What is AI?

Introduction

AI is an acronym for artificial intelligence. Although some people argue that assistive intelligence or augmented intelligence is more accurate, the majority of the time, AI stands for artificial intelligence.

But what does that mean? For starters, there is no universally accepted definition of AI. You can ask different experts and they will give different answers.

There are many definitions out there that define AI as computer programmes doing things that would generally be assumed to require human intelligence. These definitions are problematic as they contribute to the anthropomorphising (attributing human characteristics to inanimate or non-human entities) of these technologies which ultimately lead to misunderstandings of what they can and cannot do, and also because what is considered to require human intelligence is constantly shifting.

Essentially, AI is a field of loosely related computer-based and statistics enabled technologies that, based on the input of data (information),instructions and parameters, can produce outputs such as predictions, recommendations and content.

Many people use the OECD’s definition of AI:

This definition is quite wordy and jargony and can be seen as being inaccessible. But many prefer it as it is purely descriptive and does not make equivalences to human traits or otherwise.

To add to the complications, there is no single “AI”. AI is more of an umbrella term for a wide range of computer and data driven technologies, including software that filters spam emails away from your inbox, computer programmes that allocate hospital appointments, chatbots giving customer service advice, smart devices in the home and, self-driving cars.

While recent years have brought heightened public interest, AI is most certainly not new. The idea of intelligent machines has been around for well over a century and can date back to ancient myths and tales of automata.

Modern AI, however, is a 20^th century phenomenon, and it goes back to the 1950s when the term artificial intelligence (AI) was coined and launched as a formal field of research at the now famous 1956 Dartmouth College conference. It was initially concocted as a marketing term which has evolved over the years to mean very different things to different people. The evolution of the field can be tracked over the decades through a series of winters (troughs) and summers (peaks) corresponding to rises and falls in computing power and funding.     

The definition of the term “intelligence” is up for debate. Current AI models cannot understand or think for themselves. Rather, they are statistical models that may give the impression of intelligence based on the outcomes they produce from patterns identified in their training data. While there is speculation about the possibility of future AI models developing true intelligence, we do not currently have actual artificial intelligence.

Categories of AI

In public and media discussions of AI there are often very different understandings of what this means, or what AI is capable of. The public image of AI has often been shaped by the ways it has been presented in movies and science fiction novels as something that can think for itself and control its own actions, but this is very different from the reality of AI today. In recognising the capacities and limitations of AI compared to visions of how this might develop in the future, AI can be classified under three broad categories:

All AI in existence now is Narrow AI.

Core AI technologies

Generative AI and Agentic AI

Over the past few years, the vast majority of public conversation and media coverage on AI has related to generative AI, and more recently agentic AI. So very often when people say “AI” this is what they are referring to. Thanks to ChatGPT and subsequent GPT powered large language models like Claude, Gemini, etc., generative AI tends to be what people think of when they use the term AI.

Generative AI and agentic AI are not new. They are advanced architectures that combine and build on the core AI technologies listed above.

Generative AI

AI that creates new content such as text, images, music, or code.

Use Cases:

• Large Language Model (LLM) chatbots (e.g., ChatGPT, Claude)
• Image, audio and video generation (e.g., Midjourney, DALL·E, Suno)
• Code generation (Github Copilot)
• “Synthetic data” for training ML models. This is designed to replicate real data without containing real information (such as personal details about people or commercially sensitive information) so that models can be trained in a safe way.
• Personalised marketing content

Generative AI models are trained on very large datasets, often “scraped” from the internet. Data scraping is when data is automatically taken from websites using software that can collect large amounts of data. From this data generative AI models detect patterns and associations so that they are then able to predict what output would be an appropriate and convincing response to any given prompt. In this way they can create new content, but the new content is always a reflection of the data that the model has been trained on.

How large language models work

Let’s dive a bit deeper into how large language models (LLMs) like ChatGPT work as that is probably the tool that people are most likely to come across. LLMs utilise deep learning architectures to generate text by predicting the next word based on statistical patterns found in vast datasets, rather than through genuine comprehension.

