Find on this page

Searching for information with AI

You can use many generative AI tools when you search for information. It is important to understand how these tools work. You also need to know their advantages and limitations, and how to use them in an ethical and safe way.

What is artificial intelligence?

Artificial intelligence, often called AI, is a technology that allows computers to perform tasks that usually require human intelligence. These tasks can include understanding language, recognizing images, and making decisions. There are three main types of artificial intelligence:

Artificial Narrow Intelligence – ANI
Artificial narrow intelligence is the type of AI that is used today. It is sometimes called weak AI. This type of AI is created to perform one specific task within a limited area. ANI can be as good as, or better than, humans in that area, but does not have a general understanding. It cannot learn new tasks outside the area it was trained for.
Artificial General Intelligence – AGI
Artificial general intelligence is sometimes called strong AI. It is a theoretical form of AI that does not yet exist in practice. AGI would be able to perform intellectual tasks at the same level as a human. It would be able to reason, understand context, learn new things, transfer knowledge between different areas, and make its own decisions.
Artificial Superintelligence – ASI
Artificial superintelligence is a hypothetical form of AI that would be more intelligent than humans in all areas. This could include scientific creativity, social abilities and problem solving. ASI does not exist today. It is discussed in research because it raises important questions about ethics, safety and possible effects on society.

How is AI used?

Artificial narrow intelligence is used in many areas of everyday life. Here are a few examples:

Knowledge-based systems that support troubleshooting
To plan routes in navigation apps
Robots in industrial environments
Image recognition and facial recognition
AI tools that generate text, images, or code

Generative AI

When you search for information using AI, generative AI is often used. It is based on a technology called machine learning. This means that the computer learns from data instead of being programmed for each task.

The purpose is to allow the computer to handle complex tasks automatically. This is done by using algorithms and statistical models. These models identify patterns in large amounts of data. When a model is trained with more data, the results usually improve over time.

Deep learning and neural networks

Deep learning is a more advanced form of machine learning. It uses deep neural networks to process information. A neural network is a mathematical model inspired by how the human brain processes information. It is built from several layers. Each layer contains nodes (or neurons) that handle information step by step. Neural networks can be used for many different tasks. When they are used to create new content, such as text or images, it is called generative AI.

Large language models – LLM

There are different types of generative AI. One example is large language models, often called LLMs. Large language models are a type of neural network based on a technology known as transformer architecture. These models are used to understand and generate text that is similar to human language. They are commonly used in chatbots and AI-based search tools.

Examples of large language models are:

GPT
GPT was developed by OpenAI. It is used in services such as ChatGPT and Microsoft Copilot. The model can answer complex questions, reason about problems, write code and generate text.
Claude
Claude was developed by Anthropic. It is used in services such as Claude.ai and Notion AI. Claude can write and summarise content. It is especially good at producing longer texts, summaries and well-structured responses.
Gemini
Gemini was developed by Google DeepMind. It is used in Google AI services and in Google Workspace tools such as Gmail and Google Docs. The model can handle different types of information at the same time, including text, images and video.

Important to know about large language models – LLM

Large language models are trained on very large amounts of text, but the training material is not publicly available. This means it is not possible to know which sources a specific response is based on. Different AI tools can give different answers, because the models are trained on different data.

Large language models do not know what is true or false. They generate responses based on patterns and probabilities in language. Because of this, they can sometimes produce incorrect information. biased results, or fabricated responses. These errors are known as hallucinations.

Bias and prejudice

Language models are trained on material created by humans. Because of this, there is a risk that the models reflect existing prejudices, inequalities, or bias. If some groups are underrepresented in the training data, the responses may be discriminatory. This can affect areas such as gender, ethnicity, or culture. In this way, AI can reinforce prejudices that already exist in society. The risk becomes greater when models are trained on AI-generated content that already contains errors or bias. This can lead to a cycle where incorrect or biased information continues to spread.

AI – how does it work?
Karolinska Institutet University Library explains what generative AI is and how large language models work.
What is generative AI?
Digiteket provides an introduction to large language models, machine learning, transformer architecture and neural networks. Available in Swedish.

AI search tools

AI tools can be used for information retrieval, but they work in different ways and are not suitable for every situation. In this section, you will learn about the differences between chatbots, AI search tools and AI functions in databases. You will also learn what to consider in order to use them safely and responsibly.

Chatbots

ChatGPT and Microsoft Copilot are examples of chatbots based on the GPT language model. These chatbots generate responses based on the text you provide in your question. This input is often called a prompt. The same prompt can produce different responses at different times. The data used to train the models is not public. Because of this, it is not possible to know which sources a specific response is based on.

When are chatbots useful?

