AI Glossary: Essential Terms for Science and Research

03/11/2026

Artificial intelligence is already being integrated into many scientific practices: scientific literature monitoring, quantitative and qualitative data analysis, scientific writing, programming, communication, image analysis , and the automation of certain tasks through specialized assistants. Yet the vocabulary used to discuss these technologies often remains unclear.

For researchers and scientists, understanding these concepts is not merely a matter of technical vocabulary. It is a first step toward evaluating tools, identifying relevant use cases, and better understanding their methodological, ethical, and scientific limitations.

That is why we have compiled an AI glossary of the key concepts to know in science and research, organized into 5 sections:

The Different Forms of Artificial Intelligence:

Artificial Intelligence ▪ Analytical AI ▪ Predictive AI ▪ Generative AI ▪ Multimodal AI ▪ Artificial General Intelligence

The Fundamental Building Blocks of AI Systems:

Machine Learning ▪ Deep Learning ▪ NLP ▪ LLM ▪ Token

Interaction Tools:

Chatbot ▪ AI Assistant ▪ AI Agent

Techniques for Adapting and Connecting AI Systems:

Prompt Engineering ▪ RAG ▪ Fine-Tuning ▪ MCP

Key Risks to Be Aware Of:

AI Hallucinations ▪ Algorithmic Bias ▪ Shadow AI

Want to learn more about AI?

Discover our Focus IA section

Our articles are part of our EFFISCIANCE AI support offering, dedicated to the use of AI for science.

Discover all our articles

The Different Forms of Artificial Intelligence

From data analysis to content generation, artificial intelligence covers a wide range of uses. Discover the main forms of AI, what they make possible in research, and the opportunities they open up for the scientific practices of tomorrow.

# Artificial Intelligence

Artificial intelligence, or AI, refers to a set of methods and systems capable of performing tasks that typically draw on certain human cognitive abilities: classifying, predicting, recognizing patterns, analyzing text, generating content, or supporting decision-making. It is not a single technology, but a family of very different approaches, ranging from specialized statistical models to generative systems.

What can it be used for in research?
In a scientific context, AI must always be assessed in light of its actual use:

What data does it use?
What results does it produce?
How are those results verified?
Can they be reproduced, explained, or documented?

Limitation and caveats
AI can support certain stages of scientific work, but it is not proof in itself. The quality of its outputs depends on how the task is framed, the data, the instructions, the model used, and the validation process.

# Analytical AI

Analytical AI refers to the use of artificial intelligence to analyze, structure, and interpret existing data. Rather than producing new content, it aims to extract useful information from that data.

What can it be used for in research?
Exploring a dataset, spotting trends, segmenting observations, detecting anomalies, or identifying relationships between variables. It applies to quantitative data as well as to textual and visual data, or data from measurement instruments.

Limitation and caveats
Analytical AI does not replace scientific analysis. It helps process and organize information, but interpretation still depends on the protocol, the methodological framework, data quality, and the researcher’s judgment. An algorithmically generated result is not automatically a scientifically valid one.

# Predictive AI

Predictive AI encompasses models designed to estimate a future, probable, or not-yet-observed outcome based on existing data. It is generally considered a branch of analytical AI: all predictive AI is analytical, but the reverse is not always true.

What can it be used for in research?
Predictive AI encompasses models designed to estimate a future, probable, or not-yet-observed outcome based on existing data. It is generally considered a branch of analytical AI: all predictive AI is analytical, but the reverse is not always true.

Limitation and caveats
A prediction is not an explanation. A model can identify useful correlations without establishing a causal relationship. Its validity depends on the quality of the training data, the representativeness of the sample, the choice of variables, the stability of the phenomenon under study, and evaluation on independent data.

# Generative AI

Generative AI refers to a family of models capable of producing new content from a prompt or context: text, images, code, audio, video, summaries, or translations. Large language models (LLMs) are the best-known example when they generate text-based responses.

What can it be used for in research?
Generative AI refers to a family of models capable of producing new content from a prompt or context: text, images, code, audio, video, summaries, or translations. Large language models (LLMs) are the best-known example when they generate text-based responses.

Limitation and caveats
Generating is not proving. Generative AI produces plausible content based on learned patterns: it guarantees neither the accuracy of a claim, nor the validity of a reference, nor the soundness of a line of reasoning. Generated content must be reviewed, verified, sourced, and placed within an explicit methodological framework.

# Multimodal AI

Multimodal AI refers to systems capable of processing several types of information, or modalities: text, images, audio, video, tables, charts, or complex documents. It is less a distinct family than a capability that can apply to both analytical and generative AI.

