The word "AI" gets used as if it describes one single thing. In reality, it is a broad label covering several quite different kinds of technology. Some are actively chosen by the person using them. Others run quietly in the background of apps people already use every day without a second thought.
This creates confusion. Someone might assume that because a chatbot can write a convincing essay, the same underlying technology should be able to predict the stock market, recognise a face in a photo, or forecast tomorrow's weather.
The reality sits somewhere in the middle.
These are different tools built for different jobs, even though they all get filed under the same umbrella term.
The scale is easy to underestimate. OpenAI reported that ChatGPT alone reached 800 million weekly users by October 2025, up from 500 million in March that year (TechCrunch). That figure comes direct from the company rather than an independent source, so treat it as a rough guide. Either way, that is one type of AI. Most of those people are also using four or five others without thinking of them as AI at all.
This is where understanding the main categories of AI becomes useful. Knowing which type is doing the work behind an app or feature makes it much easier to judge what that tool is actually good for, and where it is likely to fall short.
This guide explains the five main types of AI a beginner is likely to encounter: generative AI, recommendation systems, computer vision, voice and speech AI, and predictive AI.
Along with where each one shows up day-to-day and what the evidence says about how well each one works.
1. Generative AI: the kind that creates things
Generative AI is AI that creates new content, including text, images, audio, or video, based on a description or request. It is the type most people mean when they talk about AI today. AI chatbots like ChatGPT and Claude fall into this category, as do image tools like Midjourney or DALL-E.
What makes it different
What sets generative AI apart from other types is that it produces something new each time, rather than simply sorting or retrieving existing information. It has learned patterns from huge amounts of existing content, and it uses those patterns to generate fresh output that fits what was asked for.
This does not mean the output is always accurate. It means the system is built to generate plausible new content, not to check facts before producing it. That distinction matters more than most people realise when they first start using these tools. This is covered in more depth in Can AI Be Wrong?.
How quickly it has grown
The growth of generative AI in the UK has been steep. Ofcom's Online Nation 2025 report found that ChatGPT alone had 1.8 billion UK visits in the first eight months of 2025, up from 368 million in the same period of 2024.
Generative AI is also changing how people search. The same Ofcom report found that around 30% of searches now show AI-generated overviews at the top of the results page, and 53% of UK adults say they see these summaries often. Many people are reading generative AI output daily without ever opening a chatbot.
Where it shows up
Examples include:
- AI chatbots used for answering questions or drafting text
- AI image generators used for artwork, illustrations, or marketing visuals
- AI writing assistants built into word processors and email clients
- AI summaries appearing at the top of search results
- AI video and music generation tools
- code-writing assistants used by developers
Why it gets the most attention
Generative AI is the most visible and conversational form of AI, which is part of why it dominates the headlines. It feels different to interact with because it responds in natural language and produces something that looks original.
That visibility is also why many beginners assume "AI" and "generative AI" are the same thing, when generative AI is really just one branch of a much bigger tree.
2. Recommendation systems: the kind that predicts what people want
Recommendation systems are AI that analyses a person's past behaviour, along with the behaviour of similar users, to predict what someone might want next. They are one of the oldest and most widely used forms of AI, and most people interact with them constantly without thinking of them as AI at all.
How it actually works
This is the technology behind a streaming service suggesting the next show to watch, a shopping site recommending products, or a music app building a playlist based on listening history. It does not generate anything new. Instead, it predicts and ranks from a pool of existing options.
The scale of this influence is easy to underestimate. Netflix's own researchers have stated that around 80% of the hours streamed on the platform come from its recommendation system rather than people searching for something themselves. In other words, four out of five things watched on Netflix were suggested by AI, not chosen from scratch.
That is why recommendation systems can feel eerily accurate. They are not reading minds. They are comparing patterns across millions of other users with similar habits and surfacing what tends to follow.
Where it shows up
Examples include:
- streaming service suggestions for shows and films
- online shopping recommendations
- social media feeds ranking what appears first
- music and video playlists built from listening or viewing history
- "customers also bought" sections on retail sites
3. Computer vision: the kind that recognises images
Computer vision is AI trained to interpret visual information, including identifying objects, faces, text, or scenes within images or video. It does not generate images. It analyses and recognises them.
How it actually works
A computer vision system is trained on huge numbers of labelled images until it learns to recognise patterns, such as the shapes and features that make up a human face or a street sign. Once trained, it can apply that recognition to new images it has never seen before.
This is what allows a phone to unlock with facial recognition, a photo app to automatically tag people in pictures, or a banking app to scan a cheque. It is also behind more advanced uses such as medical image analysis, where systems are trained to flag potential abnormalities for a doctor to review.
How accurate is it really?
Accuracy varies greatly between systems and between the people being recognised.
The most thorough independent study to date, run by the US National Institute of Standards and Technology, tested 189 facial recognition algorithms from 99 developers. It found that most algorithms showed higher error rates for women and for people with darker skin tones, though the most accurate algorithms showed much smaller differences.
