top of page

Glossar

Term
What It Means
Quick Example
Why It Matters
AI (Artificial Intelligence)
Computer programs that learn from data to do tasks like a human.
Chatbot answering customer questions.
Helps automate parts of your project.
Generative AI (Gen‑AI)
AI that creates text, images, audio, or code.
It drafts a logo idea or writes ad copy.
Speeds creative work.
Hallucination
When AI confidently makes up something false.
Your AI cites a fake source.
You must always check AI output.
LLM (Large Language Model)
A big AI trained on loads of text to write like a person.
GPT‑style tool writing a business pitch.
Helps draft documents quickly.
Model
A tool that predicts things based on patterns it learned.
Predicting sales from past data.
Helps make data‑informed decisions.
Training Data
Examples you teach the AI with.
10,000 labeled cat photos.
It learns from the examples.
Validation Data
Used to tune the AI while learning.
Try settings to pick the best model.
Avoids poor training choices.
Test Data
Data the AI hasn’t seen before.
New cat photos to check accuracy.
Measures real performance.
Feature Engineering
Turning raw data into useful inputs.
Turn dates into “is weekend”.
Makes AI smarter.
Overfitting
AI memorizes training data and fails on new cases.
A student crams answers then fails new test.
You want AI to generalize.
Neural Network
Layers of math that learn patterns.
Like stacked image filters.
Powers modern AI.
Transformer
Model that looks at all parts of input at once.
Reads a whole paragraph for context.
Core tech behind chatbots.
Attention
Mechanism that focuses on important bits of input.
Highlights key words in a sentence.
Improves understanding.
Latency
How long the AI takes to answer.
A pause after you ask a question.
Affects user experience.
Algorithm
A recipe of steps to solve a problem.
Cookie recipe = algorithm.
Behind every feature.
Prompt
The message or instructions you give the AI.
“Write a friendly email to parents.”
Clear prompts get better results.
Prompt Engineering
Crafting prompts to steer outputs.
Add role, style, and constraints.
Boosts quality without coding.
System Prompt
Hidden “house rules” the AI follows.
“You are a helpful tutor.”
Sets tone and behavior.
Tokens
Tiny chunks of text the model reads/writes.
“Hello” ≈ 1–2 tokens.
Affects cost, limits, and context.
Context Window
How much text the model can remember in one go.
Past messages it still “sees”.
Limits long chats and big docs.
Temperature
Controls creativity vs. consistency.
Lower = safer; higher = more creative.
Tune for stable or novel outputs.
Top‑p (Nucleus Sampling)
Picks from the most likely words up to a cutoff.
Keeps outputs fluent, less random.
Another creativity dial.
Max Tokens
Cap on response length.
Stop replies from getting too long.
Prevents cutoff or extra cost.
Stop Sequence
Characters that tell the model to stop.
Stop at “\n\nUser:”.
Keeps outputs tidy.
Grounding
Connecting answers to real sources or data.
Bot cites your handbook or website.
Reduces hallucinations; builds trust.
RAG (Retrieval‑Augmented Generation)
The AI searches your files, then writes.
Ask “Summarize our policy PDF.”
Keeps answers accurate and up‑to‑date.
Knowledge Base
Your trusted documents and FAQs.
Help center articles.
Powers grounded answers.
Embeddings
Number lists that capture meaning of text.
“Car” close to “vehicle”.
Used for search and RAG.
Vector Database
Stores embeddings for fast meaning‑based search.
Find similar questions quickly.
Makes Q&A snappy and relevant.
Fine‑Tuning
Teaching a model your style with examples.
Train on your brand voice.
Improves accuracy for your niche.
LoRA / Adapters
Light‑weight fine‑tuning methods.
Add new skill without retraining all.
Cheaper, faster customization.
Zero‑Shot
Model handles a task with no examples.
“Write a press release.”
Good baseline ability.
Few‑Shot
You show a few examples in the prompt.
Provide 3 labeled Q&As.
Boosts output quality fast.
Chain‑of‑Thought (CoT)
The model reasons step‑by‑step (often hidden to users).
“Let’s reason this out stepwise.”
Can improve complex reasoning.
