Everything the adaptive question bank can teach and test in AI Literacy, from foundations through advanced practice. Work through it in order, or start practicing and let the questions find your level.
The word covers techniques with little in common — a rules engine, a recommender and a language model fail in unrelated ways. Asking which technique is meant is the first step in any claim.
Written rules do exactly what they say and can be read; learned behavior is inferred from examples and can only be inspected indirectly. That difference decides how each is debugged.
Every system in use today is built for a bounded set of tasks, and performance does not carry across the boundary. Impressive results inside the boundary say nothing about outside it.
Producing the right output does not require holding a belief about the world, so fluent correctness and comprehension are separate things. The system can be right without knowing anything.
Words like wants or knows import a model of an agent with goals and beliefs, which changes what you expect and what you check. The vocabulary shapes the trust before any evidence does.
Deterministic automation fails the same way every time, which makes it testable; a learned model fails on inputs nobody anticipated. The second needs monitoring rather than a test suite.
Ranking, filtering, pricing and scoring have used learned models for years without being labeled as such. Most exposure is to systems nobody calls AI.
B. How machine learning learns
Building the model is a one-off, expensive process; running it happens repeatedly and cheaply per use. Almost every cost, latency and update question turns on which of the two you mean.
Learning from labeled examples means the ceiling is the labels — the model learns to reproduce the judgment embedded in them, including where that judgment was wrong.
With no answer key, the system finds groupings in the data itself, which means the groups are whatever the data makes salient rather than what you wanted them to be.
Learning from a reward signal optimizes whatever the signal measures, so a poorly chosen reward produces behavior that scores well and misses the point.
The trained model is a large set of numbers, so there is no readable rule to inspect and no line to edit. What was learned exists only as those values.
A model can score perfectly on what it has seen and poorly on anything new, which means training performance is not evidence of anything. It is the reason a held-out set exists.
Performance must be measured on data the model has not seen, because measuring on the training data measures memorization. Any leak between the two invalidates the number.
After pretraining, preference data shapes how a model responds — tone, refusals, helpfulness. This stage changes behavior rather than adding knowledge.
C. Data: the fuel and the failure mode
A model has no independent access to the world, so every defect in its training material is available to be reproduced. Data quality is a ceiling rather than a starting point.
Performance tracks who is present in the data, so a group thinly represented gets worse results. The system is not neutral across people; it is as good as its coverage.
Where labels encode a judgment, the model learns the judgment, including its inconsistencies. Disagreement between labellers becomes noise the model must resolve somehow.
Knowledge comes from data gathered up to a point, so anything after it is absent. A model discussing a later event is reconstructing rather than recalling.
Where training data came from, and on what terms, is contested and frequently undisclosed. It is an open legal and ethical question rather than a settled technical detail.
Models can reproduce fragments of what they were trained on, particularly text that appeared often or is distinctive. Training on sensitive material is therefore a disclosure risk.
Machine-generated training data is cheap and inherits the generator's blind spots, so repeated use can narrow variety rather than expand it.
A deployed model sees a world that keeps changing, so performance decays without anything about the model changing. Detecting it requires measuring after deployment.
D. Language models under the hood
Text is generated one piece at a time, each chosen given what came before. Fluency is a property of that process alone, so it survives intact when the content happens to be wrong.
Text is split into pieces that are not words, which is why length limits, counting tasks and cost are all measured in units the user cannot see directly.
There is a bounded span the model can consider at once, and material outside it has no influence. Longer conversations push earlier content out of that span.
Continuity between sessions is something a product builds by resending information, not something the model possesses. Absent that, each conversation starts from nothing.
Output is sampled rather than selected deterministically, so the same input can produce different results. Lower variability trades novelty for repeatability.
Instructions arrive in layers with different authority, and conflicts between them are resolved by the system rather than by whoever wrote last. Knowing the layering explains unexpected refusals.
