Skip to main content

ChatGPT, Claude, Gemini and other AI assistants

These services let people use everyday language to ask questions, create material and work with information. They are products built around AI models, not the models themselves, although the models are a large part of what makes them work.

Features, prices and data policies for individual services change regularly. Check the provider's current information before choosing a service or sharing personal material.

What AI assistants are

Most current AI assistants, including ChatGPT, Claude and Gemini, use a large language model (LLM) to interpret a request and generate a response. The website or app around the model adds the conversation screen, safety rules, account settings and features such as searching the internet, uploading files and generating images.

Some assistants are multimodal, meaning they can work with more than one kind of information. For example, you may be able to ask a written question about a photograph or speak instead of typing.

A fluent answer is not proof that an answer is correct. These systems generate responses from patterns and can invent facts, misunderstand a request or miss important context.

How a response is made

The details vary between services, but a typical conversation works roughly like this:

  1. You enter a prompt by typing, speaking or providing a file or image.
  2. The service combines your prompt with relevant parts of the conversation and its own instructions.
  3. The model generates a response a small piece at a time, based on patterns learned during training.
  4. If available and needed, the assistant may use a tool such as web search, a calculator or a connection to another service.
  5. The product displays the result and may save the conversation, depending on your account and settings.

An assistant is not simply looking up a finished answer in a database. It generates a new response, so the same question can produce a different answer another time.

Common ways people use assistants

  • explaining an unfamiliar subject in simpler language
  • brainstorming ideas or questions to investigate
  • drafting, shortening or changing the tone of text
  • summarising material that the person is allowed to share
  • planning the steps in a low-risk task
  • asking questions about an image, document or spreadsheet

These uses still need human judgement. A draft may sound polished while containing incorrect details, and a summary may omit something important.

WarningDo not share passwords, financial details, confidential work or another person's private information unless you know the service is approved for it.

Other assistants

There are many alternatives. Some use their own models, while others combine models made by different companies.

Being able to cite web pages can make an answer easier to check, but citations can still be irrelevant, incomplete or misunderstood.

How assistants differ

Services can look similar while behaving differently. The model is only one part of the product.

Models
A service may offer several models. Free and paid accounts may use different versions, and this can affect the quality of the response.
“Thinking” or reasoning level
Some services let you choose how much processing effort a model can use before answering. A higher setting may help with complex tasks, but it is slower and does not guarantee a correct answer. This is a product label; it does not mean the model thinks like a person.
Current information
Some assistants can search the web or retrieve connected information. Others answer mainly from learned patterns and the material in the conversation.
Tools
Tools can create images, run code, analyse files or connect to calendars and other services. Access may depend on the plan and region.
Memory and personalisation
Some products can remember selected details between conversations. Look for controls to view, change or delete saved information.
Data handling
Providers have different rules for storing conversations, training models and handling business accounts.

Open-source and open-weight models

Not every AI model belongs entirely behind a company's website. Some can be downloaded, inspected, adapted or run by other organisations and individuals.

Open source

Open-source AI aims to give people the freedoms needed to use, study, change and share an AI system. That can involve more than publishing the finished model: licences, code and information about its training also matter.

The Open Source Initiative's Open Source AI Definition explains these requirements. Ai2's OLMo project is an example designed to make model weights, training code and other parts of the development process available.

Open weight

An open-weight model makes its learned numerical parameters available to download. This can allow people to run or adapt it, but the training data, development code or licence may still have restrictions. Open weight and open source do not always mean the same thing.

Model families such as Meta's Llama and some Mistral AI models can be downloaded under their stated licences. Always check the licence for the particular version. The Hugging Face model directory lists models from many organisations, with widely varying licences and documentation.

What to consider

  • Your task: a useful service for writing may not be the best service for current facts, images or private files.
  • Accessibility: check keyboard, screen-reader, voice and text-size support with the features you need.
  • Privacy: understand whether conversations are stored or used to improve the service.
  • Evidence: check important claims using reliable independent sources.
  • Cost and limits: free and paid accounts may use different models or offer different features.