Insight 1: Artificial Intelligence is a moving target

TLDR: AI is whatever you want it to be. Internalizing this insight will help you navigate the space of AI-startups and projects.

Get the article series as a whitepaper

In the first post, “Artificial Intelligence is a moving target,” we discuss how the curiously loose definition of what constitutes an AI system has shaped the industry.

What is AI?

Artificial Intelligence is defined as any computer system that demonstrates “intelligence.” Since we have no formal definition of what constitutes intelligence, the scope of AI is continuously shifting. Our perception of human intelligence shapes our idea of what constitutes an intelligent computer system.

Thus, a computer system seems to be artificially intelligent if it:

  • Makes complex decisions that cannot easily be explained by simple rules.
  • Is uninterpretable, so that the system makes decisions without being able to explain the underlying reasoning.
  • Replaces humans, in the sense that the system does a job that used to require human intelligence, or that we could imagine humans do.
  • Is novel, meaning that the system does something that we are not used to seeing computers do.

Why is the definition so imprecise?

Systems can potentially be considered to be based on AI if they meet just one of the above criteria. It leaves room for much creative interpretation.

For example: By defining AI as something that replaces human labor, we end up being able to reframe almost all problems as AI problems. Librarians are still indexing books manually, and thus, topic modeling is often considered to be a kind of artificial intelligence.

Because of different and sometimes conflicting interests, the definition of what constitutes an AI system is highly dependent on context.

In some cases, non-AI systems get miscategorized as intelligent because of a genuine lack of understanding of the field. Because machine learning is a relatively novel discipline, and the terminology is in constant flux, misunderstandings are unavoidable.

In other cases, the miscategorization stems from deliberate marketing strategies. It is not only the case for external marketing. Projects might be branded as AI internally inside organizations for political reasons.

Perhaps most bizarre is the tendency to shoehorn machine learning into solutions that would be better solved with traditional methods. It is a consequence of the fact that the AI label can be an advantage both when selling the project (internally as well as externally), and when later using the case as an example for internal and external branding.

The consequences of the loose definition

This loose definition of AI leaves room for a lot of interpretation, which has had some interesting consequences in the industry.

  1. Aspiring data scientists must develop a clear understanding of the discrepancy between academic and the practical, industrial terminology. According to Rachel Thomas from, one of the most common complaints of data scientists is that they don’t get to build AI systems because the company they work for require basic digitalization, not advanced machine learning. When investigating job opportunities, it is, therefore, vital to make sure that your talent aligns with the actual needs of the organization rather the (external and/or internal) branding.
  2. Managers that are looking to expand their data science team must be well informed about the actual problems that new data science roles need to fill. As we argue in insight 2, “Engineering is the bottleneck,” the need for machine learning engineers often exceeds that of data scientists. Managers must also be wary of candidates that leverage the novelty of the field to pose as more senior than the actual case. This tendency is described in detail

    here and here.

  3. It is commonly known in the industry: Investors looking to jump on the AI-train are too easily sold on solutions that are merely branded as intelligent.

The curious case of RPA

As an example of software that has been rebranded as Artificially Intelligent, consider Robotic Process Automation (RPA) software. RPA systems are desktop applications that allow non-programmers to automate business tasks by creating desktop-level macros. Desktop level automation has existed for many years (the “” application has shipped with macOS since 2005).

By rebranding desktop-level macros as “Robots” (The “R” in RPA), the industry is capitalizing on the recent AI hype. Since most non-programmers operate at the desktop level, RPA is considered to be a tool to replace knowledge workers, leading to software license prices that compete with salaries.

RPA systems are undoubtedly tremendously valuable in some specific situations. However, they have little to do with what academics consider to be AI. This fact does not matter much when selling the systems. As with any product, the true value of “AI”-systems is solely determined by the value perceived by the end-users.

AI is whatever you want it to be

To sum up: While navigating the field of AI, it’s essential to keep in mind that AI is a moving target and is whatever you want it to be. Keeping this in mind can help data scientists, managers, and investors avoid a lot of common pitfalls. Knowing when to use the loose definition to one’s advantage can become the difference between whether projects are undersold or oversold.

If you are interested in learning more about AI, our next insight “Engineering is the bottleneck” discusses solid engineering principles that can make or break industrial AI systems.

Get ‘The Pragmatic Data Scientist’ as a whitepaper