Deep learning has driven the current rise of artificial intelligence and is the backbone of today's machine intelligence. Despite its successes, its limitations have only recently become apparent. This perspective argues that many of deep learning's problems stem from a common issue: shortcut learning. Shortcut learning refers to decision rules that perform well on standard benchmarks but fail in more challenging scenarios, such as real-world applications. This phenomenon is observed in various fields, including Comparative Psychology, Education, and Linguistics, suggesting it is a common characteristic of both biological and artificial learning systems. The paper discusses how shortcut learning can be identified and addressed. It introduces a taxonomy of decision rules, highlighting how shortcuts can be distinguished from intended solutions. The paper also explores the origins of shortcuts, emphasizing the role of dataset biases and discriminative feature learning. It discusses how shortcuts can be revealed through out-of-distribution (o.o.d.) generalisation tests and how they affect different areas of deep learning, including Computer Vision, Natural Language Processing, Reinforcement Learning, and Fairness. The paper emphasizes the importance of understanding shortcut learning to improve the robustness and transferability of deep learning models. It outlines actionable strategies for diagnosing and understanding shortcut learning, including careful interpretation of results, designing effective o.o.d. tests, and considering the inductive biases of models. The paper also highlights the need for research in areas such as domain adaptation, adversarial robustness, fairness, and meta-learning to overcome shortcut learning and improve the generalisation of deep learning models.Deep learning has driven the current rise of artificial intelligence and is the backbone of today's machine intelligence. Despite its successes, its limitations have only recently become apparent. This perspective argues that many of deep learning's problems stem from a common issue: shortcut learning. Shortcut learning refers to decision rules that perform well on standard benchmarks but fail in more challenging scenarios, such as real-world applications. This phenomenon is observed in various fields, including Comparative Psychology, Education, and Linguistics, suggesting it is a common characteristic of both biological and artificial learning systems. The paper discusses how shortcut learning can be identified and addressed. It introduces a taxonomy of decision rules, highlighting how shortcuts can be distinguished from intended solutions. The paper also explores the origins of shortcuts, emphasizing the role of dataset biases and discriminative feature learning. It discusses how shortcuts can be revealed through out-of-distribution (o.o.d.) generalisation tests and how they affect different areas of deep learning, including Computer Vision, Natural Language Processing, Reinforcement Learning, and Fairness. The paper emphasizes the importance of understanding shortcut learning to improve the robustness and transferability of deep learning models. It outlines actionable strategies for diagnosing and understanding shortcut learning, including careful interpretation of results, designing effective o.o.d. tests, and considering the inductive biases of models. The paper also highlights the need for research in areas such as domain adaptation, adversarial robustness, fairness, and meta-learning to overcome shortcut learning and improve the generalisation of deep learning models.