Large language model

Simplified artificial neural network

Large Language Model or LLM is a type of artificial intelligence model trained on very large collections of text and other data to understand, generate, summarize, translate, classify, and reason with natural language. LLMs are a major technology within natural language processing, machine learning, deep learning, and generative artificial intelligence.

Large language models are commonly used in chatbots, search engines, document summarization, medical education, clinical decision support systems, patient education, translation, computer programming, research, data analysis, writing assistants, customer service, and healthcare informatics. They are called "large" because they typically contain billions or even trillions of numerical parameters and are trained on very large datasets using substantial computational resources.

Examples of well-known LLM or related foundation model families include GPT, Claude, Gemini, Llama, Mistral, Command R, Grok, Qwen, DeepSeek, Falcon, Gemma, and PaLM. Some newer models are not limited to text and are better described as large multimodal models because they can process or generate text, images, audio, video, or structured data.

Overview

A large language model is a statistical and computational system designed to model language. Given an input, called a prompt, the model produces an output by estimating the probability of the next token and repeating that process until a response is generated. A token may be a word, part of a word, punctuation mark, number, or symbol.

LLMs can perform many tasks, including:

• Question answering
• Text generation
• Summarization
• Translation
• Classification
• Information extraction
• Named-entity recognition
• Sentiment analysis
• Code generation
• Mathematical reasoning
• Clinical documentation
• Patient education
• Medical literature summarization
• Decision support
• Data analysis

Although LLMs can produce fluent and useful responses, they do not understand language in the same way humans do. They generate responses based on learned patterns in data, model architecture, training methods, and the prompt provided by the user.

Terminology

Common terms related to large language models include:

Term	Meaning
Artificial intelligence	The broad field of creating computer systems that perform tasks associated with human intelligence
Machine learning	A field of AI in which systems learn patterns from data
Deep learning	Machine learning using multi-layered neural networks
Natural language processing	The field of AI concerned with human language
Foundation model	A large model trained on broad data and adaptable to many tasks
Transformer architecture	A neural network architecture central to most modern LLMs
Token	A unit of text used by the model, such as a word or word fragment
Prompt	The input instruction, question, or context given to the model
Context window	The amount of input and output text a model can consider at one time
Embedding	A numerical representation of text, images, or other data
Fine-tuning	Additional training of a model for a particular task or domain
Retrieval-Augmented Generation	A method that combines an LLM with external information retrieval
Hallucination	A false or unsupported output generated by an AI model
Alignment	The process of making an AI system behave according to human goals, values, and instructions

Historical development

The development of LLMs grew out of earlier work in linguistics, statistics, computer science, and artificial neural networks.

Important historical milestones include:

• Rule-based natural language processing systems
• Statistical language models
• N-gram language models
• Recurrent neural networks
• Long short-term memory networks
• Word embedding models such as word2vec and GloVe
• Sequence-to-sequence learning
• Attention mechanisms
• Transformer architecture
• Pre-training and fine-tuning
• Foundation models
• Large multimodal models
• Chatbot systems using reinforcement learning from human feedback

The introduction of the transformer architecture greatly improved the ability of language models to process long sequences, learn relationships between words, and scale to very large datasets.

How large language models work

Tokens and tokenization

Before text is processed by an LLM, it is converted into smaller units called tokens. Tokenization may divide text into words, word fragments, punctuation, numbers, or special symbols.

For example, the sentence:

The patient has hypertension.

may be broken into tokens such as:

The | patient | has | hypertension | .

In many models, uncommon medical terms may be split into smaller subword units. Tokenization affects cost, speed, context length, and how the model handles medical terminology.

Next-token prediction

Most LLMs are trained using a task called next-token prediction. The model sees a sequence of tokens and learns to predict the next likely token. By repeating this process, the model can generate paragraphs, tables, code, or conversational responses.

Example:

Prompt: The capital of Brazil is
Likely continuation: Brasília

This does not mean the model "knows" facts in a human sense. It has learned patterns from training data and may generate incorrect information if the prompt is ambiguous, if the model lacks reliable information, or if it overgeneralizes.

Neural network parameters

A parameter is a numerical value learned during training. Parameters determine how the model transforms input tokens into output probabilities. Larger models often have more parameters and may perform better on complex tasks, but size alone does not guarantee accuracy, safety, or usefulness.

