AI vs regular IT — what is the difference?

During public talks, I am often asked, “What is the difference between regular software and AI? Would you call rule-based systems AI?”.

I always had to resort to vague statements like “It depends” or “It is a vague border”. I was never really satisfied with my answer. So I took the time to think deeply about it and describe the boundary between regular IT and AI.

Where does IT end…

Software systems provide value to businesses by processing information or “data”. In their most rudimentary form, they allow us to create, store, update and retrieve it later on.

Retrieval, in its simplest form, performs a perfect match between numerical identifiers (e.g., your national security number). A little more advanced system may let you search on (parts) of a field value, make the retrieval case-insensitive, or perform basic pattern matching capabilities using regular expressions or similar (e.g. retrieve me all records for patients whose name start with “J” and end with “Smith”). The results may be displayed alphabetically, ordered by different fields, or grouped.

I doubt that anyone would call these systems “AI”. In the above examples, the criteria can be strictly defined syntactically, in form. There is no ambiguity in determining whether the letter “A” comes before “B” in the alphabet, salary is lower or higher than 50.000, or how grouping should occur. The functionality naturally translates into a formalised form like mathematics, SQL or code.

And AI start?

AI systems are of course also software systems that process information — so when do we call something AI instead of “regular” IT?

Fundamentally, it depends on the depth of interpretation and whether the processing (the control flow and outcome of the code) depends on the input. In other words, the processing flow is reconfigured based on the data.

AI starts when a) control flow is configured by data or b) information processing manipulates representations at a “meaning” level.

Property 1: control flow is configured by data

A first fundamental characteristic of AI systems is that the control flow is not static, like most regular IT systems, but can depend on the input data. Please note that we define “data” broadly; they could be rows from a tabular dataset or rules in a rule-based system.

Picture a simple program consisting of if-then-else statements. When this program receives input, the sequence of these if-then-else statements is typically determined by humans during software development.

Things change when the control flow is configured by data. An algorithm will determine the amount, the nesting and the specifics of if-then-else statements based on the input data.

Consider the prototype program below.

if Person.Salary > 10000 {
  Person.Influential = true
}

Traditionally, humans enter this code. When the code is output from an AI system, the decisions to pick the variable Person.Salary, the threshold value “10.000”, the output variable “Person.Influential”, and the value “true” are not left to a human designer but to an algorithm.

Examples of AI techniques that exhibit this property are machine learning, automated reasoning systems, time series forecasters, clustering and anomaly detection.

Property 2: manipulation of “meaning”

Let me clarify. With a “meaning level”, we intend an interpretation that cannot be straightforwardly derived from the data. Imagine we wish to characterise someone’s personality. One could use a theoretical framework like “OCEAN” that quantifies one’s Openness, Conscientiousness, Extroversion, Agreeableness, and Neuroticism and attribute 0–10 for each category and represent John’s personality as a vector (2, 9, 10, 6, 4).

These numbers in themselves do not represent one’s personality, of course. The vector could indicate a point in space at a particular time, demographic info — anything you like. The interpretation within the theoretical framework gives meaning to this data structure. So far, so good.

Manipulation of meaning

AI differs from IT because these data structures are also manipulated to derive new insights.

The manipulation of a data structure is restrained by its meaning.

For example, when the advice vector represents a point in space, it may make little “sense” to multiply the first two quantities (like the X and Y coordinate). If the two first components, however, represent salary and the number of connections on LinkedIn, one could say that the multiplication of these two could serve as a proxy for “influence”. In contrast, sorting does not depend on the meaning of the sorted words.

We base actions on conclusions in the meaning space

Of course, we do not simply want to derive new information: we wish to act upon it. For example, one could compute the optimal advertising message based on this vector.

In everyday conversations, we say that we personalise the advertising message to one’s personality, but in reality, we adapt it to this particular vector. And this vector is to represent — in the real world — one’s personality.

Observe the heavy influence of Plato and Aristotle in having true “ideas” reflected in numbers. However, it is but a feeble reflection of one’s personality, with two main simplifications: a) defining personality as a 5-dimensional characteristic (OCEAN), b) quantifying it as a vector.

Exploring the boundary

Imagine we wish to retrieve the “most similar records to a given record” from a database. What do we mean by “similar”? What is the meaning of similar?

1. If a record consists of a sole salary field, one could take the differences between the salaries. This difference then defines a distance metric. Sorting records by this distance metric would give you the most similar records. In this case, the meaning of “similar” can be easily defined and operationalised in a mathematical formula, being the subtraction (and absolute value).
2. But what if the record consists of an address field? Do we mean “closest in physical distance”? And if so, do we mean line-of-flight or using road infrastructure? Or the time it takes to use public transportation? In this case, we need to create more complex models of “distance” that deviate from a simple, physical Euclidean distance.
3. Maybe we intended similar communities? Imagine you are a real estate agent and wish to explore neighbourhoods identical to your portfolio. How can we quantify this information? We’d need to define what we mean by “similar” (same size of houses, same demographics of residents, similar prices, same amount of greenery, etc.) and quantify it to translate into a distance metric.

Example AI techniques that have this property include recommender systems and supervised (where a model of a class represents the concept) and unsupervised machine learning (where the distance metric should reflect meaning as in the example above).

Far reaching consequences

These properties imply an important transformation for software systems.

First, for the design of these systems, as their functionality is not fixed at design time, but at runtime. How to prove correctness? How to make them robust against adversarial attacks? How to measure performance? How to make sure the results “make sense” (as we manipulate meaning)? How to design a suitable distance metric?

Second, for the users — whether humans or machines. How to interpret results? How to avoid “hallucinations”? How to justify the decisions taken by the algorithm (e.g., in the example above, the chosen threshold or variable)?

AI opens up many possibilities, but designing and implementing AI systems requires careful craftsmanship and a critical attitude. I hope you see by now that these two properties have far-reaching consequences, that both make out their strength and weakness.