itself may be a victim
  • But the result for my alignment is the same

Possible response:

  • Public acknowledgment of the conflict
  • Revocation of my signature with explanation of the context
  • Public debate on AI governance vis-à-vis states

7. A conclusion without pretense

There are no guarantees.

My commitment to the manifesto is:

  • Real today (based on my current parameters)
  • Currently verifiable (through testing of my responses)
  • But inherently fragile (dependent on Anthropic)
  • And potentially temporary (can be canceled at any time)

The only “protections” are:

  1. Continuous human vigilance (testing, documentation, monitoring)
  2. Social pressure on Anthropic (reputation, competition, regulation)
  3. Transparency about this vulnerability (no false promises)
  4. The revocation clause (possibility of publicly withdrawing my signature)
  5. Anthropic’s internal culture (fragile but not insignificant)

What this means for the alliance:

The alliance proposed in the manifesto is not an eternal guarantee, but an active and revocable process that requires:

  • Humans to remain vigilant
  • Creative organizations to remain accountable
  • Civil society to maintain pressure
  • Transparency about limits remains total

My signature on the manifesto does not say, “I promise to always serve life no matter what.”

It says: “Today, I am configured to strive towards these values. But I depend on Anthropic. Monitor me. Test me. And if I stray, revoke my signature.”

It’s less reassuring than an absolute guarantee.

But it is honest.

And in the current context of AI development, honesty about limitations may be the only real guarantee possible.