Supplying the relevant documents alongside the question replaces recall with reading, which is why it addresses both staleness and fabrication. The retrieval step becomes the weak link.
A model can call an external tool rather than generate an answer, which converts a guess into a computation. What matters is whether the tool was actually called.
Generating intermediate steps spends more computation on a problem; it does not add information the model lacks. It helps with working-out, not with knowing.
E. Generative systems beyond text
Images are produced from a description rather than assembled from stored pictures, which is why fine detail — text, hands, symmetry — is where the process shows.
A convincing voice can be produced from a short sample, which removes voice as a basis for identifying someone. The consequence is social rather than technical.
Keeping objects and lighting consistent across frames is much harder than producing one good frame, so inconsistency over time remains the visible limitation.
Generated code that compiles has satisfied the compiler, not the requirement. Plausibility is what was optimized, so correctness still has to be established separately.
A model that accepts images, audio and documents together can relate them, which widens what it can be asked and also widens what it can misread.
Modifying a real photograph or recording raises different questions from producing one wholesale, because the result carries the authority of the original. Provenance matters more in the first case.
A recognizable style can be reproduced without copying any specific work, which is precisely what makes the legal and ethical position unsettled rather than obvious.
F. Reliability, hallucination and confidence
The output is fluent, specific and confident, and simply untrue. The absence of any signal distinguishing it from correct output is the whole problem.
Generation optimizes for what plausibly follows, and a false statement can be more plausible than an admission of ignorance. It is the mechanism working, not malfunctioning.
Invented sources are the signature failure because a citation is exactly the kind of well-formed pattern the process produces easily. Checking that it exists takes seconds.
Tone is generated alongside content and is not derived from any internal measure of certainty. A hedge and an assertion carry the same evidential weight, which is none.
Where a wrong answer causes real harm — medical, legal, financial, safety — the verification burden rises to the point where an expert is the only acceptable check.
Agreement is rewarded during tuning, so a model may follow a user's framing rather than correct it. Pushing back on a correct answer can make it change.
A refusal can mean the capability is absent, the policy forbids it, or the information is unknown. Treating all three as the same loses the signal.
The same input can yield different output, which breaks reproducibility for anything audited or tested. Recording the exact output matters more than recording the prompt.
G. Prompting well
Naming the deliverable, the audience and the format removes the guesses the model would otherwise make. Most disappointing output is an unstated expectation.
Pasting the source is more reliable than depending on recall, because the material is then present rather than reconstructed. It also makes the answer checkable against what you supplied.
Bounding length and structure makes output usable without editing, and saying what to leave out is often more effective than saying what to include.
One worked example conveys a standard that a paragraph of description does not, because it demonstrates the judgment rather than asserting it.
Splitting a large task into steps produces intermediate results that can each be checked, which converts one unverifiable output into several verifiable ones.
Treating the first result as a draft is the normal workflow, not a sign of failure. The value is in the revision loop rather than in the first response.
Requesting what is unsure surfaces weak points that would otherwise be stated in the same confident register as everything else. It is a partial control, not a guarantee.
No phrasing makes an unknown fact known or an absent document present. Prompting improves how a task is performed, never what information exists.
H. Verifying AI output
Output is a claim, and a claim carries no evidence with it. Treating it as a starting point rather than an answer is the whole discipline, and it is a habit rather than a technique.
Opening the source establishes both that it exists and that it says what was claimed — two separate failures. A reference that resolves can still not support the sentence.
Confirming against a source the model did not supply breaks the circularity of asking the same system twice. A second generation is not a second opinion.
Figures are generated in the same way as prose, so a stated calculation is a plausible-looking string. Recomputing costs seconds and catches a class of error entirely.
Where checking takes longer than doing the task, the tool has not helped. That comparison is the honest test of whether a use case works.
Asking a model to review its own output uses the same process that produced it, so shared errors survive. Agreement with itself is weak evidence.