Transformer architecture

Most modern LLMs use the transformer architecture. Transformers use attention mechanisms to weigh relationships among tokens in a sequence.

Key components include:

• Self-attention
• Multi-head attention
• Feed-forward neural network layers
• Positional encoding
• Layer normalization
• Residual connections
• Token embeddings

Transformers are especially useful because they can process language in parallel and learn long-range relationships.

Attention mechanism

The attention mechanism allows the model to focus on relevant parts of the input when generating output. In medical text, attention may help a model connect a symptom, diagnosis, medication, laboratory value, and clinical instruction within the same prompt.

Context window

The context window is the amount of information the model can consider at one time. A larger context window allows longer documents, medical notes, guidelines, or conversations to be included in the prompt. However, a longer context window does not guarantee that every detail will be used correctly.

Training process

Pre-training

Pre-training is the initial stage in which a model learns general patterns of language from large datasets. Training data may include books, websites, articles, code, public documents, and other text sources.

During pre-training, the model learns:

• Grammar
• Word relationships
• Common facts
• Writing styles
• Reasoning patterns
• Code syntax
• Domain-specific terminology
• Common question-answer patterns

Fine-tuning

Fine-tuning is additional training on a more specific dataset or task. A model may be fine-tuned for:

• Medical question answering
• Legal document analysis
• Customer support
• Coding
• Scientific writing
• Translation
• Safety behavior
• Instruction following

Instruction tuning

Instruction tuning trains a model to follow user instructions. This improves its ability to respond to prompts such as "summarize this article," "explain this diagnosis," or "write a patient education handout."

Reinforcement learning from human feedback

Reinforcement learning from human feedback or RLHF is a training method in which human preferences are used to improve model responses. Humans compare outputs, rate quality, and help train reward models that guide the system toward more useful and safer responses.

Constitutional AI and safety tuning

Some models use safety training methods such as constitutional AI, red-teaming, policy training, and refusal behavior to reduce harmful outputs. These methods aim to reduce dangerous, biased, misleading, or inappropriate responses.

Types of large language models

General-purpose LLMs

General-purpose LLMs are trained to perform many types of tasks, including writing, summarization, translation, reasoning, tutoring, coding, and conversation.

Examples include:

• GPT
• Claude
• Gemini
• Llama
• Mistral
• Command R

Open-weight LLMs

An open-weight model provides publicly downloadable model weights under a license. Open-weight models may be used by researchers, companies, and developers for local deployment, customization, or fine-tuning.

Examples include:

• Llama
• Mistral
• Falcon
• Qwen
• Gemma
• DeepSeek

Open-weight does not always mean fully open-source. Licenses, training data transparency, commercial permissions, and usage restrictions vary.

Closed or proprietary LLMs

A proprietary LLM is usually accessed through a web application or application programming interface rather than downloaded by the user. The model weights and full training details are usually not public.

Examples include:

• GPT models from OpenAI
• Claude models from Anthropic
• Gemini models from Google
• Grok models from xAI
• Commercial models from other AI companies

Domain-specific LLMs

Domain-specific LLMs are optimized for a particular field such as:

• Medicine
• Law
• Finance
• Education
• Computer programming
• Biology
• Chemistry
• Radiology
• Pharmacology

Medical LLMs require special caution because incorrect outputs may affect patient care.

Large multimodal models

A large multimodal model can process or generate more than one type of data. Inputs or outputs may include:

• Text
• Image
• Audio
• Video
• Tables
• Code
• Medical images
• Documents

In healthcare, multimodal AI may be used for radiology, pathology, clinical documentation, patient education, and combining text with images or structured records.

Examples of large language models

OpenAI GPT models

GPT models are a family of large language models developed by OpenAI. GPT stands for Generative Pre-trained Transformer. GPT models are used in ChatGPT, the OpenAI API, coding tools, writing tools, and many AI applications.

Examples include:

• GPT-3
• GPT-3.5
• GPT-4
• GPT-4o
• GPT-5
• GPT-5.1
• GPT-5.5

GPT models are known for broad language ability, reasoning, summarization, coding, conversation, and use in general-purpose assistants.

Anthropic Claude models

Claude is a family of large language models developed by Anthropic. Claude models are used for writing, coding, document analysis, reasoning, summarization, and enterprise applications.

Examples include:

• Claude 3
• Claude 3.5
• Claude 3.7
• Claude Sonnet 4
• Claude Opus 4
• Claude Opus 4.1
• Claude Sonnet 4.6

Claude models are often associated with long-context document analysis, coding, instruction following, and AI safety research.

