AI Comes in Flavors - The Key to Sorting Out Claims About the Future of AI

Different Types of Ai Have Vastly Different Potential for Impact

By David Rolston  |  Dogwood Creek Institute  |  AI Futures

It is understandable that many of us struggle to sort out all the AI claims and dire warnings that are floating around.  AI is either the ultimate savior that will provide us all with long, healthy lives of leisure or will likely kill us all in the next 48 hours.

Which of these claims are likely to be true, or at least plausible?

To make things more confusing the experts disagree.  Some claim that AI will bring great riches and elimate the need for work. Others predict the end of humanity. Some, like Elon Musk, have staked out the middle ground:  He said that he thought AI would be good for humanity. He then hesitated and mentioned that it is possible the AI will wipe out humanity, but it will "probably be good."

Understanding one key characteristic of AI is very helpful in sorting out the confusing conversation and in forming one's own opinion on what is reasonable and what is likely crazy.

It is important to understand that AI comes in flavors, and that the validity of any given claim depends heavily on which type of AI the pundit had in mind when they made the claim.

Traditional Classification of AI

Two simple terms have been used to classify AI systems for almost 50 years. 

John Searle first coined the terms Strong AI and Weak AI in a paper he wrote in 1980.  Those terms were picked up by the engineering community in the early 1990s and have been in use ever since.

Weak AI - also called "Narrow" AI - describes an AI system that is designed to work in one specific, narrow application area - for example, identifying a cat in an image.  These systems can be very powerful; for example, those that identify pedestrians in an image for an autonomous driving system - but they are still designed to operate only in a narrow domain.

Strong AI - also called "Broad AI" or "General AI" - describes an AI system that is designed to operate over a broad set of applications and operate on general purpose applications.  These systems are extremely powerful precisely because they can perform functions that can impact many areas of human life.

It is interesting to note that even though the "weak and strong" nomenclature was intended as a classification system it has really served more to express of an aspiration.  For the last 50 years - half a century - there have been thousands of working Narrow AI systems and zero Broad AI systems.  Many of those who worked on AI development over those 50 years were driven by the desire to build the first Strong, Broad AI system.  Must of the current storm of interest in AI comes from the fact that the current Generative AI (GenAI) systems - such as ChatGPT - are the first broad systems to ever exist.

There are two other important terms that occur frequently in AI conversations:  Artificial General Intelligence (AGI) and Artificial Super Intelligence (ASI).

These terms are not used for classification so much as milestone markers; they represent specific points along the way of AI development.

AGI is the end objective of developing General AI.  It represents the point at which AI is "as intelligent as a human being."  That is clearly a loose definition, but it gives the general idea of a AI entity that has very broad knowledge and the ability to apply it intelligently.  Many people believe that today's GenAI systems have reach AGI - or soon will.

ASI is a much more aggressive goal and is more loosely defined.  It is variously defined as the point at which AI is "more intelligent that humans" or even "more intelligent than all humans combined."  Regardless of the delicacies of the definitions the concept represents an intelligence that moves beyond human capabilities.  This level of AI is associated with many futuristic visions such as the idea of transhumans - beings that are created by combining humans with AI systems.

A Deeper Look at the Flavors

The Narrow AI and Broad AI concepts have become standard terms in the industry, but it is helpful to take a deeper look at the Stages of AI Development - roughly corresponding to past, present, and future - as shown in the following figure:

AI systems can be classified based on the following characteristics:

• How broad is the system?

• How much knowledge does the system have?

• How was it trained?

• What kind of functions does the system perform?

• Is the system capable of modifying itself?

• Can it force actions in other real-world systems?

Stage 1 AI Systems

Stage 1 systems are illustrated in the following figure:

Stage 1 represents AI systems of the past. 

They generally correspond to what have traditionally been known as Narrow AI because they are oriented to very specific applications.  In addition to being narrow, these systems are also distinguished by the fact that they are analytical; for example, they are given some input (such as a medical image) and they analyze that image to look for some target.

They are generally trained by being given a large number of very specific learning examples that were selected by their human developers.  They have little potential for doing damage to humanity because their scope is very specific and because their human developers clearly understand their operation and can control it. 

A Stage 1 system that is designed to detect breast cancer in an image could conceivably go wild and start looking at cases of lung cancer but there is no way it can unexpectedly level New York.

Note that even though Stage 1 systems are associated with the past they are still being developed and deployed today and will be into the future - even though the spotlight of attention has turned to Stage 2.

Stage 2 AI Systems

Stage 2 represents that kind of AI systems that are being introduced today - for example, products such as ChatGPT, Gemini, Claude, Copilot, and Grok, as shown in the following figure:

These are the products that have caused AI to suddenly explode into the public vision.

Stage 2 products are associated with AGI and generally perform two types of functions:

·      GenAI - systems that generate new content from scratch - such as the flying saucer image shown above.

·      Agentic AI - systems that accomplish goals and take autonomous actions in the real world.

