Much like autonomous driving, there is a misunderstanding outside of the AI community about the true, current capabilities of generative AI and the models that are emerging. At Engenai we spend a fair bit of our time working with clients to help them understand the importance of the quality and quantity of their data in determining the relevance of the AI’s output, as well as how much human curation is still needed to review genAI content before publishing.
Let’s begin a dialogue in the AI community on how we classify generative AI autonomy into different levels or stages. These would describe the extent to which the AI system can operate independently and generate content without human intervention. These levels could be used as a framework to help understand the capabilities and limitations of AI systems, and generate a shared vocabulary. The levels of generative AI autonomy can vary depending on the context and application, but let’s begin our framework discussion with the following:
Level 0 – No Autonomy
The AI is completely controlled by humans and cannot generate any content on its own. It generates content or responses only when explicitly prompted or directed by a human operator. There is no initiative or creativity on the part of the AI.
Level 1 – Assisted Autonomy
The AI can generate content, but it is still heavily guided by humans. The AI might be able to generate text, but it would need to be prompted with a specific topic or style. The AI system can assist a human user in generating content by providing recommendations, correcting errors, or suggesting improvements based on predefined rules or patterns in the data. The human user still maintains control and makes final decisions.
Level 2 – Partial Autonomy
The AI can generate content with some degree of independence within certain predefined boundaries or guidelines. They can understand context and generate coherent responses or content without constant human supervision. Additional human intervention may be required for complex or exceptional cases. For example, the AI might be able to generate images, but it still needs human oversight to ensure that the content is safe and appropriate.
Level 3 – Conditional Autonomy
The AI can generate content with a high degree of autonomy, but it is still subject to certain conditions set by humans. Humans may still need to set high-level goals and provide occasional guidance. The AI can adapt to changing contexts and generate content or responses without constant human oversight. For example, the AI might be able to generate text that is factually accurate, but it would need to be programmed to avoid generating text that is biased or discriminatory.
Level 4 – High Autonomy
The AI can generate content with a high degree of autonomy and is capable of operating independently in most situations. The AI is not subject to any significant conditions set by humans, only requiring human intervention for rare, exceptional cases. They can understand and respond to a wide range of inputs, adapt to novel situations, and make complex decisions. However, the AI is still monitored by humans to ensure ethical or legal compliance.
Level 5 – Full Autonomy
The AI can operate entirely autonomously, making decisions and generating content without human involvement or oversight. It exhibits a high level of creativity, adaptability, and problem-solving abilities. The AI models are self-training and no longer need human input to make judgements on the content created. Additional information can be drawn from connected, external sources. However, this level of autonomy is still considered theoretical and may not be achievable in the near future. Full autonomy raises significant ethical, legal, and safety concerns and may not be achievable or desirable in all contexts.
Of course, the level of autonomy in generative AI can vary widely depending on the specific AI model, its training data, and the design choices made by developers. Additionally, ethical considerations, safety measures, and regulatory guidelines play a crucial role in determining the appropriate level of autonomy for a given application.
—
For this article, we used Level 1 generative AI, assisted autonomy. It suggested the initial levels of genAI autonomy and we wrote the article using those initial talking points. Generative AI, returned content that is helping define its own levels of autonomy. Imagine what it will do at Level 4.
