Back
Keentune
AI Literacy curriculum 26 chapters
·
192 concepts
·
free
Everything the adaptive question bank can teach and test in AI Literacy, from foundations through advanced practice. Work through it in order, or start practising and let the questions find your level.
Start practising AI Literacy
New here? Read the AI Literacy guide
A free 16-minute primer — the mental model, the mistakes beginners make, and what to practise first.
•
AI names a family of techniques, not one product.
•
Hand-written logic versus patterns learned from data.
•
Systems good at one task versus the general intelligence they are not.
•
Statistical association is not comprehension or belief.
•
Why "it thinks / wants / knows" misleads about what happened.
•
Deterministic automation and learned models fail differently.
•
Recommendations, filters, pricing and scoring you already touch.
B. How machine learning learns
•
Building the model once, then running it many times.
•
Learning a mapping from examples someone labelled.
•
Finding groupings without an answer key.
•
Learning from reward signals rather than labels.
•
What "the model" physically is after training.
•
Memorising the training set instead of generalising.
•
Why performance must be measured on unseen data.
•
Aligning behaviour with preference data after pretraining.
C. Data: the fuel and the failure mode
•
A model inherits the defects of what it was trained on.
•
Who is in the data determines who it works for.
•
Ambiguous or biased labels become model behaviour.
•
Why a model does not know about events after its data ends.
•
Where training data came from, and why that is contested.
•
Models can reproduce fragments of their training data.
•
Machine-made training data and the drift it can introduce.
•
The world changing out from under a deployed model.
D. Language models under the hood
•
The core mechanism behind fluent text generation.
•
Text split into pieces, and why that affects counting and cost.
•
The working span a model can attend to at once.
•
Continuity is a product feature, not a model property.
•
Why the same prompt can produce different answers.
•
Layered instructions and which one wins.
•
Giving a model documents instead of trusting recall.
•
When a model calls a calculator, search or database instead of guessing.
•
Step-by-step generation as more computation, not new knowledge.
E. Generative systems beyond text
•
How prompts become images, and the artefacts that remain.
•
Synthesis and cloning from short audio samples.
•
Temporal consistency as the current hard problem.
•
Plausible code that compiles is not verified code.
•
Models that read images, audio and documents together.
•
Targeted modification of real media versus whole synthesis.
•
Reproducing a recognisable style, and the questions that raises.
F. Reliability, hallucination and confidence
•
Fluent, confident output that is simply not true.
•
Generation optimises plausibility, not truth.
•
Invented sources, authors and page numbers as a signature failure.
•
Tone carries no information about correctness.
•
Where a wrong answer costs enough to require an expert.
•
Agreeing with the user rather than correcting them.
•
Distinguishing "cannot" from "will not" from "does not know".
•
Reproducibility problems when the same input varies.
•
Naming the deliverable, audience and format explicitly.
•
Pasting the source beats hoping the model remembers it.
•
Bounding length, structure and what to leave out.
•
Showing one worked example instead of describing it.
•
Splitting a large task into checkable steps.
•
Treating the first output as a draft to be revised.
•
Requesting what the model is unsure about.
•
Why no prompt wording can make an unknown fact known.
•
Treating output as a claim requiring evidence.
•
Opening the source rather than trusting the reference.
•
Confirming against a source the model did not supply.
•
Recomputing figures instead of trusting stated math.
•
When checking costs more than doing it yourself.
•
Asking the model to check itself is weak evidence.
•
Where a human specialist is the only acceptable check.
•
Logging prompt, output and check for accountable work.
I. Evaluation and measurement
•
What a benchmark score does and does not tell you.
•
Test items leaking into training data inflate scores.
•
General capability does not transfer to your specific task.
•
Reading the failures, not just the aggregate accuracy.
•
Accuracy hides the cost asymmetry between error types.
•
Rating rubrics, agreement, and their expense.
•
Performance measured after deployment, not only before.
•
Data, design choices and deployment context each contribute.
•
A model faithfully reproducing an unjust past.
•
Worse performance for groups thinly present in the data.
•
Neutral-looking features that stand in for protected traits.
•
Competing fairness criteria that cannot all hold at once.
•
Model outputs shaping the data that trains the next model.
•
Where automated decisions produce unequal outcomes.
•
What debiasing techniques can achieve, and what they cost.
K. Privacy and confidentiality with AI
•
Input leaves your device and may be retained.
•
Whether a service learns from what you submit.
•
Client, patient, employee and student data boundaries.
•
Models deducing attributes nobody disclosed.
•
Why removing names does not make data anonymous.
•
What a provider keeps, for how long, and who can see it.
•
Where the computation happens changes the exposure.
•
Untrusted content issuing instructions the model obeys.
•
Malicious instructions hidden in a page, file or email the model reads.
•
Corrupting training data to plant behaviour.
•
Inputs crafted to make a classifier decide wrongly.
•
Probing a deployed model to recover what it learned.
•
Weights, datasets and libraries as trusted-by-default components.
•
Limiting what an AI-driven action can reach.
•
Approval before irreversible automated actions.
M. AI-enabled threats and misuse
•
Cheap, fluent, personalised fraud at volume.
•
Cloned voices and faces used against people who trust them.
•
Synthetic accounts and content manufacturing consensus.
•
Fake tools and "AI" claims wrapped around a scam.
