Newton Institute of Technology Logo
Newton Institute of Technology Building Your Dream Career

AI vs Machine Learning vs Data Science: Explained

ICT learning guide Three connected fields

AI vs Machine Learning vs Data Science: What’s the Difference?

Understand · Compare · Practise

Learn the difference through one school ICT help-desk example, a working spreadsheet exercise and a small Python classifier that makes mistakes you can investigate.

Understand the relationship

Different purposes. Useful overlap.

TopicAI · ML · Data Science

LevelBeginner

IncludesExcel + Python exercises

Updated10 November 2026

Quick answer

What is the difference?

Artificial Intelligence (AI) creates systems that perform tasks associated with intelligence. Machine Learning (ML) is a branch of AI that learns patterns from data or feedback. Data Science analyses information to answer questions and guide decisions; it can use ML but does not have to. [1] [2]

  • Compare the three
  • Work through a school example
  • Try Excel and Python

AI vs ML vs Data Science: Quick comparison

Feature Artificial Intelligence Machine Learning Data Science
Purpose Perform tasks that require intelligent behaviour Learn useful patterns from examples or feedback Investigate data and produce evidence-based findings
Output An intelligent feature or application A trained predictive or classification model An analysis, chart, report, explanation or model
Training data Sometimes; not all AI uses ML Learns from data or interaction Sometimes; basic analysis needs no model training
Example Help a learner troubleshoot a computer problem Categorise new help-desk messages Find the most common help-desk problems
Methods Search, logical reasoning, rules, ML and more Classification, regression, clustering, reinforcement learning Cleaning, descriptive statistics, SQL, visualisation, modelling
Quality check Is the system useful, safe and accurate for its task? Does it perform well on genuinely new cases? Are the data and conclusions trustworthy?

On a small screen, swipe horizontally to read the full table.

The labels describe fields and methods, not three mutually exclusive products. A trained ML classifier is already an example of AI; putting it inside an assistant does not magically make it AI for the first time.

What is Artificial Intelligence?

AI concerns the design of systems that perform tasks we associate with intelligence. A language assistant may interpret a question and produce an explanation. A chess program may search possible moves. An image system may recognise an object. These are different approaches to intelligent behaviour. [1]

Some AI systems use rules or search rather than a model trained on examples. But do not call every automated task AI. A computer that sends an email at 4 p.m. every Friday simply follows a schedule; no intelligent capability is established by that action alone.

AI example: A school’s ICT assistant interprets a learner’s description of a problem, asks a suitable follow-up question and offers safe troubleshooting guidance. It may combine ML, written procedures and ordinary software rules.

What is Machine Learning?

In ML, a training process uses examples or feedback to build a model that can make predictions, classify information or identify patterns. Unlike a handwritten list of rules for every situation, the model learns relationships from the training process. [1] [3]

A common beginner example is spam filtering. The model studies messages labelled spam or not spam and then estimates a category for new messages. It can be wrong, especially when a new message differs from the examples it has seen.

Three ways of learning are useful to recognise:

  • Supervised learning: Learn from examples with known answers, such as labelled help-desk requests.
  • Unsupervised learning: Find patterns or groups without supplied answer labels, such as customers who have similar buying habits.
  • Reinforcement learning: Learn from actions and feedback, such as an agent improving at a game through rewards.

A responsible ML workflow defines the task, prepares examples, separates development and evaluation data, trains a model, tests it, examines mistakes and monitors it after deployment. A high score on the training examples alone is not evidence that the model will perform well in daily use. [3]

What is Data Science?

Data Science brings together data collection, cleaning, statistics, programming and subject knowledge to answer questions. It is not limited to building ML models. A useful project may end with an Excel chart and a clearly explained finding. [1] [4]

Consider a school asking which ICT problems students report most frequently. Someone must record the reports consistently, correct errors, count them, compare categories and explain what the counts mean. That is data-analysis work within a Data Science approach. An ML model is optional.

Data Science can also study uncertainty, build forecasts, run experiments and use ML. Its distinguishing feature is the question being investigated and the quality of the evidence, not the size of the dataset or whether a fashionable tool is involved.

How are AI, ML, Deep Learning and Data Science related?

The simplest accurate relationship is:

Artificial Intelligence ⊃ Machine Learning ⊃ Deep Learning.