It is important to emphasise that this process involves no real understanding of the world; the model does not "know" anything in the human sense. For instance, while it can state that a dog is an animal, it possesses no internal concept of what a dog, or any other living creature, is.  

This statistical nature makes LLMs highly susceptible to outliers; if you present a familiar riddle but alter the wording slightly, the model often fails because it recognises the pattern of the words rather than the underlying logic of the puzzle. It cannot work through the new variation as a human would. In predicting which outputs are statistically most likely to be convincing, LLMs are also prone to produce inaccurate or entirely false statements but to do so in an authoritative way. They are trained to give answers even when the answers are unknown.

While LLMs can simulate back and forth conversations, it is important to remember that they do not understand the words that are produced and they are not able to feel or care. While they have been designed to imitate human communication and even in some cases to simulate empathy, they do not truly possess these capacities: LLM chatbots are not your friend, your therapist or your life companion. They are merely a sophisticated tool for pattern matching that simulates conversation without empathy, intent, or true cognitive ability. In some cases, such as AI companions, they are designed to say whatever is predicted to keep people engaged, regardless of whether it is true or beneficial. This is because the models are collecting data through those interactions which may be used for commercial purposes (such as marketing, research or to train future AI models).

Agentic AI

AI that acts autonomously toward goals with limited input or direction - decides, plans, adapts.

Use Cases:

• Autonomous software agents (e.g., AutoGPT, Devin)
• Virtual personal assistants with memory and reasoning
• Game-playing agents (e.g., AlphaGo, OpenAI Five)
• Smart home automation (multi-step decisions)
• Robotic systems operating with minimal human control

A key development with agentic AI is that rather than simply producing recommendations or predictions, or classifying information to inform human decisions or actions, agentic AI models can take their own actions.

Agentic AI models are goal-oriented. Rather than being programmed to follow specific instructions or complete particular tasks, they are able to operate with a high level of autonomy to determine a series of actions to take in order to achieve the overall goal set. They use multi-step reasoning to break down problems into smaller tasks, and are able to adapt their processes in response to data, feedback or contextual changes.

Current state of play (June 2026)

Generative and Agentic AI systems are frequently positioned as foundational general-purpose technologies capable of addressing a wide range of use cases. Large generative AI models have typically been designed and developed to be “general purpose”, that is they are intended to be able to be applied in a wide range of contexts and for a variety of purposes, but they are not developed specifically for any particular purpose or context. This can create particular risks when these models are then used in contexts where the potential impacts or limitations have not been considered.

This positioning of generative and agentic AI as general purpose is reinforced by significant marketing investment, which has generated considerable hype around the sector. This is often used to suggest that organisations need to embrace AI or risk being left behind.

Narratives centred on inevitability and framed by the fear of missing out (FOMO) can create uncertainty within organisations. The prevalence of polarised viewpoints, ranging from technological utopias to dystopias, often obscures a clear understanding of the technology’s actual capabilities, limitations, and potential impacts. While AI features prominently in mainstream discourse, it is essential to distinguish between marketing claims and operational reality.

Governments, businesses and organisations of all types are navigating these narratives. While there is pressure to accelerate adoption to realise potential efficiencies, a critical review of current deployments reveals a mixed landscape of benefits and challenges, including instances of failed implementation and unproven returns on investment. The “Gen AI paradox” refers to the disconnect between widespread adoption of AI tools and the lack of tangible returns in terms of financial or productivity impacts. This is underscored by McKinsey’s QuantumBlack research: while nearly 90% of organisations studied use AI, 8 in 10 reported no return on investment. Gartner, meanwhile, finds that around half of generative AI projects are abandoned due to poor data quality, inadequate risk controls, escalating costs or a lack of clarity around value to the organisation.

Organisations often find themselves balancing the pressure to adopt against the need for due diligence.

The question for leadership is not merely whether to adopt, but what specific value is being pursued. If “being left behind” implies missing the opportunity to deploy unproven systems that prioritise speed over safety, then a cautious approach may be prudent. It is vital to consider whether automation streamlines processes without eroding trust or compromising organisational values.

