Types of artificial intelligence: ANI, AGI, ASI and other classifications

Artificial intelligence can be classified in different ways. ANI, AGI, and ASI describe capacity levels, but they do not explain on their own how a system learns, what kind of result it produces, or how much it can act without human intervention.

An AI can also be predictive, generative, conversational, rule-based, trained through supervised learning, or function as an agent capable of using tools.

Therefore, when we ask how many types of artificial intelligence exist, the answer depends on the criteria used.

In this guide you will learn about the main classifications, their differences and which ones correspond to existing technologies or concepts that remain hypothetical.

For an overview of how it works, applications, advantages and risks, you can consult our complete guide to artificial intelligence.

How artificial intelligence is classified

There is no single classification capable of describing all AI systems.

We can group them according to:

• Their ability to work in one or more domains.
• The technology or approach used.
• The way they learn.
• The type of result they produce.
• The information format they process.
• Their ability to act autonomously.
• The function for which they have been designed.

The same system can belong to several categories.

For example, an assistant can be:

• Specialized artificial intelligence.
• Model based on Deep Learning.
• Generative AI.
• Conversational AI.
• Multimodal system.
• Agent with access to tools.

Each term describes a different aspect.

Types of AI according to their capabilities

The ANI, AGI, and ASI classifications compare the breadth of an AI's capabilities with the flexibility of human intelligence.

It is a useful conceptual framework, but its boundaries do not have universally accepted evidence.

ANI or narrow artificial intelligence

ANI means Artificial Narrow Intelligence, translated as narrow, specialized or limited artificial intelligence.

Describe systems designed to perform specific tasks within specific contexts.

Among its possible functions we find:

• Classify images.
• Recommend products.
• Translate texts.
• Detect anomalous operations.
• Generate content.
• Calculate routes.
• Recognize voice.
• Predict a demand.

The word “narrow” does not necessarily mean that the system can only execute one action.

A multimodal model can work with text, images, documents, audio, programming, and tools. However, it still depends on a defined architecture, data, instructions, and environment.

Characteristics of the ANI

• It is designed or trained for specific objectives.
• Their performance varies depending on the task.
• It does not automatically possess all human abilities.
• It can outperform specialists in specific activities.
• It does not generalize unlimitedly to any situation.
• Needs evaluation and supervision.

Under this classification, currently deployed systems are usually included within specialized AI, although some are very versatile.

AGI or artificial general intelligence

AGI means Artificial General Intelligence or artificial general intelligence.

It is used to describe a system capable of learning, reasoning, and applying knowledge in a wide variety of tasks and domains, including new situations.

A potential AGI should showcase capabilities such as:

• Learn new tasks with little information.
• Transfer knowledge between domains.
• Adapt to unfamiliar situations.
• Plan complex objectives.
• Solve various problems.
• Use tools.
• Correct errors.
• Maintain strong performance in many contexts.

There is no universal definition or test to certify that a system is AGI.

Therefore, different researchers, companies and organizations may use the term differently.

Some focus on achieving human-comparable performance in many tests. Others require adaptation, memory, autonomy, continuous learning, or the ability to function in real environments.

AGI does not necessarily mean awareness

General intelligence and consciousness are not equivalent concepts.

A system could show a great ability to solve tasks without demonstrating that it possesses emotions, subjective experience, or consciousness.

Therefore, a table on ANI and AGI should not use awareness as a primary technical requirement.

Is there currently an AGI?

There is no widely accepted consensus that any system has achieved artificial general intelligence.

Current models can offer very broad capabilities, but they continue to show irregular profiles: they excel at certain tasks and fail at others that a person might consider simple.

That is why it is more accurate to evaluate concrete capabilities than to absolutely state whether or not a tool is “intelligent like a human”.

ASI or artificial superintelligence

ASI means Artificial Superintelligence or artificial superintelligence.

Describe a hypothetical scenario in which a system would far exceed human capabilities in a wide variety of disciplines.

It could include superior performance in:

• Scientific research.
• Engineering.
• Planning.
• Math.
• Technological development.
• Strategy.
• Problem solving.
• Knowledge creation.

ASI does not represent a currently available technology.

There is also no established way to know how it would develop, what capabilities it would have, or whether it would arise as a consequence of an AGI.

His study is primarily related to debates about:

• Security.
• Control.
• Objective alignment.
• Concentration of power.
• Governance.
• Systemic risks.

It is important to differentiate these debates from the real limitations and risks of the systems already in use.

Comparison between ANI, AGI and ASI

The table should not be interpreted as an automatic scale by which every system has to move forward.