That means ML is inside the broader field of AI, while Deep Learning is a specialised kind of ML using multilayer neural networks. Many modern Generative AI models use Deep Learning. [2]

Data Science is a partly overlapping field, not another small circle wholly inside AI. It includes statistical analysis and reporting without AI; it may also use ML models to make predictions. These boundaries are helpful teaching categories rather than universally fixed definitions. [1] [4]

Where does ChatGPT fit? ChatGPT is an AI application built around large language models trained using deep-learning methods. It is one application of AI, not the definition of the entire subject. [7]

One school ICT support desk: understand all three with the same records

Imagine a fictional school computer lab where students report problems to an ICT support desk. An employee records each message and places it into a category. This small learning dataset contains twelve reports:

Message Category
WiFi keeps disconnecting Network
No internet on my laptop Network
Wireless network is unavailable Network
The WiFi login is failing Network
Internet drops during lessons Network
Printer is out of paper Printing
The printer cannot print Printing
Printing is stuck in queue Printing
The printer shows an error Printing
I forgot my password Accounts
Password reset does not work Accounts
My user account is locked Accounts

On a small screen, swipe horizontally to read the full table.

All twelve messages are invented educational examples; no real student records are used.

Part 1: Data Science finds the most common problem

Count the three labels:

Problem Number of messages Share of all messages
Network 5 41.7%
Printing 4 33.3%
Accounts 3 25.0%
Total 12 100.0%

On a small screen, swipe horizontally to read the full table.

For example, the Network share is 5 ÷ 12 × 100 = 41.7% (rounded to one decimal place).

What have we learned? Network issues occur most often in these twelve examples. We cannot say they dominate across a school term: the sample is small, fictional and not randomly collected. The correct conclusion has to match the quality of the evidence.

This is the Data Science role: organise data, calculate the numbers and explain their meaning and limits.

Part 2: Machine Learning learns to categorise a new message

Suppose we want a program to suggest a category whenever a new report arrives. The twelve labelled sentences become training examples. A simple method converts words into numbers and trains a text classifier.

For a more revealing experiment, we then give the model nine separate, hand-written examples not used to fit it. The model correctly classifies eight and misclassifies one:

New example Correct category Model’s result
WiFi is slow Network Network
No connection to internet Network Network
Wireless is missing Network Network
Printer needs paper Printing Printing
Printing job failed Printing Printing
My document is stuck in the print queue Printing Printing
Forgot account password Accounts Accounts
My login needs resetting Accounts Network — incorrect
Cannot access user account Accounts Accounts

On a small screen, swipe horizontally to read the full table.

The incorrect result is instructive. A training message used the phrase “WiFi login”, so the word login may pull the classifier towards Network even when the user actually means an account. The model has learnt word patterns, not the full meaning of every sentence.

Do not interpret eight correct answers out of nine as a dependable 88.9% real-world accuracy. This is a tiny demonstration with hand-selected inputs, not an independent or representative test. A genuine evaluation would require many more examples and carefully protected validation and test sets. [3]

One more failure: ask about “My keyboard is broken.” The model outputs Network even though a keyboard fault belongs to none of the three defined categories. A real application should recognise unfamiliar cases and offer an Other / needs human review route, supported by suitable training data and evaluation.

This is the Machine Learning role: learn a pattern from examples and make a prediction — while revealing where that prediction can fail.

Part 3: AI builds an intelligent support experience

An ICT assistant might combine the classifier with a troubleshooting knowledge base and a conversational interface. If the model suggests a Network problem, the assistant could ask whether other devices connect to WiFi, explain the next safe check and offer to escalate the matter.

The classifier itself already counts as ML-based AI. The wider support assistant is an application of AI that may combine several technologies. The school might not need such an assistant at all if a simple form and an ICT technician solve the problem efficiently.

The distinction: Data Science tells the ICT team what kinds of problems are being reported; ML learns to suggest categories; an AI application may help interpret and respond to new requests.

Exercise 1: Analyse the data using Excel or Google Sheets

This activity works without Python. It is suitable for beginners learning percentages, spreadsheets and data interpretation.