What can it be used for in research?
This capability opens up promising possibilities: analysis of experimental images, assisted reading of charts, exploration of scientific documents that combine text and figures, help with describing figures, or combined processing of textual and visual data.

Limitation and caveats
A multimodal model may misinterpret an image, confuse visual elements, overlook the experimental context, or produce a convincing but incorrect description. It supports analysis without replacing domain expertise or validation through established methods.

# Artificial General Intelligence

Artificial general intelligence, or AGI, refers to a hypothetical system capable of learning, reasoning, and transferring knowledge across a wide variety of tasks and domains without being designed for a specific purpose. Still theoretical today, it could represent a major turning point by adapting to new problems and applying its capabilities at a scale, speed, and level of versatility potentially exceeding those of humans in many fields.

What can it be used for in research?
AGI fuels discussions about the automation of complex tasks, scientific discovery, and the solving of interdisciplinary problems. It could one day assist researchers with work that requires greater adaptability and reasoning.

Limitation and caveats
There is no consensus definition or evaluation method. Current systems, even highly versatile ones, are not considered AGI. Claims presenting a technology as AGI should therefore be treated with caution.

The Fundamental Building Blocks of AI Systems

These concepts describe the main foundations that enable AI systems to learn from data, process language, and produce results. Understanding them makes it easier to assess their potential in research.

# Machine Learning

Machine learning is a branch of artificial intelligence that enables a system to learn from data. Rather than following predefined rules, the model identifies patterns in a dataset to produce predictions, classifications, or recommendations. It is the technical foundation of most of the types of AI described above, whether analytical, predictive, or generative.

What can it be used for in research?
Classifying images, detecting anomalies, analyzing time series, predicting experimental outcomes, segmenting observations, or exploring large volumes of data. It is one of the longest-established forms of AI in scientific practice, in use well before generative tools emerged.

Limitation and caveats
A model that performs well on a test set may prove unreliable in another context. Its use therefore requires explicit attention to data quality and representativeness, variable selection, overfitting, and the conditions for generalization. A good score on evaluation data is not enough to validate a model for scientific use.

# Deep Learning

Deep learning is a subset of machine learning based on neural networks with multiple processing layers. It is particularly well suited to complex data: images, language, sound, signals, sequences, or large document corpora.

What can it be used for in research?
Recognition and segmentation of scientific images, natural language processing, signal analysis, prediction of biological structures, or mining of large corpora. Most of today’s generative models and large language models are built on these approaches.

Limitation and caveats
Deep learning generally requires large volumes of data and significant computing resources, which raises questions about cost and environmental footprint. Its inner workings are also difficult to interpret, with direct implications for transparency, reproducibility, and trust in published results.

# NLP

NLP, which stands for Natural Language Processing, encompasses the methods that enable a computer system to analyze, transform, or generate human language.

What can it be used for in research?
Analyzing corpora of articles, extracting concepts, classifying documents, identifying named entities, coding interview transcripts, summarizing texts, or supporting literature monitoring. It is particularly useful in disciplines where literature, archives, and textual data play a central role.

Limitation and caveats
NLP is not limited to generative AI. Many methods that predate LLMs remain better suited to certain tasks, particularly when reproducibility, transparency, or control over processing are priorities. Choosing an LLM by default for a classification or extraction task can lead to a result that is less reliable and less well documented than a conventional method.

# LLM

An LLM, or Large Language Model, is a language model trained on massive text corpora. It can generate, rephrase, translate, summarize, or structure language, produce code, analyze documents, or be integrated into specialized tools. At its core, it works by predicting sequences of tokens, that is, words or word fragments.

What can it be used for in research?
Exploring a topic, rephrasing a text, preparing a summary, generating an article outline, assisting with programming, translating, or adapting content for a non-specialist audience. LLMs are the technical building block underpinning most of the generative AI uses described above.

Limitation and caveats
The fluency of a response says nothing about its accuracy. An LLM can produce structured, convincing text even when the facts, references, or reasoning are wrong. It should never be used as a single source of truth: bibliographic references, numerical data, and conclusions must be verified against primary sources.

# Token

A token is a unit of text that a language model uses to analyze and generate content. Depending on the model and the language, it may correspond to a word, part of a word, a punctuation mark, or a space. Before processing a prompt or document, the model converts the text into a sequence of tokens.

What can it be used for in research?
Understanding tokens provides better insight into how LLMs work and where their limits lie. In particular, the number of tokens determines how much text a model can take into account in a single request, which matters when analyzing a long article, querying a corpus, or designing a RAG system. Tokens also affect the cost, and sometimes the speed, of processing.

Limitation and caveats
A token does not necessarily correspond to a word. The token count varies depending on the language, the model, and the structure of the text. It measures the amount of text processed but says nothing about its quality or scientific complexity.