ANI, AGI and ASI are conceptual categories based on the breadth of capabilities.

Types of AI according to their technical approach

We can also classify artificial intelligence according to the technology used to produce its results.

AI based on rules and knowledge

These systems use rules, logical representations, or knowledge introduced by specialists.

A simple example would be:

If conditions A and B are met, he recommends option C.

They can be used in:

• Expert systems.
• Rule-based diagnostics.
• Configurators.
• Decision engines.
• Validations.
• Planning.

Its advantages are traceability and control.

However, maintaining thousands of rules can be difficult when the problem changes or contains too many exceptions.

Machine Learning

Machine Learning allows you to build models that learn relationships from data.

Instead of manually writing all the rules, examples are used to adjust a model.

It can be applied to:

• Classification.
• Regression.
• Anomaly detection.
• Segmentation.
• Recommendation.
• Language processing.
• Artificial vision.

Deep Learning

Deep Learning is a subfield of Machine Learning based on multi-layered neural networks.

It is frequently used in tasks related to:

• Language.
• Images.
• Audio.
• Video.
• Multimodal models.
• Content generation.
• Recognition of complex patterns.

It does not replace all Machine Learning models. For certain problems, a simpler technique can be faster, more interpretable, and more efficient.

The relationship between these concepts is explained in the guide on how artificial intelligence works.

Types of AI according to their learning method

Within Machine Learning we can differentiate several methods.

Supervised learning

The model learns using examples that include a correct answer or label.

For example:

• Emails classified as spam or legitimate.
• Tagged images.
• Operations marked as fraudulent.
• Homes accompanied by their price.

It is mainly used for:

• Classification.
• Regression.
• Prediction.

Unsupervised learning

The model receives data without previously labeled responses and searches for structures or groupings.

It can be used for:

• Segment customers.
• Group documents.
• Detect patterns.
• Explore data sets.
• Reduce dimensions.
• Identify atypical behaviors.

The system finds relationships, but a person must interpret what the groups obtained mean.

Reinforcement learning

An agent learns by interacting with an environment.

Each action can produce a reward or penalty. The goal is to find a strategy that maximizes the accumulated reward.

It is used in:

• Robotics.
• Games.
• Systems control.
• Planning.
• Optimization.
• Research on agents.

Self-supervised learning

Self-supervised learning allows you to create training signals from your own data without having to manually label each example.

For example, a model can learn by trying to:

• Predicting a hidden part of a text.
• Reconstruct an incomplete image.
• Relate content representations.
• Identify the next part of a sequence.

This approach has played an important role in the development of large language models and foundational models.

Types of AI according to the result they produce

Another particularly useful classification differentiates systems based on the output they generate.

Predictive AI

Predictive AI calculates what outcome might be produced from certain data.

It can generate:

• A probability.
• A numerical value.
• A forecast.
• A risk estimate.
• An expected demand.
• An alert.

Examples:

• Sales prediction.
• Delivery time estimate.
• Risk of non-payment.
• Breakdown forecast.
• Probability of customer abandonment.

Their results can usually be compared later with what actually happened.

Generative AI

Generative AI produces new content from an input.

It can generate:

• Text.
• Images.
• Audio.
• Music.
• Video.
• Code.
• Presentations.
• Three-dimensional models.

It should not be confused with AGI.

A generative system can create content in various formats and continue to be a specialized artificial intelligence.

Check out our guide on what is generative artificial intelligence to learn about its models, operation and limitations.

Discriminative or classifying AI

Discriminative models separate examples, calculate categories, or estimate relationships between an input and an output.

They can be used for:

• Identify objects.
• Classify emails.
• Detecting diseases in images.
• Recognize intentions.
• Differentiate between normal and anomalous operations.

“Generative” and “discriminative” describe different approaches, although the same product can combine both.

Recommendation systems

They analyze information about users, products or content to sort options that could be relevant.

They are used in:

• Electronic commerce.
• Audiovisual platforms.
• Music.
• Social networks.
• Advertising.
• News.
• Training.

They can combine collaborative filtering, predictive models, content, and context.

Conversational AI

Conversational AI processes language to maintain interactions via text or voice.

May include:

• Chatbots.
• Assistants.
• Care systems.
• Voice interfaces.
• Agents with tools.

Not all conversational AI is generative. Some systems follow predetermined flows, intentions, and responses.

Artificial vision

Machine vision analyzes images or videos.

You can perform tasks such as:

• Image classification.
• Object detection.
• Segmentation.
• Optical character recognition.
• Follow-up.
• Quality inspection.
• Scene interpretation.