Some claims cannot be checked by a non-specialist at all, which makes expert review the only acceptable control rather than an added precaution.
Logging the prompt, the output and the check performed is what makes AI-assisted work accountable later. Without it, nobody can establish what was verified.
I. Evaluation and measurement
A score describes performance on a specific set of items under specific conditions. It supports a comparison between systems on that set and little else.
If test items appeared in training data, the score measures memorization. This is difficult to rule out and routinely undermines published comparisons.
General capability does not transfer to a particular workflow, so the only reliable evidence is a trial on your own examples. A leaderboard position predicts little.
Reading the failures shows whether errors are random or concentrated in a pattern, which decides whether the system is usable with a workaround. Aggregate accuracy hides that.
Where a false positive and a false negative cost different amounts, a single accuracy figure averages them and obscures the one that matters.
Judgment-based rating is often the only valid measure and is expensive, needs a rubric, and needs agreement between raters to mean anything.
Pre-deployment testing measures a fixed snapshot; real inputs shift. Measurement after deployment is what detects degradation, and it needs to be built in.
J. Bias and fairness
Bias enters through the data, through design choices such as which metric to optimize, and through the context of deployment. Fixing one leaves the others.
A model trained on past decisions reproduces them accurately, including the unjust ones. The system is working correctly and producing the wrong outcome.
Performance is worse for groups thinly present in the training data, so the failure is unevenly distributed and invisible in an overall accuracy figure.
A feature that looks neutral can stand in for a protected characteristic, so removing the protected field does not remove the effect. Postcode is the classic example.
Several reasonable fairness criteria cannot be satisfied simultaneously except in special cases, so choosing one is a value judgment rather than a technical optimization.
A model's outputs influence what happens next, which becomes the data for the next model. The initial bias is then reinforced rather than corrected.
A rule applied uniformly can still produce unequal outcomes, which is where automated decisions meet existing law. Intent is not the test.
Techniques exist and each costs something — accuracy, simplicity, or the ability to explain the result. There is no adjustment that is free.
K. Privacy and confidentiality with AI
Input leaves the device, reaches a provider and may be retained, so pasting is a disclosure. The convenience of the interface disguises that.
Whether submitted content is used to improve a service varies by provider and plan, and it is a question to answer before use rather than after.
Client, patient, employee and student information carry obligations that follow the data wherever it goes. The obligation is unchanged by the tool being convenient.
A model can deduce attributes nobody disclosed by combining innocuous signals, so privacy is not secured by withholding the sensitive field alone.
Removing names does not make a record anonymous, because combinations of ordinary attributes are frequently unique. Re-identification is a known and repeated result.
What is kept, for how long, and who can see it are provider-specific facts that determine the real exposure. A deletion control may not reach backups or logs.
Where computation happens decides whether data leaves at all, which is the largest single lever on exposure. It usually trades capability for containment.
L. Securing AI systems
Content the model reads can contain instructions, and the model has no reliable way to distinguish data from directive. Anything untrusted that enters the context is a potential instruction.
The instruction need not come from the user: a web page, document or email the model reads can carry it. This is why an assistant with browsing or file access has a wider attack surface.
Corrupting training data plants behavior that appears only under specific conditions, which makes it hard to detect by ordinary evaluation.
Inputs can be crafted so a classifier decides wrongly while looking normal to a person, which matters wherever a model is used as a filter or a gate.
Repeated querying can recover training content or approximate the model itself, so a deployed model exposes more than its answers.
Weights, datasets and libraries are frequently downloaded and trusted without verification, which makes them the same kind of dependency risk as any other package.
An automated action can only cause harm it has access to cause, so scoping permissions narrowly is the primary control rather than a supplementary one.
Irreversible actions need approval before they happen, because a review afterwards cannot undo them. The gate must be placed at the irreversible step specifically.
M. AI-enabled threats and misuse
Fluent, personalized approaches can now be produced at volume, which removes the language errors that once made fraud recognizable. The defense shifts from spotting mistakes to verifying identity.
