In today’s column, I examine the vital premise that not all AI is the same. The issue is that most indications about how AI is doing this or doing that are depicted as though all AI is basically the same. But not all AI is the same. There are important differences and distinctions.
The advent of generative AI and large language models (LLMs) has become the popular form of AI that many tend to think of when someone says that AI has been used to perform some amazing feat. There are other types of AI. This is important to recognize, since any lessons learned or attempts to use AI on a similar task will be contingent on knowing which type of AI was used for a particular setting. To some degree, we are often misleadingly comparing apples to oranges, namely, treating AI that works one way with other AI that works differently. Fortunately, a new taxonomy of AI in the public sector provides helpful insights for differentiating among types of AI.
Let’s talk about it. This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
AI Lumped Into One Bucket
First, let’s drift into the near past. You might remember that before generative AI, a popular form of AI was known as expert systems or knowledge-based systems. Those are still around, though in a lesser fashion and without the pizazz of LLMs. A twist is that the rising tide of hybrid AI, also known as neuro-symbolic AI, is resurrecting the best of rules-based systems and commingling them with LLMs; see my in-depth analysis at the link here.
You likely wouldn’t see anyone differentiating AI these days. The overall semblance of AI has blurred into an amorphous blob, and just about everyone is claiming that they are using AI. It doesn’t seem to matter whether the AI is truly AI. Merely tout that you are using AI — the world will be eager to hear what you have to say. Stock prices go up when a company announces they are leaning heavily into AI, even though no one knows what type of AI they are using and how they are employing AI.
I’ve previously laid out the numerous and varied definitions of AI; see the link here. You might be surprised to know that there isn’t a single, global, fully accepted definition of AI. Thus, you can generally get away with calling anything AI and likely find a definition that will support your contention. It seems unimaginable that we all speak of and worry about AI, yet there isn’t a commonly accepted definition of what AI is. Makes your head spin.
AI Governance Is Crucial
No matter which type of AI someone chooses to use, they need to think carefully about the crucial role of AI governance. The AI being adopted in organizations should be suitably overseen. Rushing into AI might seem expedient, but if the overarching elements of AI governance haven’t been suitably planned and realized, the odds are that the AI is going to create more problems than it solves.
AI governance applies to both the private sector and the public sector. Readers might recall that I’ve closely covered AI governance for commercial enterprises and for governmental entities. The nature of AI governance in the public sector has many similarities to the private sector, but at the same time, there are highly notable differences. One way to conceptualize the difference is that commercial AI governance primarily seeks to govern organizations that create or deploy AI in pursuit of business objectives, whereas public-sector AI governance generally seeks to govern the exercise of governmental authority through AI while dutifully preserving democratic values and the rule of law.
I previously discussed AI governance in the public sector; see the link here, including providing this succinct description:
- AI governance in the public sector: “Public-sector AI governance encompasses the principles, structures, practices, and stewardship through which a public-sector entity ensures that its development or procurement of AI systems, and the fielding, use, and eventual retirement of those AI systems, are undertaken in a properly documented and effective way while safeguarding democratic values, individual rights, public trust, and the rule of law.”
One notable upshot is that if we are going to insightfully identify best practices associated with AI in the public sector, we ought to be clear-cut about what type of AI is being considered. This is again the apples versus oranges conundrum.
Public Administration Research
Contemporary studies conducted by public administration scholars and researchers tend to explore how AI is used in and impacts the public sector, but those studies typically do not sufficiently differentiate what type of AI is being used. In other words, a cited case study or even a large-scale survey says that AI is beneficial in certain ways and that there are downsides to be dealt with, but this is done without articulating what AI was actually put into use.
A recent paper entitled “A Technical Typology of AI Systems in Public Administration” by Jonathan Rystrøma, Chris Schmitzb, Nathan Daviesa, Gerhard Hammerschmid, Albert Meijer, Chris Russell, arXiv, June 30, 2026, made these salient points (excerpts).
- “Research on artificial intelligence (AI) in the public sector often treats ‘AI’ as a single category, neglecting technical distinctions between different AI systems. But these distinctions affect how different systems impact core public values like accountability, procedural justice, and non-discrimination.”
- “Our analysis draws on a specific sample of the most highly cited papers shaping public administration and digital government scholarship on AI (2019–2025).”
- “Currently, researchers risk either underspecifying their scope by using no taxonomy at all — invoking ‘AI’, ‘algorithms’, or ‘automated decision-making’ generically — or relying on an unsuitable existing taxonomy, which does not clearly track such properties.”
- “There is a clear need for a typology of AI systems targeted at clarifying their public-value implications.”
- “We introduce a typology of five categories of AI systems: hand-coded, glass-box, black-box, general-purpose, and agentic systems.”