Google Gemini models

Gemini is a family of AI models developed by Google and Google DeepMind. Gemini models are designed for language, reasoning, coding, multimodal input, and integration with Google products and developer tools.

Examples include:

• Gemini 1.0
• Gemini 1.5
• Gemini 2.0
• Gemini 2.5 Pro
• Gemini 2.5 Flash
• Gemini 2.5 Flash-Lite
• Gemini 3.5 Flash

Gemini models are commonly discussed as multimodal foundation models for text, image, code, and agentic workflows.

Meta Llama models

Llama is a family of open-weight large language models developed by Meta Platforms. Llama models are widely used in research, local AI systems, enterprise deployments, and open-weight AI development.

Examples include:

• Llama
• Llama 2
• Llama 3
• Llama 3.1
• Llama 3.2
• Llama 4 Scout
• Llama 4 Maverick
• Llama 4 Behemoth

Llama models are important because they helped popularize open-weight large language models for local and customizable AI systems.

Mistral models

Mistral AI develops open-weight and commercial language models. Mistral models are used for general language tasks, coding, enterprise AI, and multilingual applications.

Examples include:

• Mistral 7B
• Mixtral
• Mistral Large
• Mistral Large 3
• Mistral Medium
• Mistral Medium 3.5
• Devstral

Mistral models are often noted for efficient performance, open-weight releases, multilingual use, and developer applications.

Cohere Command models

Cohere develops language models and enterprise AI systems. The Command model family is used for chat, retrieval, summarization, long-context tasks, and business applications.

Examples include:

• Command R
• Command R+
• Command A
• Aya

Cohere Command models are commonly used in enterprise search, retrieval-augmented generation, and multilingual workflows.

Other notable LLM families

Other important LLM or foundation model families include:

• Grok from xAI
• Qwen from Alibaba
• DeepSeek
• Falcon
• Gemma from Google
• Yi
• Phi from Microsoft
• BLOOM
• PaLM
• ERNIE Bot
• Jurassic from AI21 Labs

Comparison of selected LLM examples

Model family	Developer	Common type	Example models	Common uses
GPT	OpenAI	Proprietary foundation model	GPT-4, GPT-4o, GPT-5.1, GPT-5.5	Chatbots, coding, writing, reasoning, education, applications
Claude	Anthropic	Proprietary foundation model	Claude Sonnet, Claude Opus	Long-context analysis, writing, coding, reasoning, enterprise work
Gemini	Google DeepMind	Proprietary multimodal foundation model	Gemini Pro, Gemini Flash, Gemini Flash-Lite	Search, productivity, coding, multimodal AI, agentic tasks
Llama	Meta Platforms	Open-weight model family	Llama 3, Llama 4 Scout, Llama 4 Maverick	Research, local AI, fine-tuning, enterprise deployment
Mistral	Mistral AI	Open-weight and commercial models	Mistral 7B, Mixtral, Mistral Large, Devstral	Multilingual AI, coding, enterprise AI, local deployment
Command R	Cohere	Enterprise-oriented language model	Command R, Command R+, Command A	Retrieval, summarization, long-context tasks, business search
Qwen	Alibaba	Open and commercial model family	Qwen language and multimodal models	Multilingual tasks, coding, research, applications
DeepSeek	DeepSeek	Open-weight and commercial models	DeepSeek language and reasoning models	Coding, reasoning, research, local deployment

Common capabilities

Large language models can perform a wide range of tasks, including:

• Generating explanations
• Answering questions
• Summarizing long documents
• Translating between languages
• Drafting letters, reports, and educational materials
• Extracting structured information
• Creating tables
• Generating computer code
• Debugging software
• Producing outlines
• Classifying documents
• Rewriting content for different audiences
• Tutoring students
• Assisting with research
• Explaining medical terms
• Creating patient education materials

Limitations

LLMs have important limitations. They can produce answers that appear confident but are wrong, outdated, incomplete, biased, or unsupported.

Common limitations include:

• Hallucination
• Outdated knowledge
• Bias from training data
• Overconfidence
• Incomplete reasoning
• Poor citation accuracy
• Inability to guarantee factual correctness
• Vulnerability to prompt injection
• Difficulty with rare facts
• Difficulty with rapidly changing information
• Possible misinterpretation of medical or legal instructions

In clinical settings, these limitations are especially important because incorrect information can harm patients.