Interaction Tools

The following terms describe different ways of interacting with an artificial intelligence system. Since these labels vary from one vendor to another, we use them here as practical reference points, mainly based on the degree of autonomy given to the system.

# Chatbot

A chatbot is a conversational interface that lets users interact with a computer system through questions and answers. Older chatbots rely on predefined scripts, while more recent ones are built on language models capable of generating open-ended responses.

What can it be used for in research?
Directing users to resources, answering frequently asked questions about a platform or infrastructure, helping navigate internal documentation, or providing first-line support with software or a procedure. It is a support tool rather than a means of scientific production.

Limitation and caveats
Directing users to resources, answering frequently asked questions about a platform or infrastructure, helping navigate internal documentation, or providing first-line support with software or a procedure. It is a support tool rather than a means of scientific production.

# AI Assistant

An AI assistant is an AI system configured with predefined instructions and enriched with data and documents provided by the user. It is dedicated to a specific task (writing, rephrasing, summarizing, organizing ideas, generating code, transforming a document, or structuring an analysis) and acts as a specialized aid for that task. It goes beyond a simple question-and-answer exchange while keeping humans at the center: the user makes the decisions, step by step.

What can it be used for in research?
For a researcher or PhD student, an AI assistant can provide support with time-consuming tasks: clarifying a research question, drafting an article outline, rephrasing an abstract, creating an interview guide, explaining a script, or adapting content for a non-specialist audience.

Limitation and caveats
An AI assistant speeds up production without guaranteeing the scientific relevance of the output. Its effectiveness depends on the quality of its inputs (instructions and documents provided) and on how each request is framed. Expert review and systematic verification of factual content remain essential. Apparent time savings can turn into lost time if review is neglected.

# AI Agent

An AI agent is a system capable of chaining together multiple actions to achieve a goal. Whereas an assistant waits for instructions at each step, an agent can use tools, query databases, organize a sequence of tasks, produce files, or trigger actions in a digital environment.

What can it be used for in research?
Querying a document corpus, extracting and structuring information, producing a summary or report, automating recurring literature monitoring, or orchestrating certain steps of an analysis workflow. This is the level at which AI stops being a one-off tool and becomes embedded in a process.

Limitation and caveats
However, this autonomy remains relative and must be governed. The more actions an agent chains together, the greater the risk of compounding errors. It therefore becomes essential to document each step, check sources, limit access rights, and clearly define what falls under automation, assistance, and human validation.

Techniques for Adapting and Connecting AI Systems

These concepts describe different ways to guide, specialize, or connect an AI system. They operate at different levels and address different needs.

# Prompt Engineering

Prompt engineering refers to designing, writing, and refining the instructions given to a generative AI model. A prompt can be a simple question or a structured instruction specifying the context, objective, expected format, level of detail, and quality criteria.

What can it be used for in research?
In a scientific setting, prompt engineering helps obtain more relevant and actionable responses. For example, a researcher can ask the AI to rephrase a hypothesis, compare two methodological approaches, propose an outline for a literature review, or identify potential limitations of a protocol.

Limitation and caveats
A well-written prompt does not guarantee an accurate response. The quality of the output also depends on the model used, the available data, the context provided, the connected tools, and the verification method. It is also important to avoid entering confidential or sensitive data into a prompt without first checking how the tool processes that data.

# RAG

RAG, which stands for Retrieval-Augmented Generation, is an architecture that combines information retrieval from a vector document database with response generation by a language model. The system first retrieves relevant passages from a corpus, then passes them to the model to produce a contextualized response.

What can it be used for in research?
Querying a database of articles, locating information within publications, synthesizing a corpus, exploring internal documentation, or building an assistant specialized in a scientific field. In particular, RAG can help link a response more directly to the documents used to generate it.

Limitation and caveats
RAG does not eliminate errors or hallucinations. A response can remain unreliable if the corpus is incomplete, outdated, or poorly structured, if the wrong passages are retrieved, or if the model misinterprets the sources. The quality of the output therefore depends on indexing, retrieval, the corpus used, and the faithfulness of the generated response.

# Fine-Tuning

Fine-tuning refers to adapting an existing AI model to a particular task, domain, or format using specific data. Rather than training a model from scratch, a pretrained model is further trained on a targeted corpus.

What can it be used for in research?
Specializing a model for a classification task, a document type, discipline-specific vocabulary, or a particular response format. Fine-tuning can be useful when general-purpose models do not adequately meet a specific need and the organization has a sufficiently large volume of high-quality training data.

