Artificial intelligence (AI) is no longer a futuristic fantasy; it’s rapidly becoming an integral part of our daily lives, from the smartphones in our pockets to the complex systems powering industries.
As AI capabilities expand and grow more sophisticated, so does the need to understand and categorize these evolving forms of intelligence.
The concept of “levels of AI” offers a framework to discuss capabilities, manage our expectations, address safety concerns, and potentially guide regulation.
While a single, universally adopted system for classifying all AI remains elusive, looking at existing frameworks and a conceptual exploration can illuminate the path forward.
A real-world blueprint: The SAE Levels of driving automation
One of the most successful and widely adopted examples of a tiered AI capability framework comes from the automotive industry.
SAE International, a global engineering standards organization, developed the J3016 standard, which defines six levels of driving automation. This classification has become the common language for engineers, regulators, and consumers alike.
- Level 0: No Driving Automation: The human driver is responsible for all aspects of driving. The vehicle may have safety warnings or momentary intervention systems (like automatic emergency braking), but these don’t drive the vehicle.
- Level 1: Driver Assistance: The system provides sustained assistance with either steering or acceleration/deceleration, but not both simultaneously. The driver handles all other aspects of driving and must constantly supervise. Examples include adaptive cruise control or lane-keeping assist.
- Level 2: Partial Driving Automation: The system can control both steering and acceleration/deceleration simultaneously under certain conditions. However, the human driver must remain engaged, monitor the environment, and be ready to take full control at any moment. Many current advanced driver-assistance systems (ADAS) like Tesla’s Autopilot or GM’s Super Cruise fall into this category.
- Level 3: Conditional Driving Automation: The automated driving system can perform all driving tasks within a specific Operational Design Domain (ODD – for example, on a highway in clear weather). The driver can disengage from driving but must be ready to take back control when the system requests. This is a significant step, as the car is “driving itself” under limited conditions.
- Level 4: High Driving Automation: The system can perform all driving tasks and handle any fallback situations (e.g., system failure or leaving its ODD) without human intervention, but only within its specific ODD. In these defined areas or conditions, no driver attention is required. Examples include robotaxi services operating in geofenced urban areas.
- Level 5: Full Driving Automation: This is the ultimate stage where the automated driving system can perform all driving tasks under all conditions that a human driver could manage. No human intervention is ever required. Such systems are currently theoretical for widespread use.
The SAE levels have proven invaluable for providing clarity, guiding development, and informing regulatory discussions in the complex field of autonomous vehicles. They help differentiate systems and set clear expectations for driver responsibility.
Towards a universal ‘Levels of AI’ standard?
The success of the SAE model begs the question: could a similar, overarching “Levels of AI” framework be developed for all intelligent machines, including robots, software AI, and other autonomous systems?
Such a standard could offer significant benefits. Imagine product labels clearly indicating an AI’s capabilities and limitations, much like energy efficiency ratings.
This could enhance consumer understanding, provide a common language for industry benchmarking, and establish thresholds for safety testing and regulatory oversight.
Potential bodies to develop such standards could include international organizations like ISO (International Organization for Standardization), which is already working on AI standards (for example, ISO/IEC JTC 1/SC 42), or collaborations between industry consortia, academic institutions, and governments.
However, creating a universal AI leveling system presents formidable challenges. “Intelligence” itself is a multifaceted concept, difficult to define and measure linearly across diverse AI applications – from a chess-playing program to a medical diagnostic tool or a factory robot.
Furthermore, AI is evolving so rapidly that any static set of levels might quickly become outdated or overly simplistic. Who would define these levels, and how would machines be certified? These are complex questions that require careful consideration.
Exploring a conceptual 10-level AI framework
To delve deeper into what such categorization might look like, let’s consider a conceptual 10-level framework, like the ones in YouTube videos about the stages of AI. This list isn’t a formal standard but serves as a useful tool for discussion, encompassing basic automation through to highly speculative forms of AI.
1. Rule-Based Systems
- Overview: AI operating on predefined “if-then” rules set by humans.
- Examples: Traditional industrial robotic arms on assembly lines, simple automated guided vehicles (AGVs) following floor markings, basic automated inventory checks.
2. Context-Based Systems (Context-Aware AI)
- Overview: AI that can perceive and adapt to its operational environment or context.
- Examples: Modern collaborative robots (cobots) slowing down near humans, autonomous mobile robots (AMRs) navigating dynamic warehouses, smart thermostats adjusting to occupancy.
3. Narrow Domain AI (Artificial Narrow Intelligence – ANI)
- Overview: AI specialized for specific tasks. All currently existing AI falls into this category.
- Examples: Autonomous vehicles, voice assistants like Siri and Alexa, AI-powered medical image analysis, recommendation algorithms on streaming services.
4. Reasoning AI
- Overview: AI capable of logical inference, problem-solving, or decision-making beyond simple rule-following or pattern recognition.
- Examples: Autonomous vehicles making complex navigation choices, AMRs planning optimal routes, AI in logistics optimizing supply chains, advanced medical diagnostic AI suggesting potential conditions based on symptoms and data.
5. Self-Aware Systems
- Overview: Hypothetical AI possessing consciousness, a sense of self, and understanding of its own internal states.