As noted above, the paper looked at the most highly cited research works in public administration and AI during the time span of 2019 to 2025. The description of the AI was often broad and bland. It is challenging to reuse the findings of that body of research when you cannot know which of the AI was apples and which was oranges. You might abide by recommendations for oranges when you are avidly engaged in adopting apples in your entity.
Typology Of AI
The last bullet point in the above excerpts mentioned that the authors had derived a typology of AI that consists of five categories. They recommended that, on a go-forward basis, public administration research should incorporate those five categories. For example, a public sector case study ought to dive deeply enough into which AI was adopted so that the specific category would be apparent. Likewise, a survey of AI usage across-the-board in the public sector might wish to showcase the five categories and collect data about usage within each category accordingly.
The authors go into detail about the five categories, and you are encouraged to read the paper to get the nitty-gritty. Here are some quick one-liners from the paper that might whet your appetite and provide initial clarity about each category:
- 1. “Hand-coded systems. The first layer is hand-coded systems: systems whose decision rules are authored in code rather than learned from data.”
- 2. “Glass-box systems. The second layer is glass-box systems: systems whose decision rules are learned from data, but whose learned logic remains inspectable — at least by experts.”
- 3. “Black-box systems. The third layer is black-box systems: systems whose learned logic resists meaningful inspection, even by experts.”
- 4. “General-purpose systems. The fourth layer is general-purpose systems: systems pretrained on general tasks — such as next-token prediction — that can be adapted to diverse downstream applications through mechanisms like transfer learning or natural language instructions.”
- 5. “Agentic systems. The final layer of the typology is agentic systems: AI systems that can pursue complex and general goals, act with autonomy, and affect their environment.”
As I will get to in a moment, those who are highly versed in AI might grumble about the five categories and wonder whether it is feasible to jam a complex AI into a particular category. No worries, I will address this momentarily.
Example Of The Value Add
To give you an idea of how the use of the typology could materially enhance the value of AI-focused public sector analyses, suppose that a study made this overarching conclusion:
- Study finding: “Government agencies are increasingly using AI.”
That’s interesting but somewhat indefinite. If the research had made use of the typology, perhaps their conclusion might have stated this instead:
- Study finding: “Government agencies increasingly are deploying glass-box predictive models, more so than any other category of AI. Meanwhile, pilot or experimentation with general-purpose AI LLMs is next in line and will likely emerge in the next year or so.”
I would assert that the second description provides much more value. By leveraging the topology, the findings are more explicit and readily usable. The initial non-typology-based approach was rather nebulous and simply stated that AI was increasingly being used. That is marginally helpful. Once the typology had been used, the findings became much more illustrative, revealing that glass-box predictive models were predominant and that general-purpose AI LLMs are anticipated to be the next wave of adoption.
AI Complexities Can Cloud Things
A typology is not a free lunch. One notable difficulty in making use of a typology of AI is that the field of AI is rapidly evolving and tends to shift toward blending types of AI together. It is a bit of a head-scratching challenge to place a complex AI into just one category. This raises the thorny question of whether to list a specific instance as being in two or more categories, or whether a threshold is required to do so aptly.
For example, assume that an AI is 10% hand-coded and 90% glass-box. Does this AI belong in both categories or just one category? I’ll give you a moment to mull this over. Take your time. Okay, you might claim that the AI should be listed as a glass-box type because that accounts for the preponderance of the AI at the indicated 90%. A counterargument is that this unfairly leaves out the 10% that was hand-coded. Someone might rightfully assume that the AI is 100% a glass-box type if you don’t list the AI in both categories. What to do?
Well, if you list the AI in both categories, some might assume that it is an even 50/50 split. We know that’s not the case. The glass-box is at 90%, while the hand-coded is merely 10%. Thus, being in both categories might be misleading.
On and on this goes. There are no easy answers, and the key will be that any use of a typology should be done judiciously. Those leveraging a typology should spell out exactly what assumptions they made and what parameters they opted to use.
The World We Are In
I greatly applaud efforts to formulate taxonomies and typologies regarding AI. Keep up the good work.
My sincere hope is that discussions about AI governance, especially in the public sector, will gradually shift away from regulating AI as a monolith and move toward regulating different classes of AI according to their capabilities, levels of autonomy, and impacts on public values. This mirrors developments in other domains, such as the EU AI Act, where regulatory obligations increasingly depend on the nature and risk profile of the AI system rather than simply on whether it is labeled “AI.” For my analysis of the legal aspects of AI, including the EU AI Act, see the link here.
A final thought for now. The famed philosopher and literary critic George Henry Lewes made this pointed remark: “Science is the systematic classification of experience.” The AI field needs to mature and establish useful universal definitions and classifications. It’s for the sake of science and for the good of humanity.
