Beneficial Artificial Intelligence in the Context of […]
The following is concerned with the topic of beneficial artificial intelligence. “Beneficial”, in this context, means that systems that use AI act in a way that is desired by humans. Therefore two models will be suggested: the first approaches a classification for aspects of beneficiality in artificially intelligent systems. The second one explores methods to capture and counter ethical problem spaces. The defined models will be explored in three speculative use cases to exemplify and refine them.
Mar - July 2019
Hochschule für Gestaltung, Schwäbisch Gmünd
The impact areas of AI are growing constantly as agents advance in capabilities.
As soon as AI systems influence our daily life – and in many ways they already do – it becomes essential to design them to act in a beneficial way: now and looking forward. Establishing ethics as an integral part of design, economics, engineering and development will take time and lots of dedicated work. The constant increase in intelligence makes work on ethical systems a relevant issue long before artificial superintelligence has been attained. A system does not need to be superintelligent to cause serious issues. For example, as the autonomy in the operation of those agents increases, entirely new questions, for example, in the area of accountability for action, will arise.
we believe that…
Technology & ethics have to be established as a singular vision to enable beneficial outcomes of AI. We believe that the future development of AI must serve an overall purpose. It must preserve and enhance human life, as well as preserve and enhance the planet and its eco-system. There are great chances for AI to do so, but there are also numerous risks where AI could lead to an extremely negative outcome. It is of utmost importance to focus our efforts on the positive chances AI can enable us and to specifically target the risks that might occur, so that they can be avoided. We do not believe it is a viable option to consider reversing or degenerating the development of AI, as the pace of development is already fairly high. This unavoidable development strengthens our focus and the overall relevance on working towards the beneficial use of AI.
Beneficial Artificial Intelligence in the Context of […]
Approaching beneficial AI requires approaching what is to be considered "beneficial". Maximizing beneficiality in an artificial agent’s behaviour can be seen as a fundamental goal in ensuring its safe operation and the promotion of goals considered broadly beneficial. To structure different actions and cluster steps relevant in approaching a system’s beneficial behaviour, the diverse aspects that contribute to the process can be classified in a model. The layers represent key areas of focus when designing beneficial agents. This model is not a conclusive instruction to achieve beneficial behaviour, but rather is supposed to give guidance and establish beneficiality in the developer’s thinking.
Beneficiality depends on the use case, those affected by an agent's operation and their values
Beneficial AI is a multidiciplinary endeavour that cannot be solved, only approached
Beneficial behaviour has to be established from the beginning as an integral part of the system
… Suggesting a Framework to
approach beneficial AI
A completely bias-free system is obviously an unattainable goal, yet it is essential, that this goal is pursued best possible.
Make sure the system avoids doing harm to humans in every situation to the best extent possible.
The system should value a persons privacy wherever it is not required, that data is shared.
Benefits, that occur through the system should not be exclusive to a certain few, but available to the general public, as far as this is feasible.
The capabilities the system is given should be restricted to exclusively those that are required to achieve its goals.
The metaphorical “stop-button” should be included in the systems design, so that it is possible to detain the system in case of unintended consequences. Backups will enable recovery to safe states.
The safe operation of an agent has to be tested in regular operation as well as edge case scenarios. Possible failure has to be communicated transparently.
An agent should start exploring spaces that can be deemed “safe” and controllable, as it proves beneficial in operation the testing environment can gradually be improved to cover broader scenarios.
The system must enable qualified operators to have access to its architecture.
Qualified operators must be able to comprehend how the system acts.
It must be able to trace back the reasoning for every resulting action of the system.
Failed operations should be communicated clearly by the system, so that further investigation is possible.
The monitoring entity should retain the capability to intervene with a running system at all times.
Regulation for who is to be held accountable in case of failure, non-beneficial actions or other unintended consequences of the system need to be in place.
An entity must be employed that permanently oversees the systems actions and consequences thereof.
Constant review and verification of a systems robustness based on current standards must take place so ensure beneficial behaviour.
In case of a crucial failure of the system, measures to prevent further consequences must be in place – absolute dependability on the system must be avoided.
Crafting around beneficiality is an interdiciplinary, iterative process involving many diciplines
The debate about how a “common beneficial” should be defined takes place on multiple layers across many disciplines and spans a long timeframe. Though general rough understanding of what is beneficial exists in many cases (e.g., no human should come to harm from an action), it is important to note that actions that are considered beneficial by an individual are not necessarily the most ethical actions. It can be assumed that there is no such thing as an “absolute perfect beneficial” for all affected parties in a situation. Consequently a lot of consideration has to be placed on defining the agent’s value system to guide the process of finding a sufficient "beneficial" for its actions. Furthermore ensuring beneficial behaviour is a recursive endeavour, not a linear process. All of the pillars have to be regularly revisited as circumstances change (e.g. an agent’s level of intelligence increases).
Problem space are complex ethical issues in the way of beneficial AI
Problem spaces can only be approached and their effects minimized, but not always solved.
Problem spaces are often hidden patterns and have to be actively uncovered to be approached.
Problem spaces have many causes and have to be approached by multiple measures.
Regarding the potential issues that arise with the creation of artificially intelligent systems we suggest a classification for bigger ethical problems. Approaching these issues is essential, as they conflict with the goal of an agent’s beneficial behavior in action. The concept of problem spaces and the suggested approaches represent an ongoing process and shall serve as guidance and inspiration for further work by ourselves and others. Problem spaces are not necessarily the most obvious problems one might think about when developing an artificial agent, but those with potential long term consequences that can result in serious implications for individuals and society. Problem spaces are high level patterns, existing across a broad range of similar scenarios and manifest themselves as concrete problems in different use cases.
… Approaching problem spaces
We define a problem space as a class of issues that arise when working on intelligent artificial agents that are supposed to act in a beneficial way. Problem spaces are not simple, concrete issues, but rather a space of underlying patterns that manifest themselves in more concrete problems in different use cases.
As problem spaces are abstract issues with many different causes, a single approach is not suitable: problem spaces have to be separated into individual smaller issues that can then be approached by different actions. These actions are the connecting piece between abstract problem spaces and idealized goals.
Each problem space has a counterpart that describes an ideal state. This idealized goal is based on what is considered beneficial and / or morally desirable. It’s not necessarily the complete elimination of the original problem – it rather describes a sufficient stage of a feature, so that it is beneficial.
in the Context of […]
To refine the models and demonstrate how the process of framing and approaching problem spaces can be applied to the creation of artificially intelligent systems we created three rough use cases. Since the underlying objective is to design the applications in the use case in a beneficial way the ‘pillars of beneficiality’ were used as guidance in regard to what should be achieved. The three use cases are set in different scenarios, each taking one step further into the future. The “beneficial layer” in each use case explains the different counter measures that help countering the problem spaces.