The practical takeaway for beginners is that computer vision is not one fixed level of quality. The face unlock on a modern phone and a cheap security camera's person detection are both "computer vision", but they can perform very differently.
Where it shows up
Examples include:
- phone face unlock
- photo organisation apps that group pictures by person
- document and cheque scanning apps
- automatic photo tagging on social media
- quality control cameras on factory production lines
4. Voice and speech AI: the kind that understands speech
Voice and speech AI covers systems that convert speech to text, text to speech, or recognise spoken commands. Voice assistants like Siri or Alexa rely heavily on this type, as does any app with a dictation or voice-command feature.
How it differs from generative AI
Understanding what was said is a distinct skill from generating a new, fluent response. Many modern voice assistants now combine both abilities, which is why they can sound conversational, but the listening part and the responding part are technically separate jobs handled by different systems working together.
A simple way to picture it:
One system is doing the listening and transcribing.
A second system, often generative, decides what to say back.
How well it copes with different voices
Speech recognition works noticeably better for some voices than others. A Stanford University study published in PNAS tested speech recognition systems from Amazon, Apple, Google, IBM, and Microsoft. On average, the systems made errors on roughly 19 words in every 100 spoken by white speakers, but 35 in every 100 spoken by black speakers.
Accents, background noise, and unusual phrasing all push error rates higher too. This does not mean voice AI is unusable. It means the polished performance seen in adverts is a best-case scenario, and real-world accuracy depends heavily on whose voice is being heard.
Where it shows up
Examples include:
- voice assistants such as Siri, Alexa, and Google Assistant
- dictation features in phones and word processors
- automatic captions on videos and live streams
- voice-to-text messaging features
- call centre systems that route calls based on spoken requests
5. Predictive AI: the kind that forecasts outcomes
Predictive AI analyses patterns in data to estimate what is likely to happen next. It is widely used behind the scenes in finance, logistics, and healthcare, but beginners encounter simpler versions of it regularly too.
How it actually works
This is the technology behind a phone keyboard predicting the next word, a maps app estimating arrival time based on traffic patterns, or a weather app forecasting tomorrow's conditions. In every case, the system is looking at historical data and current signals, then producing a probability-based estimate rather than a certainty.
That is why predictive AI is described in terms of likelihood rather than fact. A weather forecast saying there is a high chance of rain is not a guarantee. It is the system's best estimate based on the patterns it has seen before.
The quiet workhorse example: spam filtering
One of the clearest examples of predictive AI working well is the spam filter. Google states that Gmail's AI-powered defences block more than 99.9% of spam, phishing, and malware, stopping nearly 15 billion unwanted emails every day.
Almost nobody thinks of their spam folder as AI. Yet it is one of the most reliable, longest-running AI systems most people use, and it has been quietly making predictions about every incoming email for years.
Where it shows up
Examples include:
- predictive text on phone keyboards
- traffic and arrival time estimates in maps apps
- weather forecasting
- spam email filtering
- fraud detection systems used by banks
6. What about narrow AI, general AI, and "the four types"?
Anyone who searches for types of AI will quickly run into a different set of labels: narrow AI, artificial general intelligence (AGI), superintelligence, reactive machines, and limited memory AI. These are the textbook categories, and they answer a different question.
The textbook categories in plain language
The academic classification sorts AI by how capable it is, not by what job it does. Narrow AI means any system built for a specific task, which covers every AI tool that exists today, from ChatGPT to a spam filter. Artificial general intelligence means a hypothetical system that could learn any intellectual task a human can, and superintelligence means a hypothetical system beyond that.
The "four types" often mentioned (reactive machines, limited memory, theory of mind, and self-aware AI) describe stages of sophistication. Only the first two exist. The other two are theoretical.
Why this guide uses everyday categories instead
The textbook labels are accurate, but they are not very useful for deciding whether to trust a chatbot's answer or understanding why a streaming app knows what someone wants to watch. Telling a beginner that both ChatGPT and a spam filter are "narrow AI" is true, and also explains almost nothing about how differently they behave.
The five categories in this guide sort AI by the job it does, because that is what actually helps day to day. For a deeper look at how these systems learn in the first place, see AI vs Machine Learning.
7. Why the distinction actually matters
Understanding these categories helps explain why an AI tool that excels at one task can be useless at another. A generative chatbot is not built to reliably predict stock prices, and a recommendation system is not built to write an essay. Each type is trained for a different kind of task, using different methods and different data.
This also explains a common point of confusion. When people say "AI is everywhere now," much of what they mean is generative AI specifically becoming visible and conversational, while the other types have quietly existed in the background of everyday apps for years. Recommendation systems, predictive text, and spam filters were doing their jobs long before chatbots became mainstream.
They simply never announced themselves as AI.
Knowing the category also helps set realistic expectations. A chatbot that gets a fact wrong is not malfunctioning; it is doing exactly what generative AI is built to do: produce plausible text, not verified truth.