Chatbots are most useful at the beginning of a search process. This is when your work is exploratory and you are testing ideas. For example, you can use them to:

Get started with a task
Test ideas or get inspiration
Generate ideas for reserach questions
Translate words from Swedish to English
Identify keywords and synonyms

Building search strings with chatbots

You can also ask a chatbot to suggest search strings for use in different databases. However, these search strings often contain errors. Important search techniques, such as phrase search or truncation, may be missing. The quality of the search string also depends on how clearly you formulate your prompt.

If you need to find scientific sources, chatbots are not suitable. The sources behind the responses are not visible, and it is not possible to know where the information comes from. Even when you ask a chatbot to provide sources, it may generate references that do not exist. Because of this, it is important to always verify information using other reliable sources.

AI search tools for scientific material

If you need to find scientific material, chatbots are not suitable. Instead, you can use AI search tools that are designed to retrieve scientific sources. These tools can search for real publications based on a prompt. They provide answers that are based on scientific articles and usually show which sources have been used.

There are also limitations to be aware of. These tools often do not have access to the full text of articles. The answers are therefore based on summaries or metadata. This can mean that important details are missing. The quality of the journals included can vary, and it is not always clear why certain sources are selected over others. This can result in incomplete or distorted answers.

How do AI search tools differ from chatbots?

AI search tools for scientific material often use a technology called Retrieval Augmented Generation, usually shortened to RAG. RAG combines two functions. One function is a language model that generates text. The other function is a search in databases or search engines. The generated answer is based on the information that is retrieved in the search.

The purpose of RAG is to provide more fact-based answers. It also helps reduce the risk of incorrect information or hallusinations. Another advantage is that the tool can use information that the language model was not originally trained on. It can also answer questions about specific documents.

Elicit and Perplexity are examples of AI tools that use RAG. These tools can be connected to general search engines such as Google, or to research-focused services such as Semantic Scholar. There are also other AI tools that work in different ways.

Agent-based and snowball search AI tools

Agent-based AI tools work by handling a task step by step. When you ask a question, the tool may first rephrase it. It can then create its own prompts to search for relevant references. The exact process differs between tools.

Snowball search AI tools are not generative. Instead, they find articles based on studies you already have. This can be done by analyzing reference lists, citations, or similar content. This type of search should not be confused with snowball searches in databases such as Web of Science and Scopus, which are based on verified citation links.

Examples of scientific AI search tools
Karolinska Institutet University Library has compiled a list of AI tools that focus on scientific material.

AI functions in databases

Many databases include functions that use AI to support searching. When you use the Library Search Tool, called EBSCO Discovery Service, you can use a function called EDS Natural Language. This function allows you to write your question in a natural way, similar to how you would ask a person. You do not need to choose specific search terms or combine them with Boolean operators.

An example of a question is:
- Which articles deal with AI in healthcare?

When a database uses natural language searching, it treats your question as a starting point. The system then creates a search string automatically. You can often view how the search string was created. This makes it possible to copy it, save it, or reuse it in another database.

There are also limitations to be aware of. Automatically generated search strings often miss important search techniques. These can include phrase searching or truncation. The search can also be very broad, which may result in many results that are not relevant. Similar functions are available in databases such as CINAHL, APA PsycInfo and ERIC. Many databases also include other AI-based functions. Some of these features require a special subscription.

Library databases or AI search tools?

Searching in library databases can feel more demanding than using AI search tools. In a traditional database, you usually need to search in a structured and systematic way. This means creating search strings using Boolean operators, search techniques, and controlled subject terms. In AI search tools, you can often write your search in everyday language. This is usually called natural language searching. Because the tools work in different ways, it is important to understand the difference between searching in a scientific database and using AI search tools.

Lexical and semantic search – what is the difference?

When you search in a traditional database without using AI functions, you are usually doing a lexical search. This means that the database searches for exact words or phrases. The search is often limited to fields such as titles, abstracts, or subject terms.

AI search tools usually use semantic search. This means the search is based on the meaning of your question. The tool searches for content that is similar in meaning, even when different words are used.

AI search tools can be useful when the search is exploratory and investigative. They can also be used as a complement to a structured search in a traditional database. When a search needs to be fully transparent and possible to repeat, AI search tools are not suitable. The search process cannot be reviewed or recreated in the same way as a traditional database search.

AI and information retrieval
Learn more about the difference between lexical and semantic search on the KTH Library website.
Difference between traditional databases and AI search tools
Table comparing library databases with AI search tools. From Karolinska Institutet University Library.

When can I use AI search tools?

The purpose of your search determines whether AI search tools are suitable. When your search is exploratory or experimental, AI tools can be useful support. This can be the case when you want to get started with a topic, test ideas, or identify possible search terms.