Tool Use / Function Calling
AI calls tools like search, math, or databases.
“Check stock price, then explain.”
Gets up‑to‑date, correct info.
Agent
An AI that can plan steps, call tools, and iterate.
Draft → check facts → revise.
Automates multi‑step tasks.
Autonomy Level
How much freedom the agent has.
From “suggest only” to “execute”.
Balance speed with control.
Human‑in‑the‑Loop (HITL)
People review or guide AI steps.
Approve email before sending.
Keeps quality and safety high.
RLHF
Training with human feedback on good vs bad answers.
Raters prefer polite, helpful replies.
Aligns AI with human values.
DPO (Direct Preference Optimization)
Newer way to learn from preferences without full RL.
Learn from A vs B choices.
Efficient alignment tuning.
Safety Policy
Rules on what AI may or may not do.
Block harmful instructions.
Protects users and compliance.
Guardrails
Checks that shape or block outputs.
Filter hate speech or PII.
Reduces risk.
Content Moderation
Screening for harmful content.
Flag bullying in chat.
Safer spaces for students.
Bias
Systematic unfairness in data or outputs.
Stereotyped job ads.
Must be identified and reduced.
Fairness
Treating groups equitably.
Balanced training examples.
Ethical and often legally required.
Explainability (XAI)
Making AI decisions understandable.
Show key features used.
Builds trust and accountability.
Transparency
Clear info on how AI was built/used.
Model card and data sources.
Enables responsible adoption.
Model Card
A short “label” describing a model.
States limits and use cases.
Guides safe, proper use.
Data Sheet (for Datasets)
Notes about how data was collected and cleaned.
When/where/what/consent.
Informs risk and bias checks.
Red Teaming
Actively testing AI for failures.
Try to trigger bad behavior.
Finds weaknesses before launch.
Prompt Injection
Malicious text that hijacks the model.
Hidden “ignore rules” in a page.
Key security risk in RAG/agents.
Jailbreak
Tricks to bypass safety rules.
Role‑play to force bad output.
Monitor and prevent abuse.
Data Leakage
Sensitive info appears in outputs.
Model reveals a password.
Breach risk; avoid and detect.
PII (Personal Data)
Info that identifies a person.
Name, email, ID number.
Protect under GDPR and school rules.
Data Minimisation
Only collect what you truly need.
Ask ZIP, not full address.
Reduces privacy risk.
Anonymisation
Remove personal identifiers.
Strip names and emails.
Safer sharing and analysis.
Differential Privacy
Add noise so individuals can’t be identified.
Noisy counts in reports.
Protects privacy with utility.
Federated Learning
Train across devices without moving raw data.
Phones learn keyboard suggestions.
Privacy‑friendly training.
Synthetic Data
AI‑generated training examples.
Create rare case examples.
Fills data gaps safely (if done well).
Data Augmentation
Expand data with simple tweaks.
Paraphrase questions.
Improves robustness.
Watermarking
Hidden marks in AI media.
Detect if image is AI‑made.
Supports trust and attribution.
Responsible AI
Building AI that is safe, fair, and useful.
Policies, audits, and training.
Essential for schools and startups.
Accessibility
Features that help all users participate.
Text‑to‑speech, captions.
Expands inclusion and reach.
Multimodal Model
Understands/creates multiple media types.
Describe an image from a photo.
Richer, more natural interfaces.
VLM (Vision‑Language Model)
Connects images and text.
“Explain this chart.”
Enables visual Q&A.
ASR (Speech‑to‑Text)
Turns voice into text.
Transcribe a lecture.
Faster notes; supports hearing loss.
TTS (Text‑to‑Speech)
Turns text into natural audio.
Read an article aloud.
Supports reading and language needs.
OCR
Extracts text from images or PDFs.
Read a scanned worksheet.
Prepares docs for AI search.
Image Generation
AI creates pictures from prompts.
“Draw a Roman helmet.”
Prototypes visuals fast.
Diffusion Model
A popular way to generate images.
From noise to a clear picture.
Higher quality visuals.
Evaluation (LLM Evals)
Ways to score AI quality.
Rubrics for helpfulness and safety.
Tracks improvements.