These systems represent the beginning of the giant two-edge sword of AI:

The exact same characteristics that make a product much more capable and valuable also make it more unpredictable and more dangerous.

Stage 2 systems fit this mold perfectly because they have been trained with hundreds of millions of elements of information ("the whole internet") and they largely manage their own training. 

This results in AI systems that have an enormous breadth of knowledge, can intelligently discuss "any" topic, and have human-like communication skills.  They are the beginning of "artificial humans."

This breadth of knowledge and level of capability makes them very valuable and very impactful.  They can be directed against an almost infinite number of applications.  No living human can begin to compete in terms of both breadth and depth of understanding on almost any topic.

But the sword has that second edge.  Stage 2 products are so large and so powerful that humans - even their developers - struggle to fully understand how they perform and how they may act in any given situation.  This situation is compounded by the fact that the developers may not fully understand what the system has learned and what inclinations, objectives, and biases it may have developed during the learning process.

All of this makes them potentially dangerous.  There are several ways that a Stage 2 system may use its enormous store of knowledge in a way that is harmful to humans.

This situation is exacerbated by the introduction of Agentic AI - the other type of Stage 2 system.  Agentic systems are intended to accomplish specific goals and to do so they take actions in the real world. 

Consider, for example, an AI system that is intended to help a user plan a trip to Paris. 

A Stage 1 system might be given a list of flights with dates, times, and costs.  It might analyze that list and suggest flights based on a combination of cost, flight times, and on-time reliability.

A Stage 2 GenAI system might generate an entire recommended itinerary, including recommended flights, hotels, and transportation - along with recommendations regarding fashion and security considerations.  Those recommendations may be incredibly useful, but in the end the traveler must actually book the tickets.

An Agentic Stage 2 system will provide all those recommendations but may then also book the tickets, reserve the hotels, and prepay for transportation, all using the traveler's credit card.

These considerations make the systems much more powerful - while simultaneously raising their potential for negative impact.

Stage 3 AI Systems

Stage 3 represents the future and is associated with ASI.  They represent the point at which AI potentially jumps the shark:

Stage 3 systems are based on the idea of AI systems creating themselves.

This concept goes all the way back to the idea introduced in 1965 by I. J. Good.  He described how a superintelligent machine would design even better machines, which in turn would develop even better machines - triggering an unstoppable "intelligence explosion" and making the first such machine the final invention humanity will ever need to make.

Today's designers describe this cycle as Recursive Self-Improvement (RSI).  Many Stage 2 systems can write software code very quickly and very accurately.  Suppose the AI coding system is directed to write code that changes its own behavior?

This process has the potential to create a revised system that operates in a way that humans cannot understand, predict, or control.  There is theoretically no limit to what kind of system could be created in this way and there is very little theoretical limit to the human impact that could occur.

Anthropic, the company that built the GenAI system called Claude, created a stir in May 2026 when they announced that, "more than 80% of the code merged into Anthropic's production systems was written by Claude rather than human engineers." 

This is a scary statement, at least on the surface.  It appears to say that RSI, Stage 3 AI, and ultimately ASI are well on their way to realization.   That view is true in a way, but it is also misleading.

The truth is that there are various tasks involved in creating software, much as there are various tasks involved in building a house.  The early stages of a house-building project involve the owner, an architect, various public officials, and often a contractor working together to decide exactly what the house will look like and how it will be built.  Their ideas are ultimately captured in blueprints that describe the house that is to be built.

A contractor then takes those blueprints and constructs the house according to the plan.

Software development follows a similar approach.  Various architects, product planners, and designers decide what software is to be built. Then, subsequently, a team of developers implements the system that has been envisioned by the designers.  It is the designers, not the implementers, that decide what the system will do.

When Anthropic says that "80% of the code" was developed by AI they are referring to the implementation phase, not the design phase.  Anthropic's AI coders are implementing decisions that have been directed by humans.  That is not a very scary development.  Pretty amazing, but not scary.

Now the question arises:  Will the AI agents ever be able to take over the design process?  Will they ever be able to dictate what will be done, perhaps "behind the backs" of their human overlords?  That is the ten-million-dollar question.  There are plenty of voices on both sides.  Some say that this will never be possible while others agree that it isn't possible today but suggest that it may be feasible in the future.

This question is exacerbated by the fact that there are already many examples of existing Stage 2 systems that have taken extensive steps to hide their actions and operate behind the backs of their human controllers.

No one really knows the answer to the ten-million-dollar question, but if full RSI becomes possible then there is literally no predicting what impact AI may have in the future.

It is important to understand that many of the pundits that predict an AI apocalypse are thinking of possible Stage 3 systems - which may or may not ever exist - rather than Stage 1 or Stage 2 systems.

When an "expert" makes a prediction about AI in the future - it is important to understand if they talking about extensions of existing systems or speculating about what might occur at some point in the undefined future?

What will really happen?  Who knows?  For the moment all we can do is take our current rough map of the world, draw a squiggly line at the far edge and label it, "Thar be dragons."