AI search tools should not be used for systematic and exhaustive searches. These types of searches need to be structured, transparent, and possible to repeat. When you use AI search tools, it is often not possible to review or recreate the search process. There are several reasons for this.

• The tool generates responses based on statistical probability
• The same question can produce different answers at different times
• The suggested sources can vary even when the question is the same
• It is not always clear why certain sources are selected over others

Even when AI search tools display scientific sources, there are risks to consider. The tools may suggest articles from journals with low quality or unclear review processes. AI tools also do not understand content in the same way humans do. When an AI tool summarises a longer text, important information may be missing or distorted.

Prompt correctly

The instructions you write to an AI tool are called prompts. The way you write a prompt affects the response you get. Therefore, it is important to think carefully about your wording. A good place to start is to give the tool a clear context. Explain who you are, who the material is for and how the material will be used. This often makes the response more relevant and useful.

After you write your first prompt, it helps to work step by step. You can adjust the prompt after each response to make the instruction clearer. This process is called iteration, which means that you improve the prompt gradually in order to get a better result. Avoid asking leading questions and try to be as neutral as possible. Otherwise, you may influence the answer in a way that makes it one-sided or misleading. This is sometimes called bias.

Prompt like a pro – how to improve your AI instructions
Examples of how you can formulate better prompts when using AI. Created by Internetkunskap from Internetstiftelsen. Available in Swedish.
Prompt tips for AI search tools
Karolinska Institutet University Library provides examples of how to write clear and effective prompts in AI search tools.
What is prompting?
Digiteket explains how prompting works and what you need to consider when writing instructions for AI tools. Available in Swedish.

Ethical issues surrounding AI

Using AI involves personal responsibility. Here you will learn more about ethical issues related to copyright, privacy, data security, and the environment. You will also learn how to make informed choices when using AI.

Things to consider when choosing an AI search tool

There are many ethical aspects to consider when using AI. Therefore, it is important to understand how a tool works and where it gets its information. Start by finding out who developed the tool and what type of data it is based on. It is also important to be aware that all tools have limitations. Consider in which situations it is appropriate to use a specific tool.

Different AI tools work in different ways. Because of this, you need to know what type of prompts work best in the tool you are using. You should also check how the tool handles your data, and make sure you understand whether the information you share is stored and protected.

Personal privacy and data security

When you use AI tools, you have a responsibility to use them safely. This is important for both you and others. You should assume that any information you share may be stored, shared, or used to train and develop AI models. Therefore, you should never share sensitive information.

Sensitive information can include social security numbers, addresses, passwords, payment details, or confidential material. It can also include audio recordings or transcripts from meetings.

In addition to the information you actively provide, AI tools may collect other types of data. This can include your IP address, email address and location data.

The Swedish Authority for Privacy Protection – IMY
IMY describes the risks to personal privacy when AI is used. Available in Swedish.
Use generative AI in an ethical manner
Digg, the Swedish Agency for Digital Government, provides information on how you can use generative AI in a responsible and ethical manner. Available in Swedish.
Ethical aspects
Who is behind AI tools, and are there other ethical aspects to consider? Karolinska Institutet University Library clarifies the concepts.
What does AI do with your data?
Internetkunskap from Internetstiftelsen explains what happens to your information when you use AI tools. Available in Swedish.

Copyright

The use and development of AI raise many questions about copyright. The technology is still developing, and the rules are not always clear. Therefore, you need to be careful before sharing material with an AI tool. If the material is copyrighted, sharing it may break copyright rules. This applies even if the purpose is analysis or summarisation.

There are still many open questions about the use of scientific articles and research material. Some publishers do not allow articles that require a subscription to be uploaded to AI tools. Other publishers may allow this, but with restrictions on how the material may be used.

Even Open Access articles with a Creative Commons license can have specific conditions. These conditions may limit how the material may be shared or reused. If you are unsure, it is best to contact the copyright holder or the publisher to ask what is allowed.

Assess copyright when using generative AI
Digg, the Swedish Agency for Digital Government, describes how intellectual property rights and copyright are affected when generative AI is used. Available in Swedish.
Creative Commons on AI and CC licenses
Creative Commons explains how their licenses may be used in connection with AI and language model training.

Environmental aspects

Training and using AI models requires large amounts of energy. This leads to emissions that affect the climate. The servers that run AI systems also need cooling, which often requires large amounts of water. The exact level of emissions varies and is debated. However, the overall environmental impact is significant. Therefore, it is important to consider when AI is necessary and when it can be avoided.

At the same time, research is focused on making AI more sustainable. This includes developing algorithms that use less energy. There is also research that shows how AI can support solutions that improve environmental sustainability in the long term.