  1. Open a blank spreadsheet. Enter Message in A1 and Category in B1.
  2. Copy the twelve reports above into A2:B13, preserving the order and labels.
  3. In an unused cell, enter =COUNTIF(B2:B13,"Network"). The answer should be 5.
  4. Replace Network with Printing, then Accounts. The answers should be 4 and 3.
  5. Calculate the Network share using =COUNTIF(B2:B13,"Network")/COUNTA(B2:B13) and format that cell as a percentage with one decimal place. The result should be 41.7%.
  6. Create a two-column summary table of categories and counts, highlight it and choose Insert → Chart (or Insert → Column Chart, depending on the application).
  7. Write one sentence describing the highest count and a second sentence explaining why the data cannot represent a whole school term.

COUNTIF counts cells that meet a condition; COUNTA counts non-empty cells. [5] If the spreadsheet uses semicolons to separate function arguments, replace the commas with semicolons.

Check your understanding: Have you trained an ML model? No. You performed data analysis. That work can be valuable on its own.

Exercise 2: Train a tiny text classifier in Python (optional)

If you have Python installed, install the library in your learning environment:

terminal · setup
python -m pip install scikit-learn

Now run the following complete example. The training data, evaluation examples and expected labels are stated openly so you can reproduce and challenge the results. CountVectorizer builds word-count features, and MultinomialNB learns a basic classifier. [6]

On a phone, swipe horizontally inside the code window to read long lines.

support_classifier.py
from sklearn.pipeline import make_pipelinefrom sklearn.feature_extraction.text import CountVectorizerfrom sklearn.naive_bayes import MultinomialNB training_messages = [ “WiFi keeps disconnecting”, “No internet on my laptop”, “Wireless network is unavailable”, “The WiFi login is failing”, “Internet drops during lessons”, “Printer is out of paper”, “The printer cannot print”, “Printing is stuck in queue”, “The printer shows an error”, “I forgot my password”, “Password reset does not work”, “My user account is locked”,]training_labels = [“Network”] * 5 + [“Printing”] * 4 + [“Accounts”] * 3 model = make_pipeline(CountVectorizer(), MultinomialNB())model.fit(training_messages, training_labels) test_cases = [ (“WiFi is slow”, “Network”), (“No connection to internet”, “Network”), (“Wireless is missing”, “Network”), (“Printer needs paper”, “Printing”), (“Printing job failed”, “Printing”), (“My document is stuck in the print queue”, “Printing”), (“Forgot account password”, “Accounts”), (“My login needs resetting”, “Accounts”), (“Cannot access user account”, “Accounts”),] correct = 0for message, expected in test_cases: predicted = model.predict([message])[0] is_correct = predicted == expected correct += is_correct status = “OK” if is_correct else “WRONG” print(f”{status}: {message} -> {predicted} (expected {expected})”) print(f”Correct on these examples: {correct} of {len(test_cases)}”)print(“Outside the categories:”, model.predict([“My keyboard is broken”])[0])

Verified demonstration result: eight of the nine prepared examples are classified correctly. “My login needs resetting” is misclassified as Network, and the out-of-category keyboard problem is also assigned Network. This output was reproduced with scikit-learn 1.8.0; outputs may differ if data, software or preprocessing are changed.

What should you improve next?

Instead of adding more rules to hide the two failures, investigate them. Add more varied, properly labelled messages, include an Other category where appropriate, and evaluate the model against realistic reports it has not already been tuned on. The goal is not to obtain a flattering classroom score; it is to build a system that fails safely when the question is unfamiliar.

If Python reports No module named sklearn, check that the package was installed into the same Python environment used to run the script. See the official installation guide.

Two further examples from everyday life

A maize farmer

Data Science may compare rainfall, soil records and harvest quantities. ML may be used to estimate future yields from sufficiently good historical data. An AI assistant may explain estimates or help a farmer navigate agricultural information. A prediction should not be mistaken for a confirmed crop diagnosis.

A matatu operator

Data Science may reveal which departure times carry the most passengers. ML may estimate future demand using repeated observations. An AI scheduling tool might combine those estimates with route, capacity and timing constraints. If a spreadsheet already answers the operator’s question, a new ML project may be unnecessary.

Neither scenario requires all three approaches. Start with a real question and choose the simplest method that answers it reliably.

