Artificial intelligence can classify satellite images, model hazards, translate languages and generate convincing text. To use it well, you need more than good prompts: you need to understand how models are trained, where their outputs can fail, and when AI assistance becomes academic misconduct.
IGCSE & IB Diploma Programme
Use of AI
Understand how artificial intelligence developed, how modern models learn, where geography uses AI—and how to protect evidence, authorship and academic integrity.
Central question
How can we use AI as a powerful analytical tool without giving up evidence, independent thinking or academic integrity?
Knowledge
- major stages in AI history;
- machine learning, neural networks and transformers;
- training, validation and inference;
- strengths, limitations and uncertainty.
Skills and judgement
- evaluate AI-generated evidence;
- identify bias and spatial inequality;
- document AI assistance transparently;
- distinguish support from plagiarism.
1. AI is a field, not one machine
Artificial intelligence is a broad field concerned with computer systems that perform tasks associated with perception, language, reasoning, prediction, planning or content generation. The systems used today are usually narrow AI: they can be highly capable within a defined task but do not possess human consciousness, lived experience or general common sense.
2. Technical foundations
Training data and features
A dataset contains examples. A feature is an input variable, such as elevation, temperature, a pixel value or a word token. A label is a target category or value, such as “forest”, “urban” or a predicted rainfall total.
Parameters and loss
A model contains adjustable numerical parameters. During training, an algorithm compares predictions with known answers using a loss function. Optimisation methods adjust the parameters to reduce average loss. This improves statistical fit; it does not create awareness or guarantee causal understanding.
Training, validation and testing
Data should be separated. Training data adjust the model. Validation data support design choices. Test data estimate performance on unseen examples. Testing on training examples alone encourages overfitting: memorising detail instead of learning a pattern that generalises.
Neural networks
A neural network is a layered mathematical function. Each artificial neuron combines inputs, weights and a non-linear activation. Deep networks can build representations in stages—for example, from edges to shapes to land-cover classes.
Inference
After training, the model applies its learned parameters to new input. This stage is called inference. The result may be a class, numerical estimate, generated sequence or probability distribution.
Correlation is not causation
Machine-learning models often exploit statistical associations. A model may predict accurately while using the wrong signal. Researchers therefore examine data provenance, possible confounding factors and whether results remain reliable in a different place or time.
3. How a large language model produces an answer
- Tokenisation: input text is divided into tokens—whole words, word fragments or symbols.
- Embeddings: tokens are converted into numerical vectors that represent learned relationships.
- Transformer layers: attention mechanisms estimate which tokens are relevant to one another in context.
- Next-token probabilities: the model assigns probabilities to possible continuations and selects one according to its settings.
- Autoregressive generation: the process repeats token by token until the response is complete.
- Human verification: claims, calculations, sources and suitability must still be checked.
Large language models can produce fluent explanations because they learn complex language patterns. They can also generate a hallucination: a confident-looking statement, quotation or reference that is false. Fluency is therefore evidence of language modelling—not evidence that a claim is correct.
4. A short history of AI
- 1950 — Alan Turing: “Computing Machinery and Intelligence” replaced the vague question “Can machines think?” with the imitation game, later called the Turing test.
- 1955–1956 — Dartmouth: a research proposal used the term artificial intelligence and imagined machines using language, concepts and self-improvement. The summer project became a symbolic starting point for AI as a field.
- AI winters: early promises exceeded available computing power, data and methods. Funding and interest fell during periods in the 1970s and again around the late 1980s and early 1990s.
- 1997 — Deep Blue: IBM’s chess system defeated world champion Garry Kasparov in a rematch, demonstrating specialised search and computing power.
- 2012 — deep learning breakthrough: large neural networks trained with powerful processors dramatically improved image recognition and encouraged rapid investment.
- 2016 — AlphaGo: a system combining neural networks, search and reinforcement learning defeated Lee Sedol 4–1.
- 2017 — transformers: the paper “Attention Is All You Need” introduced a transformer architecture based on attention, becoming foundational for modern language models.
- 2022 — generative AI becomes widely accessible: conversational systems brought large language models to a mass audience and intensified debates about education, work, copyright, misinformation and power.
5. AI, geography and spatial data
Spatial bias matters
Data are rarely distributed evenly. Wealthier neighbourhoods may have more sensors, mapped buildings or online activity. Rural places and minority languages may be underrepresented. If absence of data is treated as absence of need, AI can reproduce spatial inequality.
How should performance be evaluated?
- Accuracy: the share of all predictions that are correct. This can be misleading when one class is rare.
- Precision: of the cases predicted positive, how many were correct?
- Recall: of the real positive cases, how many did the model find?
- Uncertainty: how confident is the estimate, and is that confidence calibrated?
- External validity: does the model work in another location, season or population?
- Fairness: are error rates and benefits distributed equitably among groups and places?
6. Critical questions for every AI system
- Purpose: What problem is being solved, and who defined it?
- Data: Where did the data come from? Who is missing? Was consent obtained?
- Method: Which model and evaluation measures were chosen, and why?
- Evidence: Is the output reproducible, sourced and independently checked?
- Impact: Who gains, who carries risk, and can a decision be challenged?
- Transparency: Can people tell when AI was used and understand its role?
- Resources: What computing, energy, water and hardware are required?
- Accountability: Which human or institution remains responsible?
7. Plagiarism, authorship and AI
IB and Cambridge International
The International Baccalaureate does not treat AI output as the student’s own work. AI-generated material included in work must be made clear and appropriately referenced. The IB emphasises transparency and ethical use rather than a general ban. See the official IB statement on AI in assessment and education and the IB Academic integrity policy.
Cambridge International states that inappropriate use of generative AI to create or enhance coursework without acknowledgement risks being treated as plagiarism or malpractice. Teachers must be able to authenticate submitted coursework. See Cambridge International guidance on generative AI in coursework.
Always follow the most recent instructions from your teacher, school and examination board. Rules may differ between practice, homework, coursework and examinations.
Your tasks
Quick check
- Distinguish training from inference.
- Why can high accuracy hide weak performance?
- What is overfitting?
- How does spatial bias enter a geographic model?
- Why can a language model hallucinate a reference?
- When can AI-generated content become plagiarism?
Open the answer guide
1. Training adjusts parameters using examples; inference applies fixed learned parameters to new input. 2. In an imbalanced dataset, predicting the common class can produce high accuracy while missing rare but important cases. 3. Overfitting occurs when a model fits training detail but generalises poorly. 4. Sensors, labels, internet activity and samples may cover places or groups unevenly. 5. It predicts statistically plausible tokens and does not automatically verify a source. 6. When AI-generated material is presented as the student’s own or is used without the disclosure and referencing required by the assessment rules.
Primary and official sources for further reading
- Alan Turing Digital Archive: “Computing Machinery and Intelligence” (1950)
- Dartmouth Summer Research Project proposal (1955)
- IBM Research: the Deep Blue project
- Google DeepMind: AlphaGo
- Vaswani et al.: “Attention Is All You Need” (2017)
- Introducing ChatGPT (30 November 2022)
- International Baccalaureate: AI, assessment and academic integrity
- Cambridge International: generative AI in coursework