Limitation and caveats
Fine-tuning changes the model’s behavior, but it is not a reliable way to give it up-to-date knowledge or to ensure the traceability of its responses. To query an evolving document corpus on a regular basis, RAG is often better suited. Fine-tuning also requires careful attention to data quality, usage rights, confidentiality, bias, and evaluation of the resulting model.

Key takeaway:
RAG connects a model to external sources, while fine-tuning adapts the model using additional training data. The two approaches can be combined, but they address different needs.

# MCP

MCP, which stands for Model Context Protocol, is a protocol that enables an AI application to connect in a structured way to external tools, data, and resources. In particular, it can facilitate access to databases, documents, software, APIs, or domain-specific tools.

What can it be used for in research?
Connecting an AI assistant to an internal corpus, a bibliographic database, an analysis tool, a programming environment, or a laboratory workflow. The system can then retrieve information or use certain functions without the user having to perform each step manually.

Limitation and caveats
MCP does not automatically make a model more reliable, more capable, or more autonomous. The security and quality of the system depend on the connected tools, the access rights granted, the data made accessible, and the controls in place. A sound deployment requires limiting permissions, protecting sensitive data, logging actions, and reviewing the outputs produced.

Key Risks to Bo Aware Of

The risks associated with artificial intelligence are not limited to the errors produced by models. A system can also reproduce biases or be used without a clear framework within an organization. The following concepts help clarify these risks and how to mitigate them.

# AI Hallucination

An AI hallucination is a response generated by an artificial intelligence system that appears coherent but contains information that is false, fabricated, imprecise, or unverifiable. The term mainly applies to generative models.

What can it be used for in research?
A hallucination may take the form of a nonexistent bibliographic reference, a misattribution, an approximate definition, an incorrect interpretation, or an unsourced statistic. The risk is especially high when AI is used for a literature review, a method, a protocol, or a scientific conclusion.

How can this risk be limited?
References, statistics, quotations, and reasoning generated by AI must be verified against reliable sources, ideally primary ones. Generative AI can help explore, structure, or rephrase information, but it is not a scientific source in itself.

# Algorithmic Bias

Algorithmic bias refers to a systematic distortion in the outputs produced or influenced by an artificial intelligence system. It can stem from the data, the model, the optimization criteria, or the context of use.

What can it be used for in research?
A model trained on unrepresentative data may perform less well for certain populations, languages, disciplines, or data types. It may also reproduce imbalances present in the scientific literature or in the databases used to train it.

How can this risk be limited?
The data must be documented, performance assessed across relevant subgroups, results compared with other methods, and the model tested on independent data. Human oversight remains necessary, but it must be accompanied by rigorous methodological evaluation.

# Shadow AI

Shadow AI refers to the use of artificial intelligence tools within an organization without official approval, visibility, or oversight. It is not necessarily malicious: it often arises when a team tries to meet a need quickly without having suitable tools or rules in place.

What can it be used for in research?
A researcher or team may use an external tool to summarize an article, analyze data, edit a text, or generate code without checking how that tool processes data. The risk is particularly high when unpublished results, personal data, protocols, code, or confidential information are shared.

How can this risk be limited?
Organizations need to map existing uses, define a clear policy, identify approved tools, and classify data by sensitivity level. Training teams and providing opportunities to test new uses in a controlled environment help limit risks without stifling valuable initiatives.

Understanding the basic vocabulary of artificial intelligence lays an initial foundation for keeping up with its rapid developments and interpreting AI news with discernment. These reference points help teams better understand the possibilities and limitations of AI, build a shared AI culture, and use these tools wisely. This culture is the starting point for adopting AI in a way that is useful, reliable, and tailored to research practices.

Taking AI Integration Further

Building on these fundamentals, the next step is to turn an understanding of AI into concrete applications. This requires choosing the right tools, identifying relevant use cases, and defining a framework suited to the needs and constraints of each organization.

This is precisely the approach RITME has taken for over 35 years in supporting the scientific community, through a catalog of specialized software solutions, training programs and support services tailored to scientific environments.

EFFISCIANCE our AI support offering, follows this same approach. Designed by scientists, for scientists, It combines three complementary services:

Consulting: to clarify use cases, identify the right tools, and establish a realistic, secure framework aligned with your scientific objective.
Training: to build your teams’ skills and embed practical uses of generative AI into everyday scientific practice.
Integration: to turn identified use cases into concrete solutions, whether specialized assistants, AI agents, or scientific process automation.

To structure this approach, we have developed a comprehensive, modular AI Pathway. From introductory sessions to training, through use case identification and the integration of assistants or agents, it adapts to the maturity level, objectives, and constraints of each organization.

Ready to make your organization’s AI transformation a success?

Contact our experts to discuss your AI project.