Hallucination

Hallucination in AI refers to generated information that is false, fabricated, misleading, or not supported by reliable evidence. In medicine, hallucination may include incorrect drug doses, invented citations, inaccurate diagnoses, or unsafe treatment suggestions.

Examples of AI hallucination include:

• Inventing a medical guideline
• Giving a wrong drug dose
• Misstating a contraindication
• Creating a fake journal reference
• Confusing similar diseases
• Overstating the benefits of a treatment
• Omitting urgent warning signs

Prompting

Prompt engineering is the practice of writing inputs that guide an LLM toward useful outputs. A prompt may include instructions, context, examples, constraints, and desired output format.

A simple prompt:

Explain hypertension in simple language for a patient.

A more specific prompt:

Create a 1-page patient education handout on hypertension at a 6th-grade reading level.
Include symptoms, risks, home blood pressure monitoring, lifestyle changes, medicines, and when to call a doctor.

In healthcare, prompts should clearly state the audience, purpose, reading level, and safety requirements.

Retrieval-Augmented Generation

Retrieval-Augmented Generation or RAG is a method that connects an LLM to external information sources, such as databases, guidelines, medical references, or document collections.

A RAG system usually has three steps:

1. Retrieve relevant documents or passages.
2. Provide the retrieved information to the LLM.
3. Generate an answer based on the retrieved context.

RAG can reduce hallucination and improve accuracy when implemented carefully, but it does not eliminate the need for source quality control, validation, and human review.

Fine-tuning and customization

LLMs can be customized for specific tasks through:

• Fine-tuning
• Instruction tuning
• Retrieval-augmented generation
• Prompt templates
• System instructions
• Tool use
• Human feedback
• Domain-specific datasets

In healthcare, customization may be used for:

• Medical coding
• Clinical note summarization
• Patient education
• Prior authorization support
• Research literature screening
• Medical terminology expansion
• Care navigation

LLMs in healthcare

LLMs have many potential uses in healthcare, but they require careful oversight. They may assist clinicians, patients, educators, researchers, and administrators.

Potential healthcare uses include:

• Patient education
• Medical education
• Drafting discharge instructions
• Summarizing clinical notes
• Summarizing medical literature
• Explaining laboratory results
• Supporting clinical documentation
• Assisting medical coding
• Supporting triage workflows
• Generating differential diagnosis lists
• Creating plain-language explanations
• Translating health information
• Supporting public health communication
• Assisting in research protocol development

Clinical decision support

LLMs may be used in clinical decision support systems, but they should not replace clinician judgment. In clinical use, they must be evaluated for accuracy, safety, bias, workflow fit, privacy, and regulatory compliance.

Appropriate clinical safeguards include:

• Human clinician review
• Reliable source grounding
• Clear uncertainty statements
• Audit trails
• Data privacy controls
• Bias testing
• Regular model evaluation
• Avoidance of unsupervised diagnosis or treatment decisions
• Clinical governance

Medical education

In medical education, LLMs may help students and teachers by generating:

• Explanations of medical concepts
• Practice questions
• Case-based learning scenarios
• Differential diagnosis exercises
• Anatomy summaries
• Pharmacology tables
• Patient communication scripts
• Study plans
• Glossaries

However, learners should verify AI-generated content against textbooks, guidelines, peer-reviewed literature, and faculty instruction.

Patient education

LLMs can help create patient education materials at different reading levels and in multiple languages. They may help explain:

• Diseases
• Symptoms
• Medications
• Procedures
• Laboratory tests
• Lifestyle changes
• Warning signs
• Follow-up care

Patient materials should be reviewed by qualified healthcare professionals before publication.

Medical research

LLMs may assist medical research by helping with:

• Literature search planning
• Abstract screening
• Data extraction
• Protocol drafting
• Statistical explanation
• Manuscript editing
• Grant writing
• Code generation
• Hypothesis generation

They should not be used to fabricate data, citations, authorship contributions, or peer review.

Risks in medicine

Risks of LLM use in medicine include:

• Incorrect medical advice
• Unsafe medication dosing
• Missed emergency symptoms
• Fabricated citations
• Bias against patient groups
• Privacy violations
• Overreliance by patients or clinicians
• Confusion between education and diagnosis
• Incomplete documentation
• Failure to follow local clinical guidelines

LLM outputs should be considered assistive information, not a substitute for professional medical care.

