The Elements of AI Conscience.

Article Short Summary

Today's vast and complex knowledge and data is difficult to combine and use optimally, but AI technology can help understand it and create new solutions and technologies from it. But to keep up, we need an AI conscience that ensures the solutions are optimized to be beneficial and benevolent.

To create an AI conscience, we need three steps: 1. properly identify a situation and assemble relevant data, 2. use input in a way that inherently involves the AI conscience, 3. create a solution that best solves the identified situation.

Regarding the first step, we need to perceive as many facets as of the situation as possible and recognize all the present and non-present stakeholders involved in the situation. Thus, for the first step we need a variety of sensor technology, connectivity, extensive access to data about the stakeholders, and scientific and social knowledge.

For the second step it is critical to ensure the correctness, quality, and benevolence of the generated solutions. A realistically simulation provides understandability, decision tracing and exploration of different solutions to identify their implications and find the best one. Sophisticated evaluation criteria, which are basically constraints and guidelines, are used by the AI conscience to ensure the quality and benevolence of the solution generated in the simulation. Thus, for step two, we need simulation technology, scientific data and knowledge, information about people's values, needs, and wishes as well as a large set of complied evaluation criteria.

Turning to the third step, the generated output can either be an explanation or proposed actions to resolve the situation step by step. The third category of output is new technologies to achieve defined objectives. In addition, the AI must provide information about the decision-making process and statistics.

Introduction

Before we start,?let me give you a very brief definition of a human conscience, which serves as the role model: “Conscience is a moral sense of right and wrong that is considered as a guide for one's behavior.”

The last article was about ‘Why we need AI Conscience’. That article argued that we live in an ever more complex world with exponentially rising knowledge and content and that humanity struggles to comprehend and incorporate this knowledge beneficially. And to be able to use our knowledge more and more in a beneficial way, we start to use AI to process humanity’s knowledge, but we need to substantially extend its usage in the future to keep up with global calamities and individual issues. But by doing so, we would create an immensely powerful AI which is not without justifiable concerns and trust issues, mainly because humanity is not yet capable to create a safe learning dataset and thus leading to dangerous AI decisions.

The article then argued that dangers of AI can be avoided if we use a suitable evaluation system for AI decision-making, namely a conscience developed for AI. We can align AI conscience with our values, needs and wishes, thus create a safe AI and with that build the much-needed trust in this technology.

This article is the second part that introduces AI conscience, and it focuses on the question “What are the elements of an AI conscience.”. This article deals with the concept from a top-level view and aims at proposing necessary parts and approaches. Articles with more details and concrete implementation suggestions may follow.

Now, let us look at the approach we can take to create an AI conscience. When it comes to building it the first two questions should be “What are the specific objectives we try to meet?” and of course “How can we create AI conscience and meet these objectives?”. We can phrase a very simple objective using the conscience formulation as “Giving an AI the ability to be able to use conscience in any situation.” And as a note, we really want to assume here that it could be applied in any situation, whether we want to use it for technologies against climate change, for car engine optimization, data storage, work processes, craftsmanship, or just when people quarrel, or we want to search information about a book.

To meet the objective, we can derive the following general steps:

1. Input: Identifying a situation correctly and compile relevant data.
2. Processing: Use the input in a way, that inherently incorporates the AI conscience.?
3. Output: Create a solution that solves the identified situation the best way possible.

The next chapters will concentrate on these three steps and their inner workings.

Identifying a situation correctly and compile relevant data.

To ensure the broad applicability of the AI and its respective AI conscience, we must be able to perceive the situation we want to use AI conscience in. Additionally, we need to access knowledge about the situation and the involved parties as well. Ideally, the AI must also have access to the values, needs and wishes of all involved parties to develop the most beneficial solution. These inputs are integrated in the processing and decision-making stage later on.

Let us first look at the pure perception of a situation. We can build a perception by imitating the five classical senses we humans possess, namely sight, hearing, touch, smell, and taste. Other perceptions for AI could for example also be the magnetic and electric field, detection of IR and UV light, thermal sensors, air pressure and basically any other sensors we have developed. All these should be incorporated in an AI to perceive a situation in as many facets as possible, as information enables better decision making for Ais, just like it does for humans. Widening the range of perception objectifies the perception and creates a more reliable decision base. In contrast, the fewer different senses we implement the more elements of a situation would be open to discussion and we would ultimately ignore possible solutions.