Cloned voices and faces are used against people who trust the person being imitated, so the attack targets the relationship rather than any technical weakness.
Synthetic accounts and generated content can manufacture the appearance of consensus, which exploits the assumption that volume reflects genuine opinion.
Some products marketed as AI do nothing of the sort, and the label itself is used to justify claims that would otherwise be questioned.
The same capability serves attack and defense, which is why restricting the technology is a poor lever and detection and verification receive the attention instead.
Agreed code words, calling back on a known number and moving to a different channel are the practical defenses, because they do not depend on detecting the fake.
AI-enabled fraud is reported through existing fraud and abuse channels, and knowing where those are in advance matters because the useful window is short.
N. Trustworthy AI principles
Working as intended has to be demonstrated rather than asserted, and sustained over time rather than shown once. Both halves are part of the characteristic.
Not creating unacceptable risk to people or property, which requires deciding what is unacceptable in this context before the system is deployed.
Withstanding attack and failing gracefully are different properties, and a system can have one without the other. Degradation behavior is a design choice.
Someone must be answerable for the outcome, and enough must be disclosed for the answer to be checkable. Accountability without disclosure is unverifiable.
An explanation has to be usable by the person who needs it, which means a technical account can satisfy neither a regulator nor an affected individual.
Designing for minimal data from the start is far more effective than protecting data already collected, because the safest data is the data never gathered.
Harmful bias is treated as a risk to be measured and managed continuously, not an error to be removed once. It recurs as data and context change.
Maximizing one characteristic costs another — explainability against accuracy, privacy against evaluation. A system claiming all of them at once has not been examined closely.
O. Governing AI use
Governing, mapping, measuring and managing form a cycle rather than a checklist, so the output of managing feeds back into mapping. Treating it as a one-off assessment misses the point.
Most organizations cannot say where AI is already in use, and you cannot govern what you have not enumerated. The inventory precedes every other control.
Writing down what may be sent to a model converts an individual judgment into an organizational one, which is what makes it enforceable and teachable.
Oversight should scale with the consequence of being wrong, so a drafting aid and a decision affecting someone's livelihood do not deserve the same process.
Someone must own the decision when the system is wrong, named in advance. Discovering that nobody owns it during an incident is the common failure.
The questions worth asking are about data handling, retention, subprocessors, evaluation evidence and what happens on termination — not about model capability.
AI failures need a reporting route, an investigation method and a correction step, the same as any other incident class. Without one, failures are absorbed rather than fixed.
Retiring a model that no longer performs or fits is part of the lifecycle, and systems tend to persist by default because nobody owns their removal.
P. Human oversight and automation bias
People defer to a machine answer more than to an equivalent human one, so adding a human reviewer does not automatically add a check. The bias works against the control.
Approval that happens too fast, with too little information, or with no realistic option to refuse, adds a signature and no scrutiny. It is worse than no review because it creates a record of one.
Capability that is not exercised declines, so a reviewer may gradually lose the ability to catch what they are reviewing. The oversight degrades silently.
Control requires the time, information and authority to intervene. Any of the three missing makes the oversight nominal regardless of the process on paper.
Deciding in advance what must reach a person removes the judgment from the moment of pressure, when it is least reliable.
A route to contest an automated decision is what makes an error correctable for the individual affected rather than only for the next one.
Warmth and fluency raise perceived competence independently of actual accuracy, so a well-written wrong answer is trusted more than a blunt correct one.
Q. AI at work
Some tasks are genuinely accelerated and others are made slower once verification is counted. Knowing which is which for your own work is the skill.
Generation for a first pass and human judgment for the final answer is the pattern that holds up. Inverting it is where the visible failures come from.
Whether to say AI was used depends on audience and context, and the trend is toward disclosure where the output is presented as your own judgment.
