AI does not know what it does not know. It is not sentient. The brittleness of what it knows is reflected in the shallowness of what it learns. Accordingly, it is unable to identify and avoid blunders that are vivid to the human.
Its inability to notice glaring errors on its own highlights the need for humans to understand the limits of AI’s capacities, and to allow for human capabilities to review the outputs of AI, and even predict when an AI is most likely to fail.
The robustness of AI auditing and compliance at this time remains rudimentary. It is quite often the case that what and how it conflates is unexpected. It is here that many opportunities lurk, but also many dangers. It is here the need for auditing by humans becomes necessary.
The inability of AI to clear errors on its own underscores the need to develop testing that allows humans to (1) identify the limits of an AI’s capacity, (2) use human capability to review courses of action, and (3) to forecast when an AI is most likely to fail.
This is important since machine learning will drive the development of AI for some time into the future, and humans will remain deeply unaware of what an AI is learning, and how it knows what it has mastered.
Misidentification of one thing for another by AI is commonplace. AIs may confuse or conflate two objects that humans would easily distinguish. AIs are easily misled or confused by images of two objects because the images share a set of subtle characteristics. But subtle features also puzzle humans. A child may identify a sheep as a dog. Some humans simply ignore some differences.
AI learning and human learning are equally opaque. Painters like Amoako Boafo, architects like Frank Gehry, parents, and auto mechanics often act on the basis of intuition or “common sense”, and are thus unable to clearly articulate what or how they learned.
To overcome this opacity, we have developed in all cultures, professional societies, civil service academies, law and medical degrees, microcredentials, and professional certificate programs.
GraphCast is a new AI tool created by Google DeepMind that presently outperforms the European Medium Range Weather Forecasting model. The model generates a forecast in under one minute. It uses less computing power than traditional forecasting methods.
Numerical Weather prediction models take hours to produce a forecast. This data is fed to a supercomputer that executes millions of computations per second using global data from weather stations, satellites, balloons, and buoys. AI shortcuts this process because it does not try to model how the world works, and it does not try to solve complex equations.
GraphCast uses machine learning that takes on board decades of historical data. It is not as detailed as a forecast generated by the European Centre for Medium-Range Weather Forecasts. Yet it is better able to predict extreme temperatures and track storm paths. It predicted where Hurricane Lee would make landfall.
While AI models are making great strides, they will not immediately replace traditional approaches to weather forecasting, since AI models are trained on the data generated by traditional methods.
The Climate Crisis will limit the predictive power of AI-driven tools as the way weather systems evolve has changed. Hurricane Otis intensified into a category 5 hurricane in twenty-four hours.
This provokes the question – will AI systems be able to take into account atmospheric rivers, lightning sprites, mammatus clouds, melting African glaciers, and straight-line bands of high-speed thunderstorms or derechos? AI systems have never been trained on these weather structures.
AIs vary with the tasks they perform, and this impacts the techniques developers must use to create them. Different aims and functions require different training techniques.
At this time, three forms of machine learning are noteworthy: supervised learning, unsupervised learning, and reinforcement learning. Supervised learning created the AI that discovered Halicin. The MIT team used a database of 2000 molecules to train a model in which molecular structure was the input and antibiotic efficacy was the output.
This is supervised learning as the AI developers used a dataset containing molecular structures that were individually labelled according to the desired result. This is the same technique used to train AIs to recognize images.
In instances where developers are awash with lakes of data, unsupervised learning can be used to extract potentially useful insights, spot inconsistencies among reams of transactions, and notice patterns or anomalies without having any information pertaining to outcomes. Such AIs produce startling insights as much as they can produce preposterous results.
Netflix may use an algorithm to make clusters of customers with similar viewing choices. However, refining the algorithm is knotty because a particular person will have multiple interests.
Beyond data analysis, AIs have also been trained to operate in dynamic environments. The AI is not passive in reinforcement environments. The AI acts as an agent where a single decision produces a cascade of opportunities and risks.
Here feedback is the task of the reward function. This allows the AI to train itself at speeds of processors that make direct human feedback totally impractical.