While getting sensor data is quite forward, identifying involved parties and their internal values, needs and wishes is not. There are parties, or elements, in every situation, that can be perceived and recognized quite easily, but often a situation also involves or even originates from elements that are not present in the currently perceived situation. To generate decision with conscience for complex problems we need the bigger context and the broad incorporation of remote but linked parties, too. Thus, the AI must be able to map the recognized present elements to relevant non-present elements. Or in other words, it must be able to understand the contextual information of a situation.

But even, if we manage to get a significant amount of contextual information, purely perceiving a situation would still not be enough. We need to understand it. Meaning, we need to connect the raw data to a knowledge base. We need data about the elements of that situation, data how the elements are linked between each other and to other non-present elements, and data about these non-present elements. The data we are talking about is for identifying what these elements are, but also about their properties and characteristics, their behavior, history, origin, and basically everything we know about them.

Another input the AI must gather, are the goals of the involved parties, so the AI can generate a satisfying solution. But for that we need the actual objectives i.e. goals. So, we need to raise questions like: When is this challenge solved? Is there an objective goal or only individual goals? How is the weighing of these goals to each other? And what are the implications of these goals? What are the thresholds for the different goals? And we can think of more questions to which we need answers before we start to generate solutions.

While some information on goals and objectives can be given by the directly involved parties, goals also need to be implicated by the AI. As it is all too often the case, people do not fully understand their own needs, values and wishes and more specifically the interrelations of these to their environments and to the other stakeholders in the case. Thus, the AI needs to be able to look deeper into the motivations particularly the underlying interrelations of the perceived situation, the stated objectives and the correlation and implications that these aspects have. It is certain that the AI must not only perceive but process all this information in a very sophisticated way to find a solution that is beneficial to all parties, and thus act with "conscience" no matter what challenge the AI faces. And here is where the true complexity starts.

How can we manage and process this complexity with conscience in mind? Let us dive into that aspect a little deeper.

Use the input in a way, that inherently incorporates the AI conscience.?

Taking all the inputs, we need to be able to process that data in a way, that allows us to explore different possible solutions, without effecting the real world, and evaluate them according to a wide range of criteria, that are as inclusive and fair as possible. The output of the AI might affect the lives of people and might have far reaching impact. So, the processing cannot be done blindly, but must be implemented very carefully and consciously.

For any AI generated solution to be desired and trusted and to give us safety, verifiability, and comparability of different solutions, we need to test and foresee the effects and implications of these outputs. In short, for trustworthy AI solutions, the output must be generated and tested with an AI conscience that we can comprehend and approve. And this is what the processing must deliver.

The currently best way to generate and evaluate solutions to the above standards is to use a simulation, or with other words a training and test environment. In the simulation we can generate solutions of different output types (explanation, action, technology) and use them in an environment representing the real situation. This way, we can test the performance of the different solutions and we have the freedom to move exploratively between very different solutions, optimize them and create new knowledge along the way. This is preferred to one-shot solution generation, where a single output is static and only relies on previous AI training, and thus does not account for side-effects and probable deviations.

Another benefit of a simulation is that we can use the simulation itself and all the collected data from the simulation runs to explain and check the conclusions the AI drew. We can task experts to evaluate the AI’s decision-making and go to specific points in the simulation to do what-if experiments. So, a complex and sophisticated simulation environment could create just the trustworthiness and clarity needed.

However, to truly meet the level of functionalities we pursue with the simulation, the simulation cannot be just mediocre. It must meet many special requirements. Among other things, we need it to use accurate physics and simulate agents realistically, lots of variables need to be observable, the AI needs to be able to develop concepts and objects inside the simulation, the simulation procedure needs to be objective driven, and it needs to be able to host quite a large virtual world. Indeed, it is foreseeable that to be able to really use the simulation efficiently we need to build the scene in the simulation in an automated fashion by using the data from the input stage.

But with what method can we evaluate and optimize the solution generated by the AI in the simulation? As with a human conscience, this is the core of an AI conscience. How to ensure that the generated idea is beneficial and benevolent?