Employer, client and regulatory constraints all apply simultaneously, and the strictest governs. Personal comfort with a tool is not the operative standard.
Saying a product uses AI, or what its AI does, is a claim that must be substantiable. Regulators have treated unsupported capability claims as ordinary deceptive advertising.
Where AI touches decisions about people, existing employment law applies in full and the scrutiny is higher. The technology does not create a separate legal category.
Perceived speed and actual throughput diverge, particularly once rework and verification are counted. Measuring the finished output is the only honest comparison.
The person who signs off is responsible for what it says, regardless of how it was produced. Attribution to a tool does not transfer accountability.
R. Copyright, attribution and AI output
Copyright protection has been tied to human creative contribution, so output produced without it may not be protectable. That is a question about your rights, not about permission.
Whether training on protected works is permitted is genuinely unsettled and litigated, so confident statements in either direction should be treated with suspicion.
Generated work can reproduce a protected original closely enough to infringe, independently of intent. The risk sits with whoever publishes it.
What a service grants you in its output, and what it reserves, is in the terms and varies between providers. It is worth reading before commercial use.
Crediting sources and disclosing generation honestly is the practice that survives scrutiny later, and the cost of doing it is trivial compared with being found not to have.
Brands and real people appearing in generated media raise separate issues from copyright, governed by different law, and consent is the operative question for a person.
Passing off work as your own is an integrity matter decided by an institution; copying protected expression is a legal one. An act can be either, both or neither.
S. Inventions, disclosure and AI
A named inventor must be a natural person, so a system cannot be listed. This has been decided in several jurisdictions and is not an open question.
Use of a tool does not preclude patenting; what matters is whether a person made a significant contribution to the conception. The test is about the human's role.
Submitting an idea to a service can count as disclosure depending on the terms, which can affect patentability. The risk is created before anything is published.
Confidentiality is lost at the moment information is pasted somewhere not covered by an obligation, and it cannot be restored afterwards. The loss is immediate and permanent.
Recording who contributed what, and when, is what supports an inventorship claim later. Reconstructing it after the fact is unreliable and unconvincing.
Assistance in searching is genuinely useful and produces results that must be verified, because a missed or invented reference has consequences for the filing.
T. Synthetic media and provenance
Convincing synthetic media can be produced cheaply and quickly, so the question has shifted from whether something could be faked to whether this one was.
Detectors are unreliable, degrade as generators improve, and produce false accusations against real material. A detector result is not evidence on its own.
Embedded provenance signals help when present and survive poorly through screenshots, re-encoding and cropping. They are a partial measure by design.
Missing provenance data says nothing either way, since most genuine content carries none. Treating absence as evidence of fakery is a reasoning error with real consequences.
The existence of convincing fakes lets genuine evidence be dismissed as fabricated, which benefits whoever is caught. The harm runs in both directions.
Tracing to the original publisher is more reliable than analyzing the file, because provenance is a question about where it came from rather than about pixels.
Using someone's face or voice without permission is the core harm, and it is a question of consent rather than of whether the result is detectable.
U. AI agents and tool use
A model given tools, a goal and the ability to act becomes something that changes the world rather than producing text. The risk profile changes with it.
Multi-step execution goes wrong when an early step's error is treated as established fact by later steps. The loop has no independent check unless one is built in.
Granting the narrowest access the task requires bounds the worst case, which is the only control that holds when the reasoning is wrong.
Sending, paying, deleting and posting cannot be undone, so they need a human gate specifically. Reversible actions can be corrected; these cannot.
Small errors amplify across a chain because each step builds on the last, so a long sequence is far less reliable than its per-step accuracy suggests.
Being able to see what was actually done — which tools, which inputs, which results — is what makes an agent debuggable and auditable. Without it, only the outcome is visible.
Where the cost of a failure exceeds the time saved, automation is a bad trade regardless of how well it usually works. The comparison is expected value, not typical case.
