Talking about ‘AI’ – a call for more nuanced terminology

The term “AI” does not refer to a single technology—it is an umbrella term for various technologies, applications and products. To ensure a meaningful public debate on the capabilities, risks and regulation of AI, we should strive to use more nuanced terminology. This applies in particular to journalists, who shape public perception and establish new terminology.

In 2019, I launched the website aimyths.org together with Daniel Leufer. One of the eight myths was: “The term AI has a clear meaning”. Now, in 2026, the term “AI” could not be more ambiguous: a novel technology. An 80-year-old field of research. An industry. A marketing buzzword. A hype machine. Yes, even an ideology and a worldview.

The hype surrounding “AI” has been going strong since 2022: there’s a frantic push for AI adoption. Doesn’t your company use AI yet? We all need to “learn AI” right now, or else we’ll be left behind! AI literacy is the buzzword of the moment. At the start of the year, the UK’s Prime Minister, Keir Starmer, said he wanted to “mainline AI into the veins” of the nation — as if AI were a homogeneous liquid, a magical elixir.

At the same time, a growing AI-refusal movement is taking shape. For them, the term “AI” stands for the generative AI industry, with all its exploitative practices that harm both the environment and people; for a complex “apparatus” made of technology, institutions and ideologies; and for figures such as Sam Altman and an ideological conglomerate known as “TESCREAL”.

I myself completed my Bachelor’s degree in ‘Artificial Intelligence’ at the University of Amsterdam between 2010 and 2013. During my studies, I was able to witness the breakthrough of deep learning first-hand – and back then I “fell in love” with the principle of automated pattern recognition in data. In the years that followed, I observed two developments: 1. just how highly political and problematic that very pattern recognition is, and 2. how the AI hype works.

The chronic vagueness of the term ‘AI’ is fascinating. It is a “moving target” – the definition is constantly changing, depending on what the current state of the art is. “AI is always what hasn’t been achieved yet,” I once heard someone say. If even AI specialists can’t agree on how to define AI – where on earth are we supposed to start?!

I think that even if we cannot define AI, we can at least make an effort to identify specific types of AI systems by name. This article aims to contribute to that.

Of bicycles and aircraft

In their book “AI Snake Oil”, computer scientists and researchers Arvind Narayanan and Sayash Kapoor propose a thought experiment:

“Imagine an alternate universe in which people don’t have words for different forms of transportation — only the collective noun “vehicle.” They use that word to refer to cars, buses, bikes, spacecraft, and all other ways of getting from place A to place B. Conversations in this world are confusing. There are furious debates about whether or not vehicles are environmentally friendly, even though no one realizes that one side of the debate is talking about bikes and the other side is talking about trucks”

Now replace the word ‘vehicle’ with ‘AI’, and it perfectly describes what’s going on right now.

Which systems are referred to as “AI”? A chatbot like ChatGPT. A sensor on a hospital bed that detects when a patient gets out of bed. Software used by a bank to assess a person’s creditworthiness. A spam filter in an email programme. An algorithm that determines the social media feed. An image editing tool that doubles the resolution of a photo.

Some of these systems are relatively harmless in terms of their production conditions, data protection, bias and energy consumption. Others pose massive problems for social justice.

Left: AI-based sensor data processing on a small portable device; Right: Meta’s new “Hyperion” data centre is set to be as big as Manhattan

Because I love the “vehicle/AI” comparison from “AI Snake Oil” so much, I’d like to revisit it:

“Bicycles” in the world of AI are small machine learning models that run locally on a device without an internet connection, they process sensor data, are trained on small datasets, and perform a specific task very well and predictably. This is known as Edge AI.

“Aircraft” in the world of AI are so-called foundation models, including large language models (LLMs): models with hundreds of billions of parameters, generated on monopolised IT infrastructures that emit tonnes of CO2 and are based on exploitation: exploitation of the Earth’s natural resources, exploitation of workers in the Global South, and stealing of intellectual property.

Why bother differentiating AI at all?