As Rachel Coldicutt of Careful Industries notes, “FOMO is not a strategy.” Deploying AI solely due to external pressure or vendor recommendation, without a clear strategic justification, is inadvisable. Rushing to implement technology without a defined problem statement is not effective leadership.

Too many organisations deploy technology before establishing policies and strategies. This approach is inefficient and increases risk. A technology journey should begin with a people and problem-first approach. Starting with the technology often leads to “techno-solutionism” - the belief that technology can solve all problems. Instead, organisations should first identify the specific problem they aim to solve and then determine whether AI is the appropriate solution. In many cases, it may not be.

The Inevitability Narrative

The “inevitability narrative” suggests that AI is an unstoppable force that must be accepted rather than a technology driven by specific corporate and investment choices. This is evident in frequent statements suggesting that “the genie is out of the bottle” or that organisations must embrace AI or “be left behind”. This perspective is problematic as it implies that questioning or regulating AI is futile. Crucially, this narrative strips individuals and organisations of power and agency, framing resistance as pointless and discouraging active participation in shaping how technology is used. It also overlooks that AI encompasses many different technologies and approaches, and that choices are being made about which AI to develop and use for which purposes. By framing adoption as unavoidable, it also obscures the human choices involved regarding labour practices, energy consumption, and social impact. It is important to recognise that technology adoption involves social, political, economic, and environmental choices, all of which can be influenced and changed.

Strategic Approach

While AI adoption may not be inevitable, it is increasingly difficult to avoid engaging with AI. Organisations must avoid both hasty implementation without clear objectives and total disengagement. Even if organisations do not adopt AI, there are many ways that they are likely to end up interacting with AI through other service or processes, even without knowing it. Organisations also need to understand the risk implications, as many external risks still exist even when organisations opt out of adopting AI internally.

The first step is to ensure staff possess critical AI literacy. This goes beyond basic operational skills; it involves understanding the social, economic, political, and environmental dimensions of the technology. Critical literacy equips teams both with understandings of what is technically possible with AI, and of the limitations and risks of AI. In doing so it empowers teams to question and challenge implementations rather than simply accept them. This is particularly crucial at the leadership level, where accountability for technology deployment resides. Decisions must be informed and intentional.

AI is not a panacea. It is acceptable to determine that AI is not the solution for a specific problem. Furthermore, even if the technology could theoretically solve the problem, it is equally valid to conclude that the organisation’s current infrastructure, governance, or cultural readiness is not yet mature enough to support its safe and effective deployment. In such cases, postponing adoption until the necessary foundations are in place is a responsible strategic choice.

The pressure to accelerate AI adoption must never supersede an organisation’s core ethical values. Rapid implementation driven by market hype can inadvertently compromise principles regarding data privacy, fairness, sustainability and accountability, particularly if safeguards are treated as an afterthought. Ethical deployment requires a deliberate assessment of whether a specific AI application aligns with your organisational mission and social responsibilities. If a technology offers efficiency at the cost of trust or transparency or environmental impact, it represents a strategic misalignment rather than progress. Ultimately, maintaining integrity means having the confidence to delay or reject adoption until it can be executed in a manner that fully upholds your established values.

Once literacy and ethical alignment are established, the focus should shift to problem identification. Define the problem clearly, evaluate potential solutions, and determine if AI is the best fit. If an AI solution is selected, the next step is to align its use with organisational values and policies. This includes:

• Developing a responsible and appropriate use policy
• Establishing safeguards, accountability measures, and processes for redress
• Ensuring open communication between leadership and teams, with mechanisms for feedback
• Assessing data privacy and security. Consider whether the budget allows for secure, local deployments where data does not leave organisational servers. If not, clear guidance on data protection must be enforced to prevent the input of sensitive information into public tools.

Once this due diligence is complete, the organisation may be ready to deploy. While this approach may be slower than rapid adoption, responsibility and due diligence should always be prioritised over speed.