For this, we need to develop suitable metrics. In this case metrics refer to measurements, guidelines, and thresholds by which the AI processes and evaluates the solutions in the simulation, and they are intended to be comparable to human values, norms and moral. Henceforth let us call these metrics our evaluation criteria, so it is distinguishable from the performance metrics of an AI’s neural networks. These evaluation criteria should be as wide and deep as we can make them, so they consider as many factors as possible and thus be inclusive, fair, and expedient. So, let us define four broad categories to cover most factors of a solution:

1. Social
2. Ecological
3. Technological
4. Economical

These evaluation criteria cover two important perspectives we can take. First the differentiation between people as a private social person and people as a public professional person. This is a common split of our lives, that covers most if not all of it. The second perspective is between a localized view and a global view. Although there is overlapping, the social evaluation criteria reflect mostly the localized personal aspect, ecological the global social, technological the local professional and economical the global professional aspect. To develop these evaluation criteria, we will on one hand need to look at factual scientific knowledge and on the other hand at opinions and philosophical concepts.

It is with these evaluation criteria that we can ensure that the AI will process and hence act with “conscience”, as it ensures that the most relevant influence factors to humanity and environment are considered in the solution generation. Of course, we need to design these criteria very carefully, use various sources such as expert knowledge and public surveys, review and evaluate all sources and collected data, and design them so that the AI can use them in the intended way. This in how the most beneficial and benevolent solution for all involved parties can be found, and the respective best suiting output can be generated by the AI.

Create a solution that solves the identified situation the best way possible.

For different situations, different solution will work best and if we think about what kind of solutions humanity has produced so far, we can identify three different possible output types which are:

1. Explanation
2. Action
3. Technology

These three output types have a broad variety of potential and could solve many, if not all, challenges i.e. situations presented to humans, or in this case, the AI conscience. Let us look at each output type so we can understand what solutions the AI needs to be able to generate and what the AI conscience application areas could be.

Sometimes an explanation will be enough to solve a situation, meaning we can give the directly involved parties extra information about the situation and help them consider the needs, values and wishes of the other parties. In this case, the AI’s output will be the information it could access about involved parties and inferred implications according to the expansive knowledge base taught to the AI. The AI conscience will need to generate a solution, that is among other things, personalized relevant information for each party on how to collaborate or synergize with the others, while at the same time not violating the privacy of anybody.

Other times, the AI needs to generate a series of suggested actions. This will give us for example suggestions how to interact with others, what products or methods to use and what effects and implications different decision are likely to have. This time the AI must extensively explore the possible paths and evaluate their feasibility. The AI conscience needs to consider that the suggested actions do not harm people, the environment or maybe the economy, and are also sustainable, meaning they do not have short- or long-term disadvantages. But the suggested action also needs to be as efficient and effective as possible, to make them suitable alternatives to human reasoned actions.

Lastly, today we can think of a variety of situations and challenges we cannot efficiently or effectively solve by the means of knowing more or better actions. Thus, we need new concepts and technology that we can use as a solution. Especially with today’s challenges of climate change, pandemic, biodiversity loss, pollution and so on, we have a lot of opportunities, where new or advanced technologies can have a big impact. For this, the AI must access a massive knowledge base and be able to juggle with the existing technologies, but also and maybe even more importantly, the AI needs to be able to develop new concepts. The AI conscience must develop a variety of solution fit for different cultures and technological possibilities of multiple countries and needs to consider global impact and sustainability, people’s jobs and lives as well as the economic implications for local and international businesses.

By looking at the solution types we can identify what data the AI conscience has to take, process and output, to create viable solutions, but in addition we need to implement guidelines and abilities to make the AI output human evaluable. This means, among other things, we need to be able to understand and track the computation process the AI took to generate its output and we need to ensure, that we can inform users and involved parties about uncertainties and probabilistic approaches in the decision-making process. The AI Conscience needs to be able to collect and output that information in a fashion, that humans can quickly review the information.

Conclusion

Looking at the range of output and their possible impact, it becomes obvious that we need an AI conscience, or else the risk of such an AI exploiting or misdirecting our lives is much too high. The AI with a conscience will design every output in a safe simulation while adhering to the sophisticated, strict, and carefully developed evaluation criteria. This will give us safety, that the output will not hurt us or harm our environment and it will incorporate everything we need to ensure privacy and security for everybody.

There are many more thing to consider, think about and design, before we can safely create immensely powerful AIs that help us direct our lives, but I hope this introduction to AI conscience gives a first glimpse at how we can create a core part of AI technology to ensure we all have a bright future.

There will likely be another article that goes deeper into the topics of simulation, evaluation criteria, and ensuring privacy, security, and safety for an AI conscience, as well as more on how we can create a knowledge base suitable for teaching AI all of this.

Thank you for reading! Please leave a like and follow if you enjoyed the article and tell me about your thoughts on this topic in the comments.