How to choose the right approach for a problem

Your problem Start here Consider moving further when…
How many reports came in this week? Spreadsheet or SQL analysis You need deeper patterns and explanation
Which problem category will a new message belong to? Machine Learning classification You have representative labelled examples and a real need to predict
Can software talk the learner through a problem? AI application design You can verify answers, permissions and safe escalation
Should a reminder go out each Friday? Ordinary automation No AI is required for a fixed schedule

On a small screen, swipe horizontally to read the full table.

Before making a solution more complicated, ask what error would matter. A wrong chart can lead to a poor decision; an incorrect classification can send someone to the wrong support queue; a misleading AI answer can waste time or expose private information. The right evaluation depends on the use case. [3] [8]

Skills to learn first

For a complete beginner, start with spreadsheet analysis, percentages and charts. Add SQL to query databases and Python for programming. Learn how data is cleaned, then explore elementary statistics, model training and independent evaluation. Deep Learning and AI applications become easier to understand once those foundations are in place.

This is a practical route, not a universal prerequisite list. Some beginner AI programmes teach the data and programming foundations as part of the course.

For a dedicated AI-versus-Data-Science course and career choice, use NIT’s existing guide: AI vs Data Science in Kenya: Which Should You Choose?. The purpose of this article is to explain the three technical concepts, not duplicate that course comparison.

Frequently asked questions

Is Machine Learning the same as AI?

No. ML is a branch of AI. The larger AI field also includes approaches that do not involve learning a model from data.

Is Data Science part of AI?

Not entirely. They overlap: Data Science can use ML, but data cleaning, statistics and charts need not involve AI.

Is Deep Learning a separate field?

It is a specialised approach within Machine Learning, commonly based on multilayer neural networks. [2]

Is ChatGPT an example of AI or ML?

Both descriptions apply. ChatGPT is an AI application built around language models developed using Deep Learning. [7]

Can I learn Data Science without coding?

Yes, especially at beginner level. The spreadsheet exercise above is an example. Larger projects often benefit from Python and SQL.

Does a model always get better when given more data?

No. More examples do not fix incorrect labels, unrepresentative records or a poorly defined problem. Improvement must be checked through evaluation.

Which should I learn first?

Learn basic data handling and programming foundations, then explore ML and AI applications as your interests develop. You can start through either a suitable Data Science or AI course.

What to remember

A spreadsheet that counts help-desk reports gives us data analysis. A trained classifier that assigns a category to new messages demonstrates Machine Learning, a form of AI. A wider assistant that helps users troubleshoot shows how AI capabilities can be combined into an application.

The real skill is knowing why an approach works, what evidence supports its answer and when a simpler method would be better.

Learn by building real projects

Study AI or Data Science at Newton Institute of Technology

Develop practical skills in computing, Python, data preparation and applied projects at NIT in Migori Town, Kenya.

Artificial Intelligence: Learn Python, data preparation, Machine Learning and practical AI applications. Six months of institutional training followed by three months of industrial attachment.

Data Science & Analytics: Study Excel, SQL, Python, statistics, data cleaning, visualisation and introductory Machine Learning. Six months of institutional training followed by three months of industrial attachment.

Call / WhatsApp+254 713 584 858

Emailinfo@niteducation.com

LocationMigori Town, Kenya

Sources and references

  • [1] UK Defence Science and Technology Laboratory (2025). AI, Data Science and Machine Learning: A Dstl Biscuit Book — Introduction. Official guidance
  • [2] Google Cloud. Deep Learning vs Machine Learning vs AI. Technical explanation
  • [3] Google for Developers. Machine Learning Crash Course — Dividing the Original Dataset. Training, validation and testing
  • [4] Spector, A. (2024). Data Science and AI in Context: Summary and Insights. Harvard Data Science Review, 6(3). Research article
  • [5] Microsoft Support. Use the COUNTIF function; COUNTA function. COUNTIF · COUNTA
  • [6] Scikit-learn. CountVectorizer; MultinomialNB; installation guidance. CountVectorizer · MultinomialNB · Installation
  • [7] OpenAI Help Center. How ChatGPT and Our Foundation Models Are Developed. Official explanation
  • [8] National Institute of Standards and Technology. AI Risk Management Framework. Official framework

Sources reviewed: 10 November 2026. Editorial responsibility: Newton Institute of Technology Editorial Team.