•
The same capability serving defence and attack.
•
Code words, callbacks and channel checks as the personal defence.
•
Where AI-enabled fraud and abuse get reported.
N. Trustworthy AI principles
•
Working as intended, demonstrably, over time.
•
Not creating unacceptable risk to people or property.
•
Withstanding attack and degrading gracefully.
•
Someone answerable, and enough disclosure to check.
•
What a system did, in terms a person can use.
•
Designing for data minimisation from the start.
•
Harmful bias treated as a managed risk, not an accident.
•
Why maximising one characteristic can cost another.
•
Govern, map, measure and manage as a repeatable cycle.
•
Knowing where AI is actually being used in an organisation.
•
Writing down what may and may not be sent to a model.
•
Matching oversight to the consequence of being wrong.
•
Who owns the decision when the system is wrong.
•
Questions to ask a provider before adopting a tool.
•
Reporting, investigating and correcting AI failures.
•
Retiring a model that no longer performs or fits.
P. Human oversight and automation bias
•
Over-trusting a machine answer because it is a machine answer.
•
Nominal human approval that adds no real check.
•
Losing the ability to do the task the tool now does.
•
Oversight designed so a human can actually intervene.
•
Deciding in advance what must go to a person.
•
A route to contest an automated decision.
•
Warmth and fluency inflating perceived competence.
•
Which tasks AI genuinely accelerates and which it degrades.
•
Using generation for a first pass, never a final answer.
•
When to say AI was used, and to whom.
•
Working inside employer, client and regulatory constraints.
•
Substantiating what you say your product's AI does.
•
Extra scrutiny where AI touches decisions about people.
•
Measuring real throughput, not perceived speed.
•
Who is responsible for an AI-assisted deliverable.
R. Copyright, attribution and AI output
•
Why authorship requires human creative contribution.
•
The open questions about using protected works to train.
•
Generated work that reproduces a protected original too closely.
•
Reading what a service grants you in its output.
•
Crediting sources and disclosing generation honestly.
•
Brands and real people appearing in generated media.
•
An academic-integrity problem and a legal one are different.
S. Inventions, disclosure and AI
•
A named inventor must be a natural person.
•
Significant human contribution as the test.
•
Submitting an idea to a service can be a disclosure.
•
Confidentiality lost the moment it is pasted somewhere.
•
Recording who did what, for later proof.
•
Useful assistance that still requires verification.
T. Synthetic media and provenance
•
What can now be synthesised, cheaply and quickly.
•
Why detectors are unreliable and produce false accusations.
•
Embedded provenance signals and how they break.
•
Missing provenance data does not make content fake or real.
•
Real evidence dismissed as AI-generated.
•
Tracing to the original publisher rather than analysing pixels.
•
Using someone's face or voice without permission.
U. AI agents and tool use
•
A model given tools, goals and the ability to act.
•
Multi-step execution, and where it goes off the rails.
•
Granting the narrowest access the task requires.
•
Sending, paying, deleting and posting need a human gate.
•
Small mistakes amplified across a chain of steps.
•
Being able to see what the agent actually did.
•
Tasks where the failure cost exceeds the time saved.
V. Choosing and adopting AI tools
•
Starting from the task, not from the tool.
•
A bounded trial with a success criterion agreed in advance.
•
Retention, training use, subprocessors and location.
•
Subscription, verification labour and rework together.
•
Getting your prompts, data and outputs back out.
•
Tools that work for everyone who must use them.
•
Adoption fails on skills and norms, not on the model.
W. Policy, standards and the wider landscape
•
Rules scaled to the harm an application can cause.
•
Existing law in health, credit, employment and safety already applies.
•
Where telling people AI was involved is becoming mandatory.
•
Voluntary frameworks and technical standards as soft governance.
•
The same product facing different rules by geography.
•
What voluntary commitments can and cannot deliver.
•
Following the rule change rather than the product launch.
X. Compute, cost and footprint
•
One large build cost versus recurring per-use cost.
•
Specialised accelerators as the bottleneck resource.
•
Data-centre power and cooling as a real physical footprint.
•
Smaller models trading capability for speed and cost.
•
Distillation, quantisation and caching as cost levers.
•
Not using the largest model for a trivial task.
Y. People, trust and public understanding
•
What people believe AI can do versus what it does.
•
Overclaiming followed by backlash, and how to read both.
•
Task change rather than clean job replacement.
•
Who benefits from AI tools, and who is left out.
•
Learning with AI without outsourcing the learning.
•
Emotional reliance on a conversational system.
•
People affected by a system having a say in it.
Z. Talking about AI accurately
•
Using "model", "system", "product" and "agent" precisely.
•
Distinguishing demonstrated from asserted from projected.
•
Reading a claimed score for the conditions behind it.
•
One impressive output is not a capability result.
•
Describing what broke, precisely enough to be actionable.
•
Saying "unverified" instead of hedging into meaninglessness.
•
Teaching a non-technical audience without overclaiming.
Keentune is not affiliated with or endorsed by the organizations whose documentation informs these maps.
Start practising AI Literacy
All about AI Literacy practice
All exam, test, and product names and trademarks are the property of their respective owners and are used here for identification and reference only. Keentune is independent study practice — not affiliated with, authorized, or endorsed by any of these organizations.
© 2026 SportaApp LLC