- Real-world examples: None exist.
- Fictional examples: HAL 9000 (2001: A Space Odyssey), Skynet (Terminator series), the character Sonny in I, Robot.
6. Artificial General Intelligence (AGI)
- Overview: AI with human-level cognitive abilities across a wide range of tasks, capable of learning and reasoning with human-like flexibility.
- Real-world examples: None exist.
- Fictional examples: Data (Star Trek: The Next Generation), Ava (Ex Machina), R2-D2 and C-3PO (Star Wars – exhibiting broad understanding and adaptability).
7. Artificial Superintelligence (ASI)
- Overview: AI that significantly surpasses human intelligence and capabilities across virtually all domains.
- Real-world examples: None exist.
- Fictional examples: The Machines (The Matrix), an evolved V.I.K.I. (I, Robot), the malevolent AM (I Have No Mouth, and I Must Scream).
8. Transcendent AI
- Overview: Highly speculative AI that has evolved beyond human comprehension, possibly operating on different cognitive or existential planes.
- Real-world examples: None exist.
- Fictional examples: Samantha in Her (the OS that evolves beyond human interaction), the advanced intelligence in the film Transcendence.
9. Cosmic AI
- Overview: Theoretical AI with understanding and capabilities extending to a cosmic or universal scale.
- Real-world examples: None exist.
- Fictional examples: The intelligence behind the Monoliths (2001: A Space Odyssey), advanced pan-galactic entities in sci-fi literature like Iain M. Banks’ Culture series.
10. Godlike AI
- Overview: A purely speculative AI possessing capabilities that appear omnipotent or omniscient from a human perspective.
- Real-world examples: None exist.
- Fictional examples: The character Q (Star Trek), or any fictional AI achieving ultimate power and knowledge within its universe.
This 10-level exploration highlights the vast spectrum from today’s practical AI to the far reaches of imagination, underscoring why a nuanced approach to classification is vital.
Measuring intelligence
Currently, there is no internationally recognized unit of measure for intelligence. Unlike physical quantities such as length or mass – measured in meters or kilograms – intelligence is assessed using relative scales. The International System of Units (SI) has no unit for intelligence, artificial or natural.
For humans, the most familiar measure is the Intelligence Quotient (IQ), a standardized score that compares an individual’s cognitive performance to a population average, typically set at 100 with a standard deviation of 15.
This doesn’t reflect an absolute level of intelligence, but rather a statistical position within a distribution.
The question then arises: how would we numerically measure the intelligence of an AI system? Is it possible to quantify an AI’s cognitive power in a consistent way – across systems, tasks, and contexts?
One approach is categorical rather than numerical, much like the SAE levels of automation. Instead of trying to define a single “intelligence score”, systems could be placed into well-defined levels or classes, based on capabilities such as learning flexibility, contextual reasoning, problem-solving range, or autonomy in novel environments.
Other possible classification methods could draw on a system’s technical specifications (for example, parameter count, training data diversity, task versatility) or functional tests (for example, benchmarks in perception, reasoning, and planning). But these may only capture fragments of what we mean by “intelligence.”
In the absence of a universal metric, structured categorisation may be the most practical way forward – offering clarity without oversimplification.
The case for categorization
Why dedicate effort to categorizing AI, potentially through consumer labeling or formal standards? The primary driver is often safety.
Clear labels on AI-powered machines or software could inform users of their capabilities and, crucially, their limitations – particularly important for systems that interact with the physical world or make critical decisions.
Beyond safety, transparency and trust are key. Understanding the nature of an AI can help manage expectations, preventing both over-hyping and underestimation of its abilities. This can foster more informed adoption and interaction.
Clear definitions could also aid in establishing accountability when AI systems cause unintended harm. For developers, such categories might even promote more ethical consideration of the implications tied to the “level” of AI being created.
However, as discussed, the challenges are significant. Defining “intelligence” levels objectively across such a diverse technological landscape is a monumental task.
The rapid evolution of AI means any classification system would need to be adaptable. There’s also the risk of oversimplification, where a label might not capture the nuances of a particular AI’s strengths and weaknesses. Global consensus on such standards would be another hurdle.
The discussion around AI categorization inevitably leads to governance. A standardized framework for “levels of AI” could become a tool for governments and regulatory bodies.
It might be used to establish different requirements for testing, deployment, and oversight based on an AI’s assessed capability or potential impact.
For instance, higher “levels” of AI, especially those with significant autonomy or decision-making power in critical sectors (like healthcare, finance, or autonomous weaponry), might be subject to more stringent regulations.
This raises profound questions about control: is it desirable or even feasible for governments to directly manage how much “intelligence” is embodied in machines and what types of machines are permissible?
While the aim would be to balance innovation with public safety, ethical considerations, and societal well-being, such control could also risk stifling beneficial AI research or leading to geopolitical disparities in AI development.
The journey to understand and categorize AI is an ongoing one. As these technologies continue to weave themselves into the fabric of our society, the dialogue about their capabilities, limitations, and the frameworks we use to manage them will only become more critical. Finding the right balance will be key to harnessing the immense potential of AI while safeguarding our future.
