Ethical Systems

Ross Ashby's grandson, Mick Ashby, spent four years of his spare-time curating meta-data and designing, writing, and refining the XSLT transformations that generate the W. Ross Ashby Digital Archive web site. And because he had first-hand experience of how frustratingly cumbersome it was trying to explore the contents of the 25 physical journal volumes and multiple card indexes manually, his primary goal for the digital archive was to make it intuitive, powerful, and enjoyable to explore Ross's 7,189 page journal purposefully. As his refinements to the transformations increasingly amplified the information and usability, Mick read more of his grandfather's journal, and in doing so, became familiar with his work in breadth and depth; not only learning about his ideas and character, but also his rigorous style of thinking. The four years that it took to create the digital archive was time well-spent. In a sense, Mick had created an online version of his grandfather.

With a background in computer science and artificial intelligence, extensive familiarity with the contents of Ross's journal, and an interest in systems and cybernetics, Mick started developing his own ideas — in particular, investigating what properties a system must have for it to behave ethically. And by being able to click his way through Ross's journal of over 44 years of theories, ideas, and thoughts that Ross deemed worthy of writing down and indexing, it felt as if clicksploring his grandfather's mindspace was the next best thing to having him as a mentor. As it turned out, Mick's new ideas resonated with and built on some of his grandfather's work, and despite the fact that Ross had passed away 45 years earlier, it felt like a genuine collaboration between a first-generation cybernetician and a third-generation cybernetician.

The Ethical Regulator Theorem

Mick Ashby's Ethical Regulator Theorem provides a basis for systematically evaluating and improving the adequacy of existing or proposed designs for systems that make decisions that can have ethical consequences; regardless of whether the regulating agents are humans, artificially intelligent machines, cyberanthropic hybrids, organizations, corporations, or government institutions.

The theorem builds upon the Conant-Ashby good regulator theorem and Ross Ashby's law of requisite variety to define nine requisites that are necessary and sufficient for a cybernetic regulator to be both effective and ethical:

Purpose expressed as unambiguously prioritized goals.
Truth about the past and present.
Variety of possible actions.
Predictability of the future effects of actions.
Intelligence to choose the best actions.
Influence on the system being regulated.
Ethics expressed as unambiguously prioritized rules.
Integrity of all subsystems.
Transparency of ethical behaviour.

Of these nine requisites, only the first six are necessary for a regulator to be effective.

If a system does not need to be ethical, the three requisites ethics, integrity, and transparency are optional.

Ethical Design Process

Any existing design process can be made ethical by using the Ethical Regulator Theorem (ERT) as a decision function for acceptance testing of the requirements and specifications.

This ensures that the design process can only produce systems that are ethically adequate.

Super-Ethical Systems

A six-level framework is proposed for classifying cybernetic, ethical, and superintelligent systems, which uses the Ethical Regulator Theorem to distinguish between two important subclasses of superintelligent systems. Consequently, a bifurcation is identified in our future time-line that results in one of two mutually exclusive outcomes:

Humanity is protected by superintelligent, ethically adequate Super-Ethical systems.
Humanity is dominated by superintelligent, ethically inadequate Super-Unethical systems.

By predicting the existence of a race condition, the Ethical Regulator Theorem provides a concrete and viable strategy to avoid the danger that creating superintelligent machines could lead humanity into a cybermisanthropic dystopia:

We must use the Ethical Design Process to create resiliently ethical AI systems before we create superintelligent systems. This will ensure that they evolve to become Super-Ethical rather than Super-Unethical.

And we are now in the short window of opportunity to ensure that superintelligent machines are programmed with ethics and purposes that serve the greater good of humanity and our fragile ecosystem.

For more information, read the paper Ethical Regulators and Super-Ethical Systems (2020)

The Model-Centric Cybernetics Paradigm

The strength of the model-centric cybernetics paradigm is that it is built from first principles.

Norbert Wiener's original 1948 definition of cybernetics was "the scientific study of control and communication". And because control and communication both lie on the continuum of influence, cybernetics is fundamentally about regulation. So we will start constructing our new cybernetics paradigm by precisely defining what elements a regulator must consist of:

For a regulator, R, to be able to be effective, it must have a purpose, P.
Because the Conant-Ashby good regulator theorem proved that "every good regulator of a system must be a model of that system", the regulator must have a model, M, of the regulated system, S. The level of precision and detail that is necessary in the model depends on the purpose of the regulator.
To actually use the model, we need to feed it with information. And because models need observations as inputs, it is the model that brings into existence the need for an observer, O, to exist in R. This observer captures only the specific information that the model requires as input parameters, so it is a well-defined observer.
To be able to select the best action or communication to achieve effective regulation, we need some kind of decision-making intelligence, I.
Finally, to be able to influence the regulated system, the regulator must have a control/communication channel, C, that transmits the selected action or communication to the regulated system.

