What things will AI never be good at?
- Never is unlikely, but AI today is not better than humans with language and social tasks. AI's are also not as of 2022 able to recognize 'expression of genius'.
It is difficult to make predictions about technology that has yet to be invented. Moore's law forecasts continued increases in CPU capabilities, but it is unclear when or if this trend will hit physical limits.
Criteria for being "good" at something is unclear, but AI systems are often evaluated against human performance at the same task. There is no known limitation of physics or information theory that would prevent construction of an artificial brain with similar or better capabilities than the human brain, given sufficiently sophisticated manufacturing techniques. It is more feasible to predict which things AI systems are likely to get good at sooner vs. later, or to simply observe what they are already good at.
The human brain is massively parallel; with the nearly universal Von Neumann architecture, CPU cores typically execute one instruction at a time, albeit extremely quickly. The algorithms that human brains use to solve problems are only partly understood (but are an active area of research in psychology), and emerge from a combination of evolved hardware and skills learned over many years. AI systems which attempt to solve problems in the same way that humans do could be expected to have similar strengths and weaknesses, though the ability to change the hardware on which these algorithms run suggest artificial versions could improve upon human capabilities. Given the lack of complete understanding of human psychology and problem-solving techniques yet to be invented, it is unclear how many alternatives to human problem-solving techniques exist, and whether after being perfected they will be better or worse.
Adult humans have sophisticated mental facilities that handle body movement, object recognition and tracking (useful for hunting and to avoid predators and moving objects), face recognition, emotions, social relationships, language, reasoning, and creative tasks. Human reasoning is often imprecise. Human memories frequently make errors, they frequently forget things, and they tend to make mistakes or become disinterested in tasks that are repetitive and boring.
AI is already far superior to humans in things involved mathematical calculation and information storage
Computers are already far superior to humans at accurate information storage and can perform straightforward mathematical calculations extremely quickly. This enables them to perform simulations of physical phenomena that would be impractical for a human to perform. Computer-generated imagery can produce convincingly realistic pictures and video, but human creativity is generally used to decide what to depict.
Different AI technologies have different strengths and weaknesses. In the 21st century, artificial neural networks have been successfully used to solve many tasks. These systems must generally be trained with a large amount of data, and are somewhat flexible in handling situations they have not directly encountered. Given that they use complicated numerical transformations to map inputs to outputs, it is often difficult to explain why a neural network behaves in a certain way, or to repair it other than by using different or supplementary training data. Rule-based AI systems are better at explaining why they reach certain conclusions, and are more predictable but potentially more fragile when encountering novel inputs.
AI capabilities are being actively developed
AlphaGo defeated humans at the game of go by using a neural network, whereas systems that win at checkers or chess typically do so by using sheer computing power to examine huge numbers of forward-looking scenarios. Watson defeated human champions on the game show Jeopardy! using a wide variety of techniques.
Robots have demonstrated the ability to navigate around rocks on Mars and vacuum a room while on wheels. Autonomous automobiles, boats, aircraft, and sidewalk robots are under development, and it is assumed that eventually they will surpass human capabilities and cause far fewer accidents.
Natural language processing systems are primitive compared to human understanding and speech capabilities, but are improving. They are used to power (in limited ways) customer service phone systems, search engines, smart speakers, and to assist in certain tedious processes like legal document discovery. In terms of creative writing, they can produce formulaic news stories, recipes based on ingredients and training from existing human recipes, generate eclectic poetry, and make small talk (though not yet entirely convincingly). NLP systems currently perform poorly on tasks that require a deep semantic understanding of language or concepts in a general domain.
AI recognition and expression of emotions is a nascent technology, but the field of affective computing is an area of active research and development. Though humans often personify (and assign a gender to) computers - especially those that can speak - AI systems currently perform poorly at social tasks, such as negotiations and non-transactional social relationships.