Privacy and security

LLMs may process sensitive information, including protected health information, electronic health record data, and patient messages. Healthcare organizations must consider:

• Health Insurance Portability and Accountability Act compliance
• Data minimization
• De-identification
• Access control
• Encryption
• Vendor agreements
• Audit logging
• Retention policies
• Consent
• Cybersecurity
• Prompt injection attacks

Sensitive patient information should not be entered into public AI systems unless appropriate legal, privacy, and institutional safeguards are in place.

Bias and fairness

LLMs may reflect biases present in their training data. Bias may affect outputs related to:

• Race
• Ethnicity
• Sex
• Gender
• Age
• Disability
• Socioeconomic status
• Language
• Geography
• Medical access
• Insurance status

In healthcare, bias can worsen disparities if not detected and mitigated. Evaluation should include diverse patient populations and clinical scenarios.

Evaluation of LLMs

LLMs are evaluated using benchmarks, human review, task-specific tests, safety assessments, and real-world performance monitoring.

Common evaluation areas include:

• Accuracy
• Helpfulness
• Safety
• Robustness
• Reasoning
• Calibration
• Bias
• Toxicity
• Factuality
• Citation quality
• Latency
• Cost
• Privacy
• Clinical usefulness
• User satisfaction

Medical evaluation should include expert review by qualified clinicians.

Benchmarks

Benchmarks are standardized tests used to compare model performance. Examples include:

• General knowledge benchmarks
• Coding benchmarks
• Mathematical reasoning benchmarks
• Medical question-answering benchmarks
• Multilingual benchmarks
• Long-context benchmarks
• Safety benchmarks
• Bias benchmarks

Benchmarks are useful but incomplete. High benchmark scores do not guarantee safe performance in real clinical practice.

Model collapse

Model collapse is a potential problem in which AI models trained heavily on synthetic AI-generated content may gradually lose quality, diversity, or factual grounding. Maintaining high-quality human-written and expert-reviewed content is important for future AI training and retrieval systems.

LLMs and search engines

LLMs are increasingly integrated into search engines and knowledge systems. Instead of only returning links, AI-enhanced search systems may summarize information, answer questions, generate comparisons, and guide users to sources.

Potential benefits include:

• Faster information access
• Natural-language search
• Summarized results
• Multilingual access
• Personalized explanations

Potential risks include:

• Reduced source visibility
• Incorrect summaries
• Overreliance on AI-generated answers
• Reduced traffic to original publishers
• Difficulty evaluating evidence quality

LLMs and knowledge bases

High-quality knowledge bases, including medical encyclopedias, can support LLM systems through:

• Reliable source material
• Structured terminology
• Internal links
• Topic hierarchies
• Definitions
• Categories
• Disambiguation pages
• Human editorial review
• Retrieval-augmented generation

Wiki-based content can be especially useful for retrieval systems because it often contains organized pages, categories, templates, and internal links.

LLMs and WikiMD

On WikiMD, large language models are relevant to:

• Medical encyclopedia development
• Patient education
• Medical terminology
• Drug encyclopedia content
• Food encyclopedia content
• Nutrition education
• Healthcare informatics
• Artificial intelligence in healthcare
• Clinical decision support system
• Retrieval-Augmented Generation
• Medical writing
• Health communication

LLMs may assist with drafting, summarizing, translating, categorizing, and formatting health content. However, medical content should be reviewed for accuracy, safety, neutrality, and currency.

Examples of LLM prompts for healthcare education

Patient education prompt

Create a patient education handout on type 2 diabetes at a 6th-grade reading level.
Include symptoms, complications, diet, exercise, medicines, blood sugar monitoring,
and when to call a doctor.

Medical student prompt

Explain the pathophysiology of diabetic ketoacidosis for medical students.
Include insulin deficiency, counterregulatory hormones, ketogenesis, acidosis,
electrolyte abnormalities, and treatment principles.

Clinical documentation prompt

Summarize this clinic note into a problem-based assessment and plan.
Do not add information that is not present in the note.
Flag any missing information needed for safe clinical decision-making.

RAG prompt

Using only the provided clinical guideline excerpts, summarize the recommended
first-line treatment options for hypertension in adults. Cite the relevant excerpt
for each recommendation.

