AI glossary

AI terms explained in plain language.

Search…

Agent: An AI that doesn't just answer questions but carries out several steps on its own to complete a task, such as searching, filling in forms or using other programs. Agents should be supervised, as they can make mistakes.
AI Act: The EU regulation on artificial intelligence, which classifies AI systems by risk and sets requirements for them. It bans certain harmful uses and requires, among other things, that chatbots and AI-generated content be identifiable as such.
AI assistant: A chatbot or app that helps with everyday and work tasks such as writing, searching and planning. Examples include ChatGPT, Claude, Gemini and Copilot.
AI detector: A tool that tries to judge whether text or an image was made by AI. Detectors are often wrong in both directions, so their result shouldn't be used as the only evidence.
AI literacy: Understanding how AI works, what it can be used for, and what its limits and risks are. The EU AI Act requires organisations that use AI to ensure their staff have sufficient AI literacy.
AI PC: A marketing term for computers with an NPU chip for on-device AI features, such as Microsoft's Copilot+ PCs. You don't need one for ordinary use of ChatGPT or Gemini.
Algorithm: A precise set of instructions a computer follows to solve a problem. Social media algorithms, for example, decide what content you're shown.
API: A technical connection that lets programs use each other's services. With an AI API, a business can plug a language model into its own app or website, for example.
Artificial intelligence (AI): Computer programs that perform tasks that have traditionally needed human thinking, such as understanding language, recognising images or drawing conclusions. AI doesn't think like a person; it recognises patterns in large amounts of data.
Automation: Having software handle repetitive tasks so a person doesn't have to do them by hand each time. Zapier and Make, for example, connect apps and can add AI steps to a workflow.
Avatar: A digital character that represents a person in video or online. AI video tools such as Synthesia and HeyGen create talking avatars that can resemble real people.
Bias: An AI's tendency to favour or disadvantage certain groups or viewpoints because its training data was one-sided. It can show up in images, recommendations or recruitment tools, for example.
Chatbot: A program you can have a conversation with by typing or speaking. Today's chatbots are built on language models and can answer a very wide range of questions.
Citations: Links or references showing where the information in an AI answer comes from. They make checking easier, but open the source yourself, as the AI may have misread it.
Cloud service: A service that runs over the internet on a company's servers, regardless of your own device. Most AI tools are cloud services, so they also work on an older computer through a browser.
Computer vision: AI's ability to recognise and interpret images and video. It lets your phone identify plants, faces or text, and lets apps like Be My Eyes describe surroundings to blind users.
Context window: The amount of text an AI can keep in mind at once, including the conversation and attached files. If a conversation gets very long, the model may forget what was said at the start.
Copyright and AI: The question of who owns AI-made content and whether AI may be trained on other people's works. Content made purely by AI is generally not considered protected by copyright, and the issue is still being tested in court.
Custom instructions: A setting where you write standing preferences that the AI assistant follows in every conversation, such as "always answer briefly". That way you don't have to repeat the same instruction each time.
Deep learning: A form of machine learning that uses neural networks with many layers. It underpins today's language models, image generators and speech recognition.
Deepfake: A fake video, image or audio clip made with AI in which someone appears to say or do something they never did. Deepfakes are used in scams and disinformation, among other things.
Fine-tuning: Extra training of an existing AI model on specific material, so it learns, for example, an industry's vocabulary or a company's style. Ordinary users rarely need it.
Freemium: A pricing model where the basic version is free and you pay for extra features or more usage. Most AI tools work this way, and prices change often.
GDPR: The EU's General Data Protection Regulation, which sets rules on how personal data may be collected and used. It applies to AI services too, so don't paste customer data into an AI without thinking.
Generative AI: AI that creates new content such as text, images, audio, music or video. ChatGPT and image generators are generative AI.
GPT: Short for Generative Pre-trained Transformer. It's the name of OpenAI's language models and also of a common model design; ChatGPT is a service built on these models.
Hallucination: When an AI makes up information that sounds convincing but is wrong, such as non-existent sources or incorrect figures. That's why important facts should always be checked.
Image generator: An AI tool that creates a picture from a text description. Examples include Midjourney, Adobe Firefly, Ideogram and ChatGPT's image feature.
Language model: An AI program trained on huge amounts of text so it can understand and produce language. It estimates which word is most likely to come next and builds its answers that way.
LLM (large language model): A very large language model, such as the models behind ChatGPT, Claude or Gemini.
Machine learning: A branch of AI in which a computer learns from examples instead of having every rule programmed by hand. A spam filter, for instance, learns to spot spam by seeing many example messages.
Machine translation: Translation from one language to another done by a computer, for example with DeepL or Google Translate. It's fast and often good, but official and important texts still need a human check.
Memory (in AI assistants): A feature that lets an assistant remember things about you between conversations, such as your job or preferences. You can usually view and delete what it remembers in the settings.
Model training: The stage in which an AI model learns from a huge dataset. Training requires a lot of computing power and electricity, and afterwards the model's knowledge is frozen at a certain point in time unless it searches the web.
Multimodal: AI that understands and produces several types of content, such as text, images, sound and video. For example, you can show it a picture and ask about it out loud.
Neural network: A computing model loosely inspired by the brain's nerve cells. It consists of connected units that learn to recognise patterns in data.
NPU (neural processing unit): A dedicated chip in a computer or phone designed for AI calculations. It makes on-device AI features faster and saves battery, but browser-based AI services don't need it.
On-device AI: AI that runs directly on your phone or computer without sending data over the internet. It's good for privacy but needs a more powerful device and is often more limited than cloud models.
Open model: An AI model whose maker has released it for anyone to download, so it can run on your own computer or server. Openness varies: often the model is released but not the training data.
Parameter: An internal adjustable number in a model that changes during training. Large language models can have billions of parameters, but the count alone doesn't tell you how good a model is.
Personal data: Any information that can identify a person, such as a name, address, ID number, health details or photo. Don't enter your own or others' sensitive personal data into AI services without good reason.
Prompt: The instruction or question you type or say to an AI. The more clearly you explain what you want and who it's for, the better the answer usually is.
RAG (retrieval-augmented generation): A technique where the AI first looks up information in chosen material, such as a company's guidelines, and builds its answer from that. Answers are then based on current, relevant data, with fewer errors.
Reasoning model: A language model that spends extra time working through an answer step by step before replying. It's often better at maths, logic and complex problems, but slower.
Speech recognition: Technology that turns speech into text. It's used for dictation, automatic captions and meeting notes; dialects and background noise can reduce accuracy.
Text-to-speech (TTS): Technology that reads written text aloud. Today's AI voices sound very natural and are used in audiobooks, screen readers and video voice-overs.
Token: A small chunk of text that a language model processes. A token can be a whole word, part of a word or a punctuation mark; long Finnish words often split into several tokens.
Training data: The material an AI was trained on, such as books, web pages, images and conversations. Its quality and content directly affect what the AI can do and what mistakes it makes.
Voice cloning: AI's ability to imitate a specific person's voice from a short audio sample. Scammers can use it to pose as a relative on the phone, so agree on a code word with your family.
Watermark: A visible or invisible mark showing that an image, video or audio clip was made with AI. Watermarks help identify AI content, but they can sometimes be removed, so they aren't fully reliable.

AI glossary

A

B

C

D

F

G

H

I

L

M

N

O

P

R

S

T

V

W