The use of AI technology is being driven forward in many critical sectors, such as education, healthcare, research and the military. But what types of AI are we talking about here? There are significant differences between AI systems in terms of their technical complexity, transparency and predictability, production conditions, manufacturing costs, energy consumption, data protection and IT security. Distinguishing between different AI technologies has real-world implications for assessing risks, monitoring requirements and their regulation.

Surveys and data collections such as “Which companies in Germany are already using AI?” are of little value unless they distinguish between different AI technologies and unless respondents are able to tell the difference between AI and non-AI.

The German organisation Superrr writes: “How can we talk meaningfully about AI when it can mean everything (and nothing)? A digital rights perspective on technology and its regulation must be specific.”

I agree with the authors. Failing to distinguish between different types of AI is a key mechanism of the AI hype and should be regarded as deliberate misdirection: framing AI as a monolith, as something groundbreakingly new, as an unstoppable force of nature, reinforces the ‘AI inevitability’ narrative. It depoliticises the AI industry and creates a sense of helplessness. In this climate of fear, those actors who, under the banner of “AI safety”, claim to want to tame this force of nature—but who are in fact pursuing entirely different interests—can now make a name for themselves.

AI 101 – A Short Story

I get the feeling that the confusion surrounding the term ‘AI’ also stems from a lack of knowledge about the history of “Artificial Intelligence”. This section provides a simplified historic overview, intended to help contextualise the concepts of machine learning, deep learning and generative AI.

Some people think that “AI” has only existed since 2022, since ChatGPT – but who was it that filtered your spam emails and sorted your search results in the decades before that?

AI as a field of research has a history stretching back more than 70 years. By the late 1950s, neural networks capable of simple image recognition already existed. For a while, AI researchers believed that the whole world should be represented in rules (“if-then-else”), logic and knowledge databases; only in this way could intelligent behaviour be generated in computers. This approach is known as Symbolic AI or Good Old-Fashioned AI (GOFAI) and was used in so-called expert systems.

Machine Learning (ML)

The symbolic approach reached its limits: the wide variety and subtlety of human perception, language and thought could only be partially captured in rules. The idea began to take hold that data might be important for ‘artificial intelligence’. The field of machine learning gained prominence in the 1980s. In machine learning, statistical algorithms ‘learn’ patterns from existing data and represent them in the form of a model, which generalises these patterns to such an extent that it can also handle new data. This enables machine learning to solve problems that could not be solved using rule-based approaches.

In the 2000s, a new, seemingly inexhaustible source of image, text and video data opened up for AI research: the internet. At the same time, there were enormous advances in the development of graphical processing units (GPUs), computer hardware specialised in performing many simultaneous computational operations. The concept of neural networks now had the ideal breeding ground to realise its potential: the era of deep learning began.

Since 2012: Deep Learning

Deep Learning is a form of machine learning that uses multi-layered (deep) neural networks to identify patterns of varying complexity within large datasets and to generalise these to new data. The mathematics behind deep learning is a combination of linear algebra and calculus.

Deep Learning is used for tasks involving classification, prediction and pattern recognition. This gives rise to a wide range of applications: image recognition, detecting credit card fraud, predictive maintenance and customer churn prediction, recommendation engines, spell checking, speech recognition, OCR (recognising text in images), and a sensor that detects whether a specific movement has been performed.

These systems are now sometimes referred to as “traditional AI”. Traditional AI systems are narrow, i.e. they are built and optimised for a clearly defined task, e.g. “Is there a cat in this image?” or “How likely is it that the customer will cancel their subscription?”.

Problems with deep learning

1. Bias: By identifying patterns in large amounts of data, it also learns the patterns of social inequality and discrimination and reproduces them. Activists and researchers such as Cathy O’Neil, Joy Buolamwini, Sofya Noble and Virginia Eubanks highlighted the dangers of racist and sexist discrimination by deep learning systems as early as 10 years ago.

2. Lack of intelligence: Deep learning models understand neither context nor causal relationships nor the physical world, but rely exclusively on the principle of pattern recognition in (biased) datasets. This makes them prone to error.