Benchmarks
Standard tests to compare models.
Reading or math benchmarks.
Choose the right model.
Cost per Token
The price for input + output tokens.
“This reply cost €0.01.”
Helps budget usage.
Rate Limits
Caps on how often you can call an API.
60 requests per minute.
Plan for scale.
Runway (Ops)
How long your budget lasts at current use.
€200/month on AI tools.
Keeps projects sustainable.
AI Literacy
Understanding how AI works and its limits.
A class teaching students about chatbots.
Essential skill for everyone today.
AI Ethics
Principles for responsible AI.
Fairness, privacy, transparency.
Guides safe use.
AI Governance
Rules and oversight of AI systems.
Company board reviews AI risks.
Ensures accountability.
Cognitive Load
The mental effort for users.
Complex AI UI overwhelms.
Lighter load = better adoption.
Conversational AI
Systems that chat naturally with people.
Customer service bot.
Improves engagement.
Knowledge Distillation
Making a smaller model learn from a big one.
Train a light model for phones.
Saves cost and energy.
Pruning
Removing unused model parts.
Cut small weights.
Speeds up AI.
Quantization
Store numbers with fewer bits.
8‑bit instead of 32‑bit.
Makes AI cheaper and smaller.
Explain Like I’m 5 (ELI5)
AI breaks down answers very simply.
AI explains “blockchain” in kid terms.
Improves accessibility.
Continuous Learning
AI keeps learning after deployment.
Chatbot improves daily.
Adapts to new info.
Lifelong Learning
AI remembers and adapts over long time.
Tutor AI gets better with each student.
Sustains improvement.
Curriculum Learning
Train AI on simple tasks first.
Easy → harder problems.
Better training quality.
Self‑Supervised Learning
Learns patterns from raw data.
Predict next word from text.
Enables massive LLM training.
Reinforcement Learning
AI learns by rewards and penalties.
Game AI learns to win.
Great for dynamic tasks.
Few‑Shot Classification
AI categorizes from a handful of examples.
Labeling 5 product reviews.
Works when data is scarce.
Zero‑Resource Translation
Translate languages without parallel data.
English ↔ Swahili.
Expands global reach.
Knowledge Graph
A web of linked concepts.
“Rome → Italy → Europe.”
Supports reasoning and search.
Ontology
Structured vocabulary for a domain.
Terms in medicine.
Organizes knowledge.
Agentic Workflow
When AI plans, executes, and reflects on tasks.
Research assistant AI.
Future of autonomous AI.
Co‑Pilot
AI assistant that helps but doesn’t replace.
GitHub Copilot for coding.
Augments human skills.
AI Tutor
AI that adapts to a learner’s needs.
Personalized math practice.
Boosts education impact.
Socratic AI
AI that guides by asking questions.
“Why do you think that?”
Develops critical thinking.
Emotional AI
AI that senses mood from text/voice.
Detects frustration in support calls.
Adds empathy to systems.
Sentiment Analysis
AI that reads emotion in text.
Reviews marked positive/negative.
Useful in marketing & feedback.
Chat History
Memory of prior exchanges.
Bot recalls yesterday’s question.
Makes conversations smoother.
Session
One period of AI use with context.
Open app → chat → close.
Defines memory scope.
Context Switching
Changing tasks/topics mid‑chat.
From math to history.
Tests model flexibility.
Multilingual AI
Works across languages.
Translate & chat in Bulgarian + English.
Enables cross‑border use.
Cross‑Lingual Transfer
Knowledge shared across languages.
Learn English → improve French.
Efficient training.
Emergent Ability
Surprising skills appear at scale.
Model learns logic without being told.
Shows potential but needs caution.
Alignment
Ensuring AI does what humans want.
“Be helpful, not harmful.”
Central to safe AI.
Misalignment
When AI goals diverge from ours.
Optimizes clicks, not wellbeing.
Creates risks.
Singularity (Hypothetical)
A future where AI surpasses human intelligence.
Sci‑fi superintelligence.
Sparks debate, not reality yet.
AI Winter
Period when AI progress slows/funding drops.
Past decades saw hype collapse.
Context for current optimism.
bottom of page