Robotics and autonomous systems

These systems combine perception, planning and action within a physical environment.

They can incorporate:

• Cameras.
• Sensors.
• Predictive models.
• Location.
• Route planning.
• Motion control.
• Reinforcement learning.

Its autonomy depends on the design, restrictions and level of supervision.

The guide on applications of artificial intelligence It explains in which sectors these categories are used.

Types of AI according to their level of autonomy

Autonomy describes the extent to which the system can select and execute actions without a person confirming each step.

Assistance tools and co-pilots

They help a person complete a task, but the user maintains primary control.

They can:

• Suggest text.
• Summarize information.
• Recommend actions.
• Generate drafts.
• Propose code.
• Analyze documents.

The person reviews and decides which result to use.

Artificial intelligence agents

An agent can receive a goal, plan several steps, use tools, and act on a digital or physical environment.

You can perform actions such as:

• Query a database.
• Browse an application.
• Execute code.
• Create files.
• Send information.
• Check results.
• Modify your plan.

This does not mean that he has his own will. Their behavior continues to be conditioned by objectives, permissions, instructions, tools, and control systems.

Agentic AI systems

Agent AI can coordinate multiple agents or processes to complete broader objectives over extended periods.

May include:

• Division of a task into subtasks.
• Delegation between agents.
• Coordinated use of tools.
• Monitoring of results.
• Adaptation of the plan.
• Human intervention at checkpoints.

The more autonomy the system has, the more important permissions, monitoring, activity logs, and mechanisms to stop it become.

This type of system can be used in processes of automation with artificial intelligence.

Reactive machines, limited memory, and theory of mind

Another popular classification distinguishes four functional levels:

• Reactive machines.
• AI with limited memory.
• Theory of mind.
• Self-aware AI.

It should be treated as a conceptual framework, not as a universal technical taxonomy.

Reactive machines

They respond to a specific situation without building a persistent memory about previous experiences.

They can be effective within a well-defined environment.

AI with limited memory

It uses recent information, context, or stored data to produce a response.

Many modern systems could partially fit this description, although the expression is too broad to explain their technical differences.

Theory of mind

Describe the hypothetical ability to interpret other people's emotions, beliefs, intentions, and mental states in a general way.

There are systems that detect emotional signals or infer concrete intentions, but that does not prove a complete theory of the human mind.

Self-aware AI

It would represent a system with its own consciousness or subjective experience.

There is no accepted evidence that current systems are self-aware.

This category belongs primarily to philosophical and hypothetical debate and should not be confused with the ability to converse or speak in the first person.

The same system can belong to several categories

The classifications are not mutually exclusive.

A modern assistant could be described as:

• AI specialized according to its capabilities.
• Deep Learning according to its architecture.
• Self-supervised according to part of his training.
• Generative depending on the result.
• Conversational according to the interface.
• Multimodal according to formats.
• Agent if you can use tools and execute actions.

Similarly, a fraud detection system could be:

• Specialized AI.
• Supervised machine learning.
• Predictive AI.
• Discriminative model.
• Decision support system.

That is why it is not correct to simply ask:

What type of AI is this?

The most precise question would be:

According to what criteria do we want to classify it?

What types of artificial intelligence currently exist

Currently, there are real systems of:

• Rules-based AI.
• Machine Learning.
• Deep Learning.
• Predictive AI.
• Generative AI.
• Artificial vision.
• Language processing.
• Recommendation systems.
• Robotics.
• AI agents.
• Specialized multimodal systems.

Under the ANI, AGI and ASI framework, these systems are usually classified as specialized intelligences.

Some have broad capabilities and can perform many tasks, but there is no universally accepted test to certify that a system has achieved AGI.

The ASI remains a hypothetical scenario.

This distinction helps to avoid two mistakes:

• Downplay the actual capabilities of current AI.
• Attribute to general abilities, awareness or autonomy that have not been demonstrated.

Conclusion

The types of artificial intelligence depend on the criteria used.

ANI, AGI and ASI describe breadth of capabilities. Supervised, unsupervised, reinforcement, and self-supervised learning describes how certain models are trained. The predictive, generative, conversational, or visual categories explain what results they produce, while agents describe their level of autonomy.

Current systems can combine several of these categories.

Therefore, to understand AI we must analyze:

• What task does it perform.
• What data it uses.
• How she has been trained.
• What result it produces.
• What tools can you use.
• How much autonomy does it possess.
• What supervision do you need.

Frequently asked questions about types of Artificial Intelligence