3. Black box: The decision-making processes of neural networks are barely comprehensible or explainable, as they are inferred from data patterns. In AI-based decision-making systems that have concrete impacts on people’s lives and where accountability is required, this lack of explainability is problematic. Approaches to Explainable AI (XAI) remain limited.

Seit 2022: Generative KI & Foundation models

The year 2022 marked the breakthrough of a new “type” of AI based on deep learning: generative AI. Underlying techniques such as the Transformer architecture (for text), introduced in 2017, and the Diffusion technique (for images) enabled the creation of text and media with a level of realism never seen before.

AI research is increasingly focusing on the development of foundation models, i.e. large, general-purpose AI models. This has given rise to a “bigger is better” trend: the computing power required for training doubles every five months, the volume of data every eight months, and energy consumption annually. Due to these conditions, increasingly only large, industrial AI research laboratories are financially capable of participating.

With the breakthrough of generative AI, AI technology has, for the first time in history, become a consumer technology—accessible to everyone via a simple chat interface.

Issues with generative AI

As generative AI is based on deep learning, it inherits all the weaknesses and problems of deep learning and machine learning: bias, a lack of intelligence, and a lack of explainability.

And generative AI brings with it new issues. Just to name a few:

2. Copyright: Most generative AI models are trained using data that is protected by copyright and is used without the consent of the copyright holders.

3. Energy consumption: The scaling trend has a negative impact on the environment. AI data centres deplete local drinking water reserves, overload the electricity grid and generate significant CO2 emissions.

A helpful mental image for some of the AI systems discussed is this “risk pyramid”: AI technologies build upon one another, on their strengths as well as their weaknesses and risks. Note: This diagram does not include all types of AI technology; more on that later.

Cheatsheet

• Not all AI is generative AI

• Since 2012 deep learning – a form of machine learning based on so-called neural networks – has become widely established

• Generative AI is based on Deep Learning

• Traditional, non-generative AI systems are still in use

• Symbolic/GOFAI/expert systems are still in use

Differentiate, but how?

One of the standard ways of categorising AI is by the traditional fields of AI research: Natural Language Processing (covering everything related to language processing) and Computer Vision (covering everything related to image processing). A very common way of visualising this is the nested hierarchy: AI > ML > Neural Networks > Deep Learning, which is now often supplemented by generative AI.

Other ways of categorizing are possible: The European AI Act, for example, classifies AI systems by risk category. In addition to classifying AI systems by application area (“AI in pharma”), use case (“AI for image classification”), type of data or level of explainability, they could also be classified on the basis of their technical characteristics:

• How large are the AI models, how many parameters do they have: a few thousand or several hundred billion?

• What type of machine learning is used: supervised, unsupervised, self-supervised or reinforcement learning?

• Where does the AI system run: locally or in the ‘cloud’?

• Open source vs. closed source

I have been wondering: when are certain distinctions and classifications helpful? In which contexts do they lead to greater clarity and a better understanding of the opportunities and risks?

We need a linguistic “middle ground” between a rhetoric that lumps all systems together under the term “AI” and one that treats every AI system as a separate case. I would like to offer a few suggestions for differentiation here that might be suitable for the general discourse on AI:

Distinction No. 1: Not all AI is generative AI

In most cases, when people talk about “AI”, they mean ChatGPT or other chatbots and image generators – in other words, they actually mean generative AI. But not all “AI” is generative AI. As explained in the “short story” above, traditional AI based on deep learning has existed for a long time (over 10 years) before generative AI.

Traditional (non-generative) AI is not an outdated technology, but is still in use because it solves problems in certain categories (classification, prediction, anomaly detection) more effectively – that is, more accurately and more resource-efficiently – than generative AI.

Why not just call it “AI”?

Generative AI brings with it all the problems of traditional AI (deep learning) — but adds further issues on top: theft of intellectual property, hallucinations and misinformation, and immense energy consumption for training and use. That is why we should call generative AI by its name.