A recommendation engine that suggests something irrelevant is not broken either, it is just working from limited data about that particular person's habits.
There is a trust gap here worth noticing. Even as usage climbs, Ofcom's 2024 Online Nation report found that only 18% of UK adults said they trust AI-generated content. That instinct is broadly healthy. The trick is knowing which type of AI is involved, because "should this be double-checked?" has a different answer for a chatbot essay than it does for a spam filter.
8. Limitations to keep in mind
None of these AI types are infallible, and each comes with its own blind spots.
- generative AI can sound confident while being factually wrong, so any claim, statistic, or fact it produces still needs checking
- recommendation systems only know what they have observed, so they can get stuck recommending the same narrow type of content repeatedly
- computer vision can misidentify objects or faces, particularly with poor lighting, unusual angles, or in people from groups underrepresented in its training data, as the NIST testing above showed
- voice and speech AI can mishear accents, background noise, or unusual phrasing, and the Stanford research above found error rates nearly doubling for some groups of speakers
- predictive AI is only as good as the data it is trained on, and it can be confidently wrong when real-world conditions shift away from historical patterns
The safest approach is to treat any AI output, regardless of type, as a starting point rather than a finished answer. AI works best as a helper, not the final authority, and decisions that matter, such as medical, financial, or legal ones, should not rely only on what an AI system produces.
Final thoughts
AI is not one technology. It is a collection of different approaches, each suited to different problems. Generative AI creates, recommendation systems predict preferences, computer vision recognises images, voice AI handles speech, and predictive AI forecasts outcomes from data.
The evidence backs up how differently they perform. A spam filter catching 99.9% of junk mail and a chatbot confidently inventing a statistic are both AI, doing exactly what they were built to do. The difference is not quality. It is the job.
None of these types are perfect, and all of them benefit from a healthy dose of scepticism. AI can be wrong, sometimes confidently so, and the type of AI involved usually explains why.
Recognising which type is being dealt with makes it much easier to judge what a given AI tool is actually good for, rather than expecting any single tool to do everything well.
A chatbot will not reliably forecast the weather, and a weather app will not write a cover letter. Matching the tool to the task is the practical skill worth taking away from all of this. For a fuller picture of where each type earns its place, see What AI Is Actually Useful For.
Frequently asked questions
Is ChatGPT the only type of AI worth knowing about? No. ChatGPT is an example of generative AI, which is currently the most visible category, but it is only one of several types in everyday use. Recommendation systems, computer vision, voice AI, and predictive AI all run constantly in the background of apps that have nothing to do with chatbots.
What is the difference between generative AI and predictive AI? Generative AI creates new content, such as text or images, based on patterns it has learned. Predictive AI estimates what is likely to happen next based on historical data, such as a weather forecast or an arrival time. One produces something new. The other produces a probability. A chatbot is generative; a spam filter is predictive.
What are the four types of AI mentioned in textbooks? The academic classification lists reactive machines, limited memory AI, theory of mind AI, and self-aware AI. Only the first two exist today. Every current AI tool, including chatbots, also falls under the broader label of narrow AI, meaning it is built for specific tasks rather than general human-level thinking.
Why does a chatbot sometimes get basic facts wrong? Generative AI is built to produce plausible, fluent text based on learned patterns, not to verify facts against a database. It can sound confident even when it is incorrect, which is why its answers should still be checked, especially for anything important.
Are recommendation systems and generative AI the same technology? No. They are built differently and trained for different jobs. Recommendation systems rank and predict from existing options based on behaviour patterns. Generative AI creates new content from scratch. A streaming app's "recommended for you" list and a chatbot's reply are produced by entirely different methods.
Can computer vision be tricked or get things wrong? Yes. Computer vision systems can misidentify faces or objects, particularly in poor lighting or with unusual image angles. Accuracy also varies depending on how diverse and representative the training data was. Do voice assistants understand language the way a person does? Not really. Voice assistants convert speech to text and match it against likely intents, often combining that with a generative system to produce a natural-sounding reply. They are pattern-matching at speed, not understanding meaning the way a person would.
Why does predictive text sometimes suggest strange words? Predictive AI is estimating the most likely next word based on patterns from previous typing and common language use. It is a probability guess, not a guarantee, so it occasionally suggests something that fits the statistical pattern but not the actual intended meaning.
Is one type of AI "better" than the others? Not in a general sense. Each type is suited to a different job. A generative chatbot is excellent at drafting text but poor at forecasting outcomes. A predictive system is excellent at estimating arrival times but cannot hold a conversation. Comparing them head to head misses the point, since they are not interchangeable.
Will these categories stay the same as AI develops further? The boundaries are already starting to blur, with some tools combining several types together, such as a voice assistant that listens, transcribes, and then generates a reply. The underlying jobs, creating, predicting, recognising, and understanding speech, are likely to remain useful categories even as the tools that combine them become more sophisticated.