Best practices for using LLMs

Best practices include:

• Use clear and specific prompts.
• State the intended audience.
• Ask for uncertainty when appropriate.
• Verify important facts.
• Use trusted sources for medical information.
• Avoid entering private patient data into unsecured systems.
• Review outputs before publication.
• Use retrieval-augmented generation for factual domains.
• Keep humans responsible for decisions.
• Test systems before deployment.
• Monitor for bias and errors.
• Update workflows as models change.

Limitations in medical encyclopedia writing

When LLMs are used to help write encyclopedia articles, editors should check:

• Accuracy of definitions
• Correct use of medical terminology
• Current treatment guidelines
• Drug names and doses
• Contraindications
• Warning signs
• Internal links
• Categories
• Citations
• Neutral tone
• Patient safety language
• Copyright compliance

AI-generated content should not be copied from copyrighted medical sources. It should be reviewed, rewritten, cited when needed, and adapted to the encyclopedia's editorial standards.

Ethical issues

Ethical concerns related to LLMs include:

• Patient privacy
• Informed consent
• Transparency
• Bias
• Accountability
• Automation bias
• Misinformation
• Labor displacement
• Intellectual property
• Environmental cost
• Unequal access
• Safety in high-risk settings

In healthcare, ethical deployment requires oversight, validation, documentation, and clear boundaries.

Regulation and governance

Regulation of LLMs depends on jurisdiction and use case. A general-purpose chatbot may be regulated differently from an AI system used for clinical diagnosis or treatment recommendations.

Governance considerations include:

• Whether the system is a medical device
• Intended use
• Risk classification
• Data privacy laws
• Institutional review
• Human oversight
• Monitoring and auditing
• Incident reporting
• Vendor management
• Model updates
• Documentation

Environmental considerations

Training and running large models can require significant electricity, water, specialized hardware, and data center resources. Environmental impact depends on model size, training method, inference volume, energy source, hardware efficiency, and system design.

Potential mitigation strategies include:

• Efficient model architectures
• Smaller task-specific models
• Model compression
• Quantization
• Energy-efficient hardware
• Renewable energy
• Caching
• Better routing between small and large models

Future directions

Future development of LLMs may include:

• More accurate reasoning
• Better factual grounding
• Longer context windows
• Multimodal medical understanding
• Better citation and source tracking
• Safer clinical decision support
• Smaller efficient local models
• Personalized education
• Improved multilingual health communication
• Better integration with electronic health records
• Stronger privacy-preserving methods
• More transparent evaluation
• Specialized medical models
• Agentic workflows with human oversight

Advantages and disadvantages

Advantages	Disadvantages
Can summarize large amounts of text	May generate false or unsupported information
Can explain complex topics in simple language	May be biased or incomplete
Can support multilingual communication	May mishandle sensitive medical information
Can assist with writing and coding	May produce unsafe outputs without review
Can improve access to information	May encourage overreliance
Can support education and research	May fabricate citations or references

Teaching points

For students of health informatics, medicine, public health, and computer science, the most important teaching points are:

• An LLM is a statistical language model, not a human expert.
• LLMs generate text by predicting tokens from context.
• Transformer architecture is central to most modern LLMs.
• Fluent output does not guarantee truth.
• Hallucination is a major safety concern.
• Medical use requires human review and validation.
• Retrieval-augmented generation can improve factual grounding.
• Privacy and bias are major healthcare concerns.
• LLMs are tools that can support, but not replace, professional judgment.
• Good prompts, good sources, and good governance are essential.

See also

• Artificial intelligence
• Artificial intelligence in healthcare
• Machine learning
• Deep learning
• Natural language processing
• Computational linguistics
• Generative artificial intelligence
• Foundation model
• Large multimodal model
• Transformer architecture
• Prompt engineering
• Retrieval-Augmented Generation
• Chatbot
• Clinical decision support system
• Medical informatics
• Bioinformatics
• Medical education
• Patient education
• Electronic health record
• Health information technology
• Algorithmic bias
• Data privacy
• Hallucination
• GPT
• ChatGPT
• Claude
• Gemini
• Llama
• Mistral AI
• Cohere
• DeepSeek
• Qwen

External links

• OpenAI
• Anthropic
• Google DeepMind
• Meta AI
• Mistral AI
• Cohere
• National Institute of Standards and Technology: Artificial Intelligence
• FDA: Artificial Intelligence and Machine Learning in Software as a Medical Device