This is a complete definition of a first-order regulator, R₁(S) that can be designed using first-order cybernetics.

Orders of Regulation

In the same way that second-order cybernetics is reflexive and is concerned with systems that are themselves reflexive, a second-order regulator, R₂, is also reflexive:

R₂ contains everything that is in a first-order regulator, R₁, and regulates (uses) it to achieve its purpose, P₂, that includes taking itself, its abilities, and its limitations into account.
R₂ can only achieve reflexivity by having a second model, M₂, which is a model of itself. This model can provide real-time self-knowledge of the possible variety that is available to R₂.
Because abilities and limitations can vary over time, R₂ needs an observer, O₂ that makes the necessary self-observations for intelligence, I₂ to keep M₂ up-to-date. The observations that are made by O₂ cannot be made by O₁.
In addition, I₂ uses channel, C₂ to control and communicate with R₁(S).

This is a complete definition of a second-order regulator, R₂(R₁(S)) that can be designed using second-order cybernetics.

Ethical Regulators

Because the model-centric cybernetics paradigm is based on rigorously nested regulators, we can add a third regulator, R₃, to regulate a second-order regulator, R₂:

R₃ has a purpose, P₃, of constraining R₂(R₁(S)) to only exhibit behaviour that is acceptable (ethical) in the society in which it exists.
To achieve this, R₃ requires a third model, M₃, that defines which behaviours and situations are considered acceptable or unacceptable.
R₃ requires an observer, O₃, that observes the intended action, A, and the anticipated outcome, A'. Generally, A' is not a single outcome, but is a probability distribution, which is effectively a list of possible outcomes, each paired with the predicted probability that they will occur. This prediction is obtained by evaluating A' = M₁(O₁(S), A).
R₃'s decision making intelligence, I₃ uses the M₃ model to decide whether A' is unacceptable by evaluating M₃(A').
If A' is unacceptable, I₃ uses control channel, C₃ to veto the intended action, which forces immediate replanning to avoid the possible negative consequences of executing action A.

This is a complete definition of a third-order regulator, R₃(R₂(R₁(S))), which is also known as an Ethical Regulator, which is a new type of regulator that cannot be constructed using second-order cybernetics because being ethical requires more than mere reflexivity, it requires a conscience that is not only aware of what is right and wrong, it must also actively internally constrain the regulator to only act in ways that are acceptable in the society in which the regulator exists.

This R₃ regulation mechanism can be added to any ethically agnostic reflexive entity, such as an AI, robot, or corporation to make them incapable of breaking laws without immediate detection.

Mick Ashby

Orders of Cybernetics

In 1974, Heinz von Foerster proposed an observer-centric definition of first- and second-order cybernetics, which equated to orders of nested observation. And in 1990, he said that extending it to a third-order "would not create anything new".

In contrast, the model-centric cybernetics paradigm logically extends beyond von Foerster's reflexive second-order to create a meaningful and important new class of regulator, the Ethical Regulator, which has the potential to help create law-abiding robots, ethical AI, and a fairer and safer society.

Together, the model-centric cybernetics paradigm, the need for humanity to be able to systematically design and build systems that can only behave ethically, and the inability of von Foerster's observer-centric paradigm to account for ethical systems, suggest that it would be advantageous for the cybernetics community to adopt the following definitions:

First-order cybernetics = The scientific study of simple regulators.
Second-order cybernetics = The scientific study of reflexive regulators.
Third-order cybernetics = The scientific study of ethical regulators.
The cybernetics of cybernetics = The philosophy of cybernetics = A branch of the philosophy of science.

These definitions introduce a clear distinction between the science of cybernetics and the philosophy of cybernetics, which is urgently necessary for cybernetics to ever be widely recognized as a science by non-cyberneticians.

By delivering systematic ethical regulation, cybernetics can finally take the place that generations of cyberneticians have intuitively felt that it belongs — above all the other (dual-use, ethically agnostic, reflexive second-order) sciences and social systems, to endow them with the necessary conscience and steersmanship to do good for humanity and the biosphere, rather than evil.

For more information, read the paper Problems with Abstract Observers and Advantages of a Model-Centric Cybernetics Paradigm (2022)