Distinction No. 2: Distinctions within generative AI

Now that generative AI is so ubiquitous, we should also identify specific areas within this field:

Foundation models (or: General-purpose AI models)

These are generative AI models capable of performing a wide range of tasks, rather than being limited to specific tasks like traditional narrow AI. Examples include: GPT-5 (for text), Midjourney or DALL-E (for image generation) and Sora (for video).

Why not just call it “AI”?

Because they are the ‘building blocks’ from which many of today’s AI systems and products are built. Developing a foundation model is resource-intensive and costs several hundred million dollars – as a result, only a handful of players are actually capable of building them. The vast majority, however, can derive their own models by fine-tuning these foundation models. Foundation models are the main focus of many current discussions on opportunities, risks and regulation (see, for example, the latest “International AI Safety Report 2026” or the section on “General-purpose AI models” in the European AI Act).

LLMs

Large Language Models (LLMs) are the technology behind chatbots and AI agents. Technically speaking, they are simply highly sophisticated ‘autocomplete’ systems; in other words, they can predict the next word because they have learnt how human language works by analysing vast amounts of text. This is why they are also referred to as ‘stochastic parrots’. They are capable of producing highly convincing misinformation. An LLM is a form of general-purpose AI (or foundation model) and generative AI.

Why not just call it “AI”?

LLMs are at the heart of the new AI era, which is why we should call this technology by its name. At the same time, there are a multitude of AI technologies that cannot be reduced to LLMs.

AI Agents

Unlike standard AI chatbots, AI agents are equipped with a kind of “executive” capability, meaning they can not only generate text but also access operating system functions or control entire user accounts. They are used in software development, in research, but increasingly also for personal use: “Book me a trip to Berlin”. AI agents represent a worrying paradigm shift in which human control is deliberately handed over to a system that only appears to “think” and acts autonomously.

AI agents are based on LLMs, which in turn rely on deep learning. AI agents therefore inherit all the vulnerabilities of their parent technologies mentioned above (unpredictability, hallucinations, errors, bias) whilst being granted greater autonomy and access rights. What could possibly go wrong? Many IT experts are warning of the immense risks to IT security and privacy, and significant damage has already occurred. Furthermore, AI agents often run in a continuous loop, sometimes for hours or days on end – and their energy consumption is correspondingly high.

Why not just call it “AI”?

Because AI agents are currently at the centre of the AI hype and signal a dangerous paradigm shift of “human-out-of-the-loop”. As described in the risk pyramid above, they combine the vulnerabilities of large language models (LLMs) and deep learning with reduced human oversight.

Distinction No. 3: Automated Decision Making (ADM) systems

Automated Decision Making (ADM) systems refer to a category of software used in situations where decisions are made automatically on behalf of people: banks use them to decide whether to approve a loan for an individual. Companies use them in the recruitment process to automatically pre-screen applicants. These systems are based on traditional AI rather than generative AI. Please note: however, this term often also includes systems that might under certain definitions not even fall under the category of “AI”, for example because they make decisions based on pre-programmed rules.

Why not just call it “AI”?

This is because these systems represent a particularly noteworthy application of AI technology: they have a significant impact on people’s rights and participation in our world, and the use of deep learning can very easily lead to discrimination. Under the European AI Act, they fall into the category of ‘high-risk systems’.

Distinction No. 4: Edge AI

Edge AI (also known as embedded machine learning) refers to small, energy-efficient neural networks that run directly on a device—without a network connection and without linking to a ‘cloud’. They often process sensor data, such as motion data, audio data or camera data. Examples: Face ID on smartphones, “wake word” recognition in voice assistants, noise-cancelling headphones or a necklace with a “fall sensor” for elderly people.

Why not just call it “AI”?

Because they are the ‘bicycles’ of the AI world: small, resource-efficient models that are good for privacy. They are in a completely different league to the large foundation models.

Distinction No. 5: Product vs. Technology

Even if this distinction is blurry, we can say: ChatGPT comprises various AI models (e.g. GPT-5, GPT-Image, formerly DALL-E) and, when combined with a chat interface and other functionalities, it becomes the functional product known as “ChatGPT”.

