What is AI?
AI refers to the use of computer systems to perform tasks that traditionally require human intelligence and senses. Instead of a programmer assigning AI a set of step-by-step instructions, AI “learns” by using statistical techniques that allow it to improve performance on a task. This process of machine learning equips AI to generate rules and predictions on its own by analyzing large quantities of data.
How Does AI work?
Most AI technologies currently in use are considered “Artificial Narrow Intelligence” because they can only perform a specific task or a sequence of tasks autonomously by using human-like capabilities. AI technologies being researched today are moving toward “Artificial General Intelligence,” where AI will be able to apply its capabilities to perform many different tasks.
As AI capabilities increase, organizations must be mindful of the growing potential for socially unacceptable outcomes. To mitigate risks, organizations like Deloitte recommend embedding human values in the loop. This involves using a risk-based approach that considers principles like impact (beneficence, non-maleficence), justice (procedural fairness, distributive fairness), and autonomy (comprehension, control).
Artificial Intelligence can include some or all of the following features:
- Natural Language Processing (NLP) enables AI to communicate with humans. AI can read, interpret, and produce linguistic data. For example, AI can interpret a speaker’s voice and create a transcript based on the results.
- Knowledge Representation acts as memory for AI. Humans program machines to store and organize knowledge. The resulting model it creates from that knowledge can be used to solve complex problems. For example, an AI “seeing” application may learn to identify different types of currency once it has built enough knowledge.
- Automated Reasoning is the theoretical framework behind machine learning models. This framework gives machine learning models a structure to define, approach, and solve problems. The goal is to enable AI to use stored information to answer questions and draw conclusions automatically. AI that incorporates automated reasoning can create rules to understand different situations, so it can identify when something unexpected is happening, such as a data breach. As another example, a company using AI with automated reasoning capabilities can identify unauthorized purchases on a customer’s account based on that customer’s typical location and spending habits.
- Machine Learning Models are a key type of AI that learns to recognize patterns in datasets. Humans train a machine learning model by coding an algorithm to analyze and draw conclusions from an existing dataset. For example, a model can learn to recognize patterns in an image dataset, grouping similar images into labeled categories. Machine learning is a primary technique that uses automated reasoning.
- Datasets are the building blocks for machine learning models. Humans often separate a dataset into two or three segments for training, testing, and validation. The model practices on training and testing datasets, then runs on a validation dataset. Results from the validation dataset determine how well the model performs and whether it is fit to analyze new data out in the world. Common indicators of a model’s fitness are accuracy, precision, and recall.
- Algorithms are sequences of “if, then” statements, which humans design as part of Machine Learning models. These models input datasets, then run algorithms to detect patterns, adapt to new situations, and make decisions automatically.
What Does AI Do Well?
AI can process a lot of information by analyzing and finding patterns in large datasets that may be difficult or impossible for a human to identify. It can automate simple and repetitive tasks and make complex tasks more efficient. AI is a wonderful chess player, for example, because it can consider the total number of possible moves in a predictable, controlled environment at a pace and scale that a human cannot.
Artificial Intelligence technologies can extend human capabilities in several ways. The tasks they can perform may include one or more of the following:
- Simulating Human Intelligence and Sensing: AI can be designed to mimic the way humans sense, process, and interpret information. For example, an AI-enabled mobile app can interpret camera images and identify text, objects, and people for an employee with low vision. Another application might use AI to translate speech to text and interpret sign language to help both hearing and deaf or hard of hearing employees communicate during meetings.
- Automating Repetitive Tasks: AI can sift through granular data and analyze results at a fast pace. Humans can program AI to perform repetitive, high-volume tasks such as reviewing large batches of employee expense reports for errors and updating documents in large sets.
- Learning from Experience: Machine learning algorithms enable AI tools to improve automatically through ongoing analysis of new inputs (data) and results (more data), which in turn influences future predictions and results. AI tools can also learn to process images, sounds, and linguistic expressions. For example, an AI application might be trained how to distinguish different animal species一for example, dogs and cats一and then over time be able to identify different breeds within species.
Where Does AI Fail?
AI doesn’t always understand the context behind an assigned task or a problem a user is trying to solve. It identifies solutions through methodical guesswork and takes the path of least resistance. It cannot think critically like a human. Instead, AI acts based on how a human has programmed it. When we ask AI to complete complex tasks, AI’s inability to think like a human can lead to problematic results. For example:
- AI cannot raise moral objections the way a human can. For example, when analyzing candidates for job fit, AI may not understand some correlations are a result of structural inequities in an organization or in our society, rather than merit. AI may conclude that people with disabilities aren’t good fits for certain jobs simply because the datasets the technology uses in decision-making don’t include data specific to the success of people with disabilities. The AI would not have the capability to identify that the data was discriminatory and raise objections against using it. Because AI loves shortcuts, it may slyly find a way to use this conclusion in its findings, even if obvious demographic data is removed.
- Some problems are too hard for AI. AI doesn’t do well with nuanced decisions or with understanding the meaning of jokes, cultural references, and linguistic subtleties. To perform a task well, AI needs a wide range of skills and a large amount of accurate data to work with. Without having certain skills and accurate information it may struggle with tasks like employee performance reviews—something even humans find very difficult. This is especially true when there isn’t accurate data available for AI to learn from.
- AI is a confident bluffer. When we ask AI to complete complex tasks, it doesn’t know how to say “I don’t know” or ask follow-up questions. Instead of communicating that it doesn’t understand or needs more information to effectively complete an action, it will give its best guess in definitive terms. The answer may appear conclusive, even if the AI is only 51 percent sure. Assessing the validity of an AI model’s output is a complex task.
AI often makes mistakes for the following reasons:
- The problem it is trying to solve is too broad.
- It does not have enough good data to base automated decisions on.
- It was programmed using an artificial, simulated model that doesn’t match the real-world environment that it is actually working in.
