Artificial Intelligence: Bringing up baby

For awhile now, the world has eagerly explored and developed uses for artificial intelligence (AI) -- machines that can learn how to learn. In the interest of attracting investment, its pioneers bombard us with slick marketing about the amazing things it will be able to do in the near future. Actually, AI still has plenty of growing up to do. So far, though, according to Wikipedia, AI has been applied in... "

• Optical character recognition
• Handwriting recognition
• Speech recognition
• Face recognition
• Artificial creativity
• Computer vision,?virtual reality,?image processing,?Pixel-art scaling algorithms,?Image scaling?and?motion interpolation[89][90][91][92][93]
• Photo?and?video manipulation
• Diagnosis (artificial intelligence)
• Game theory?and?strategic planning
• Game artificial intelligence?and?computer game bot
• Natural language processing,?translation?and?chatterbots
• Nonlinear control?and?robotics

Other fields in which AI methods are implemented

• Artificial life
• Automated reasoning
• Automation
• Bio-inspired computing
• Concept mining
• Data mining
• Knowledge representation
• Semantic Web
• Email spam?filtering
• Robotics
• Behavior-based robotics
• Cognitive
• Cybernetics
• Developmental robotics?(Epigenetic)
• Evolutionary robotics
• Hybrid intelligent system
• Intelligent agent
• Intelligent control
• Litigation
• Oil and Gas[94][95]"

Some of these applications rely on machine learning, a type of AI, where inputs are selected and analyzed based solely on a priori algorithms programmed by humans. The machine "learns" as it gathers more data points and applies the prescribed rules, but it's not making complex decisions. Its decision trees are limited in depth by humans' capacity and stamina to scope them. The kind of AI deemed possible via, say, quantum computing, goes beyond learning gathered from the specific query and data presented. It applies computational capabilities and speed beyond those of classical computers to determine information needs, make judgments, and prioritize questions and answers to deliver optimal solutions, not just solutions.

As with all technologies from the wheel, to factory equipment, to the internet, a technology of any sort is only as useful and successful in use as the quality of its design and construction, as well as the knowledge of its users -- why to use it, what it can do, and how to use it. AI won't replace the human brain, but runs the risk of outrunning it and getting out of control if not nurtured with care.

Are we looking to AI to teach us well? I wouldn't take advice from a baby. Babies are born with a powerful system for learning about the world, but their success in it depends on what they're exposed to while "growing up". At first, babies aren't even able to tell us much about themselves, except for their most basic needs and feelings. AI might have a more powerful and faster processor than a human brain in some ways, but its success in delivering what the world needs and wants from it depends very much on its human parents and teachers for seven things:

1. What it's trying to accomplish, and how that information will be used. Why does it matter? What is the context of the goal? How important is the mission in this context, relative to other goals? What missions are forbidden (e.g., kill humans, enable human trafficking, etc.)? These are examples of situation-specific judgments, new algorithms that AI can "learn" to create, given enough "experience", but only then. The raw curiosity has to come from somewhere. How much and what kind of relevant real-world experience do the human "parents" bring to this? What assumptions and ethics are they imparting by their choice of inquiry? A learner is, after all, developing a set of "beliefs", assumptions on which they base further behavior.
2. The language(s) on which it's built and those in which it can communicate.
3. The information available to it. This includes its source, how it's gathered, how it's presented, when and where in the process it's presented, accessibility, volume, veracity, reliability, reproducibility, etc.
4. Rules (algorithms) for how to inquire and how to handle output. This might include choices about things like algorithm triggers; models for determining statistical limits at which to ascribe causality, co-occurrence, and outcome; rules determining the relative importance of both information and query based on the goal at hand (#1). Such rules might govern, for example, how to identify biases and how to address them.
5. The identification of relevant unanswered questions and missing information. This is the "learning" part, the point of a machine that learns how to learn. It will improve its skills in this area with each iteration, but like a human, it will mostly base its approach to learning and prediction on historical data and previous learning experience. These might include its own as well as those of its community-- other machines with which it interacts. It will continually rely on humans to provide information that doesn't already exist somewhere in the machine's accessible technosphere. How do we teach it to identify what else it or we need to know? How will it ask what we didn't think to ask?
6. How to resolve conflict. What happens when one machine's assumptions and rules conflict with those of another machine-- an information source guided by different parents and teachers? Rules are needed to address how those conflicts are resolved, because learning cannot continue while conflicts prevent it. So, healthy AI communities need a common language and shared ethics. See the parallels with the human world? Machines might even find their own solutions to conflict, ones less influenced by emotion, but nevertheless based on their programmed limitations and priorities.
7. How to stay safe. There will always be bad actors in the world of information technology, as elsewhere, and they too have learning machines. What protections are we able to establish and how do we teach our AI how to identify threats, avoid them, and eliminate them?

Ultimately, humans decide what shapes these AI babies. It matters very much which humans are involved, and we need to understand clearly who's supposed to be doing what. Are they individuals well-qualified to understand the reason for the machine's existence and what it's supposed to accomplish? Are they the architects following needed function with form? Are they the expert builders, or maintenance technicians, or the everyday operators tasked with fulfilling its purpose? In the current environment, too many expert builders, technicians, and operators without enough expertise in the purpose and use context at hand are being relied on to figure out the tools' needed content, functionality, and outputs. Inexperience, disagreement, inadequate communication, and/or lack of coordination in any of and among these distinct disciplines can lead to either serendipity or trouble -- usually the latter.

Just like a baby human, AI will be shaped by its inherent capabilities and by its influences. Raising these AI babies to be good citizens and valuable contributors will require us to look deep into our values -- our fundamental rules and assumptions. We'll need to get clear about our goals, the most productive ways to achieve them, and how to make the most of these tools. We'll need to agree on common language or languages to facilitate access to other systems and information sources. At a certain point, our AI babies will explore the world of information with less guidance from us. As they create their own algorithms and learn on their own, we risk losing control over who they interact with, and how. While they're young, we have the opportunity to imbue them with good core assumptions, to determine who else gets to teach them, and to guide them on how to identify good information on their own. Having invested so much, we'll do our best to protect them, and us, from harm.

Raised right, AI will mature into all that we hoped from it, and more than we imagined.