Whether a user’s prompts are stored and may even end up in the training data are all product strategy decisions made by a company, not technical necessities of LLMs.

And the flattering tone of an AI chatbot (known as “sycophancy”) is not inherent to the core LLM technology, but a consequence of product design.

Distinction No. 6: Is it even AI?

At this point, I’d like to add that a lot of what sounds like ‘AI’ or is marketed as such doesn’t actually use any AI technology. A tool that synchronises calendars might sound like AI, but ultimately it’s just a rule-based algorithm. Many problems don’t need AI or machine learning at all! Most of the software running on our computers or “in the cloud” is still just ordinary programming code – i.e. if-else statements, switch statements and database queries. And that’s a good thing, because it makes digital processes more transparent, predictable and secure.

Bringing some order to AI terminology – a difficult task

This is, of course, far from being a comprehensive taxonomy: the concepts mentioned overlap and span various dimensions.

In a field that is evolving at breakneck speed (so they say), where new trends and technologies pop up overnight, and everything is in flux and fluid – can a taxonomy, a rigid linguistic system, really do it justice? I don’t think the field of AI is moving quite that fast after all. At least not at a fundamental, underlying technological level. The ‘application/product’ level may be moving at the speed of light. But let’s not be fooled by that.

At the same time, a new “AI winter” is looming: it is becoming increasingly clear that deep learning alone is no longer making progress through scaling. Recent research, for example, is attempting to combine symbolic AI with deep learning, a field known as neuro-symbolic AI. At the same time, the value of smaller, narrow AI models and applications that can run on consumer hardware without a cloud is being rediscovered. What used to be an aeroplane might, under certain circumstances, shrink to the size of a bicycle.

This article is a call to reflect on how and when we should replace the ambiguous and highly multidimensional term “AI” with clearer terminology – in order to communicate more effectively about the capabilities and risks of a specific technology in a specific field of application.

Often, simply naming the type of AI system is not enough to distinguish between them. In the paper “What Do We Mean When We Say ‘AI?’ Why Context—Not Volume—Should Guide Emergency Medicine”, Christian Rose writes that when evaluating the use of AI, one should also distinguish between contexts: “Which tool, for which patient, in which workflow, and at what acceptable trade-off?”

Daniel Leufer, my colleague at AIMyths and Senior Policy Analyst at Access Now, comments: “You need to drill down to the level of specificity necessary to make your point in a specific context.”

Which types of AI systems are particularly relevant to you? Which categorisation makes sense to you?

Resources

More AI-critical thinkers

• Arvind Narayanan, Karen Hao, Matteo Pasquinelli, Tante, Gary Marcus

• Institutions: Ada Lovelace Institute, AI Now Institute, DAIR, AlgorithmWatch, Algorithmic Justice League

• Bias and social justice: , Timnit Gebru, Margaret Mitchell, Emily M. Bender, Alex Hanna, Abeba Birhane, Safiya Umoja Noble, Ruha Benjamin, Johanna Bryson, Joy Buolamwini

• AI and the environment: Emma Strubell, Sasha Luccioni,

• AI and cybersecurity: Meredith Whittaker, Georg Zoeller

More about “how we talk about AI”

• For the aimyths website I built a “Terminology swap tool” and a “Headline Rephraser”

• AI & Responsible Journalism Toolkit (University of Cambridge, DesirableAI)

• Book “The Need to Rename Tech”

• Margaret Mitchell: “No, ‘AI’ is not a stochastic parrot”, a plea against equating LLMs with AI

More technical classifications

• Der latest “2025 AI Index Report” from the Stanford Institute for Human-Centered Artificial Intelligence (HAI)

• Risks of General-purpose AI systems: International AI Safety Report (2026) https://arxiv.org/abs/2602.21012

• IBM’s technical introductions to several AI concepts, e.g. the one about Foundation Models: https://www.ibm.com/think/topics/foundation-models