What AI is and is not

“AI” is short for artificial intelligence. It often works in the background of apps and services people use.

AI systems find patterns in data. They use those patterns to make guesses. A guess can be a prediction, new content or a recommendation.

Three things shape how an AI system behaves. They are what it is for, how it was programmed, and the data it learned from.

There is more than one kind of AI. The HEA’s guidance on AI literacy training names four main families: rule-based systems that follow programmed logic, classic machine learning systems that find patterns in data, generative systems that produce new content, and combined workflows that mix these approaches. It introduces them with “including”, so these are the main families it names, not a closed list.

Where the line sits depends on whose definition you use. The EU AI Act defines an AI system in law, and it leaves out programs that only carry out rules written by people. In that law, a key feature of an AI system is that it infers: it works out its outputs, or builds its own models, from inputs or data. The HEA, by contrast, counts rule-based systems as one family of AI.

AI comes in different kinds. One kind makes new text, code or pictures. Another kind predicts what is likely to happen next, as spam filters do.

You may already use AI. It is at work when a streaming or shopping app recommends something. It helps a journey planner use live traffic or transport information.

You may also meet AI when you chat with a customer-service chatbot. It is there when you translate, turn speech into text, or unlock your phone with your face.

Social media “for you” pages are shaped by recommendation systems. These use data about you to pick what you see next.

One kind of AI, called generative AI, creates new content, such as text, code or pictures, based on the material in its training data. It powers tools like chatbots, coding assistants and image generators. Another kind, predictive AI, is trained to find patterns in existing data and forecast what is likely to happen next. Predictive systems sit inside recommendation engines, spam filters and personalised learning apps.

The “for you” page is an example of a recommendation system at work: it uses data about a person to make recommendations or predictions for them.

AI is not a person. It can produce replies that sound like a person wrote them.

AI is not magic, and it does not know everything. It works by applying statistics and logic to data.

A chatbot like ChatGPT does not look up a stored answer. Each time, it works out which words are likely to come next. It does this from patterns in the text it learned from. What it learned is not a store of answers.

Not every chatbot works like this. A rule-based chatbot follows rules that people wrote. Some chatbot apps also search a set of documents. A later lesson shows how to tell.

In an August 2025 online survey of 2,301 US adults, people were asked what a chatbot like ChatGPT is doing when it answers. Almost half (45%) chose “looking up the exact answer in a database”. About 1 in 5 (21%) chose “following a script of prewritten responses”. Just over 1 in 4 (28%) chose “guessing what words should come next based on patterns it learned”. Each person gave one answer, so about 2 in 3 (66%) chose a lookup or a script.

This is one online survey, from the US only. Its publisher filed the question under “Dustbin”, as interesting but not part of its main findings.

A 2026 research paper says that what such a chatbot learned is held inside its model, the part that produces the text, as “a lossy, biased compression of the training data”. When it runs, the model gives the chances of each possible next piece of text. A chatbot product can also add material held outside the model, such as knowledge bases, and feed parts of it in. The paper notes that users mix up generating text with searching.

A reply from AI can sound warm and kind. A caring tone is not proof of caring.

A warm, fluent reply is not evidence that a system has feelings or experiences.

One 2026 research paper separates two things people mean by understanding. The first is using information well across many tasks. The second is the lived understanding a person gets from having a body and a life. It says that the language models behind the chatbots of 2026 that work out likely next words can use information well across many tasks, within limits. They have no body, and they do not keep learning from experience the way a person does.

That leaves two opposite mistakes about such a chatbot. One is treating it as a mind with feelings. The other is calling it “just autocomplete” that understands nothing. Saying that it predicts the next word is accurate; it misleads when it is taken as the whole story.

A 2024 study reported that experts in brain science and consciousness research estimated that the chatbots of the time were very unlikely to be conscious. That estimate is about the chatbots of 2023, not today’s.

In July 2023 (published April 2024), 2 in 3 (67%) of 300 adults surveyed in the US did not rule out that ChatGPT has experiences such as feelings. The other 1 in 3 (33%) said it definitely does not. Each person gave one rating, from 1 (“clearly not an experiencer”) to 100 (“clearly an experiencer”), and the middle rating was 16. The more often people used ChatGPT, the more likely they were to say this.

In a January 2025 US survey of 500 people who use AI chatbots such as ChatGPT, 4 in 10 (40%) said the one they use most acts like it understands them, at least some of the time. That question asks how it seems, not what people believe.

This matters for trust. The researchers behind the 2023 survey say that believing a chatbot is conscious to some degree could raise trust in it, and that too much trust can lead to depending on it emotionally and relying on it for important decisions.

The evidence is thin in two ways: every survey here is from the US, and no survey was found for Ireland or for neurodivergent learners.