What is AI?

Defining AI is challenging and there is no single accepted definition. The technologies included within the term are constantly shifting and expanding as the science of AI advances while many older technologies once included are no longer considered AI at all. 

Even the most advanced AI today is narrow AI: tools focused on specific tasks such as pattern recognition, classification, task optimization, and anomaly detection. Experts disagree on when or even if general AI, which can perform any cognitive task as well as or better than a human, will ever be developed. General AI systems are therefore outside the scope of this Strategy. Augmented Intelligence is also a subset of AI in which AI and ML technologies, such as virtual assistants, which will assist humans by analyzing queries and providing relatable data to assist the requestor in making better decisions.

ML is currently the dominant technique in AI, both in terms of widespread application and effectiveness. Rather than defining rules to obtain a result as conventional software does, ML uses past data to identify patterns which allow it to optimize for a predefined goal. When functioning properly, ML tools can help anticipate future needs, events, trends, and risks, allowing users to yield significant efficiencies in areas such as maintenance, logistics and inventory management. However, the effectiveness of ML depends on access to sufficient, relevant and high-quality data, without which the tool’s outputs will be unreliable. As a result, the quantity and quality of the data used for ML applications is a key priority for DND/CAF.

Generative AI is a subset of ML which can produce a wide variety of novel content, such as images, videos, audio, text, code, and 3D models, in response to user prompts. It does this by identifying the structure and patterns of vast quantities of existing data, and then using these patterns to generate novel outputs with similar characteristics. Generative AI outputs can be complex, highly realistic, and at times, indistinguishable from human-authored content. Recent breakthroughs in the field, particularly in large language models and image generation, have significantly advanced the capabilities of generative AI, opening new possibilities to use the technology to solve complex problems, including assisting in scientific research.

In popular culture, AI is often characterized as a competitor to human intelligence; however in practice, the two types of intelligence are highly complementary. In many domains, human intelligence supported by AI can deliver results that are superior to those achieved separately. For example, AI can undertake tasks that are repetitive or require high levels of precision and sustained attention. This leaves human personnel free to undertake the tasks at which they excel, especially those requiring judgement, creativity, initiative, and a grasp of the strategic landscape.

Along with its promise, AI and its use of data brings about challenges and responsibilities. If an AI model is trained on biased data, the resulting prediction may reflect and perpetuate those biases, leading to real world harms. Further, AI predictions based on past data may not represent the future, failing to foresee low-probability but high-impact events. Finally, AI's data usage can also pose privacy and security risks, especially when it involves personally identifiable or otherwise sensitive government information. Consequently, the outputs of AI must always be analyzed against expert human judgement and the constraints and expectations of the organization, with an awareness of AI’s limitations. On the road to AI enablement, it is critical for the Defence Team to establish and uphold ethical, equity, and security requirements for the use of AI.