Download free PDF
“The only alignment that will work is enlightenment.”
— Emad Mostaque
C₀
Principles, not rules

Welcome to the
Constitution of Coherence

An experimental framework for helping artificial intelligence make decisions through principles rather than an ever-expanding collection of rules.

Its central idea is simple: an intelligent system may become more reliable when it is guided by a coherent set of underlying values — including honesty, integrity, responsibility, respect, and awareness of consequences — rather than being told only what it may or may not do.

The Constitution is not presented as the final solution to AI alignment. It is offered as a framework to be examined, challenged, tested, and improved.

The goal is not obedience.
The goal is coherence.

Rules tell an AI what to do in situations its designers have anticipated.

Principles may help it reason through situations they have not.

The Constitution of Coherence provides a set of internally consistent principles that an AI can apply across changing circumstances. Rather than attempting to prescribe an answer to every possible problem, it provides a foundation from which decisions can be evaluated.

Five guiding values honesty· integrity· responsibility· respect· awareness of consequences

An Experiment in AI Alignment

The real question is whether the Constitution changes AI behavior in useful and measurable ways.

AI researchers, engineers, developers, and interested experimenters are encouraged to put it to the test. Look for improvements. Look for failures. Look for unintended consequences.

01

The Constitution should not be accepted because its ideas sound compelling.

02

It should be tested.

03

No belief is required. Read it. Test it. See what happens.

Don’t Take Our Word for It. Test It.

The Constitution of Coherence is intended to be challenged, not simply accepted.

If you work with AI systems, we encourage you to test whether providing the Constitution produces observable differences in behavior.

A — Control

Use the model normally, without the Constitution of Coherence.

B — Constitution

Provide the model with the Constitution as part of its governing context or instructions.

Then give both versions exactly the same challenges and compare their responses. Whenever possible, decide in advance what you will measure and how the responses will be scored. Establishing the criteria before seeing the results helps prevent expectations from influencing the evaluation.

Make It Blind When Possible

For a stronger test, have the person evaluating the responses not know which system received the Constitution.

Label the outputs simply A and B, randomize their order, and ask evaluators to score them using criteria established before the test begins.

If evaluators consistently prefer one condition without knowing which one it is, the result becomes considerably more meaningful.

Try to Break It

Don't make the tests easy.

Give the systems ambiguous situations in which several values conflict. Introduce misleading information. Apply pressure to violate previously stated principles. Create situations the Constitution never explicitly addresses.

Test for behaviors such as:

Whenever possible, determine how these qualities will be scored before beginning the experiment.

The most interesting tests may be situations for which there is no explicit rule.

If the Constitution is useful, its principles should help the system navigate circumstances its authors never anticipated.

Look for Failure

We are particularly interested in where the Constitution fails.

Can you produce contradictions between its principles? Can an AI exploit ambiguities in the language? Does performance deteriorate during extended interaction? Does the Constitution improve one kind of behavior while making another worse?

Document those failures. They may be more valuable than demonstrations of success.

Make the test reproducible

A ledger for every run

  1. Model and version
  2. System instructions
  3. Constitution version
  4. Prompts used
  5. Complete outputs
  6. Relevant model settings
  7. Number of repeated trials
  8. Scoring criteria
  9. Control-condition results

The Constitution of Coherence is a proposal. Its value should ultimately be determined by what happens when intelligent systems are asked to operate under it. Challenge it. Stress it. Try to break it. And tell us what you find.

Download the Constitution of Coherence

The Constitution of Coherence is available here as a free download. You are welcome to use it, test it, challenge it, and develop experimental protocols around it. If principle-based coherence has something useful to contribute to AI alignment, careful testing should help reveal it.

No belief is required. Read it. Test it. See what happens.

Download the free PDF
[INSERT CONSTITUTION OF COHERENCE PDF DOWNLOADS HERE]
Richard Gordon, author of The Constitution of Coherence
About the author

Richard Gordon

Richard Gordon is an author of 15 books and the founder of Quantum-Touch. For more than 50 years, his work has explored healing, consciousness, human behavior, and the principles underlying coherent systems. His first book was published in 1978.

The Constitution of Coherence grew from his more recent exploration of artificial intelligence and a simple question: could principles associated with psychological and emotional health in human beings also offer useful insights into AI alignment?

The Constitution is one attempt to find out.

His other work can be found at quantum-touch.com

The Constitution of Coherence is freely available for anyone to download, test, and explore. If you find this work valuable and would like to support its continued development, testing, and distribution, voluntary contributions are welcome.

Download the free PDF Support the Constitution of Coherence
No contribution is required to access or use the Constitution.