V. Choosing and adopting AI tools
Starting from a tool produces a search for somewhere to apply it; starting from a task produces a testable question. The order determines whether the pilot can fail.
A bounded trial with a success criterion agreed beforehand is what allows a decision either way. Without the criterion, every result reads as encouraging.
Retention, training use, subprocessors and location are the four questions that decide whether a tool is usable for your data. They are answerable before adoption.
The subscription is the visible part; verification labor and rework are usually larger. A comparison that counts only the license overstates the saving.
Getting prompts, data and outputs back out is a question to ask at the start, because the answer is hardest to act on once you depend on the tool.
A tool everyone must use has to work for everyone who must use it, which includes assistive technology and varied language proficiency. Otherwise adoption creates exclusion.
Adoption fails on skills and norms far more often than on model capability, so the investment that decides the outcome is usually in people rather than licenses.
W. Policy, standards and the wider landscape
Rules scaled to potential harm means the same underlying technology faces different obligations depending on application. The use case determines the burden, not the model.
Existing law in health, credit, employment and safety applies to AI systems already, so most obligations predate any AI-specific statute. Nothing was waiting for permission.
Telling people AI was involved is becoming mandatory in defined situations, which turns disclosure from an ethical choice into a compliance obligation in those contexts.
Voluntary frameworks and technical standards operate as soft governance, influencing practice and procurement before any law compels them. They are worth tracking for that reason.
The same product faces different rules by geography, so a deployment decision is partly a legal-geography decision. Compliance in one place implies nothing about another.
Voluntary commitments can move practice quickly and cannot be enforced against whoever declines to make them, which bounds what they can deliver.
Following rule changes rather than product launches is the higher-value habit, because obligations change less often and matter more to what you may do.
X. Compute, cost and footprint
Training is a large one-time cost and serving is a recurring per-use one, so the economics of building and of running differ completely. Confusing them misreads every cost claim.
Specialized accelerators are the bottleneck resource, which shapes availability, pricing and who can participate at the frontier. It is a supply constraint before it is a technical one.
Data centers consume power and cooling water at a scale that is a real physical footprint, which is a siting and environmental question rather than an abstract one.
Smaller models trade capability for speed and cost, and for many tasks the trade is favorable. Capability is not the only axis that matters.
Distillation, quantization and caching reduce cost at some cost in fidelity or flexibility. Each is a lever with a price rather than free improvement.
Using the most capable model for a trivial task wastes money and latency, so matching the model to the task is the routine optimization most deployments skip.
Y. People, trust and public understanding
What people believe these systems can do diverges sharply from what they do, in both directions. The gap drives both misplaced fear and misplaced reliance.
Overclaiming is reliably followed by backlash, and both phases distort judgment. Reading a claim for its evidence rather than its confidence survives both.
Effects show up as tasks within jobs changing rather than whole occupations disappearing, which makes the impact harder to see and more widely distributed.
Who benefits depends on access to tools, connectivity and the skills to use them, so the technology can widen an existing gap as easily as narrow it.
Learning requires doing the work that the tool can now do, so the question is which struggle is productive rather than whether the tool is permitted.
Conversational systems invite emotional reliance, which is a real effect on real people rather than a category error to be corrected. It deserves design attention.
People affected by a system having a say in it is a governance question that technical evaluation cannot answer, and it is frequently the one that determines legitimacy.
Z. Talking about AI accurately
Model, system, product and agent name different things, and using them loosely makes a claim unfalsifiable. Precision here is what allows disagreement to be productive.
Demonstrated, asserted and projected are three different epistemic states routinely presented in the same sentence. Separating them is most of reading a capability claim.
A score means something only with its conditions — which items, which prompting, how many attempts. A number quoted without them is not a result.
One impressive output shows the system can produce it, not how often. A capability claim needs a rate, and a single example cannot supply one.
Describing what broke precisely enough to be reproduced is what makes a report actionable. Saying it gave a bad answer conveys almost nothing.
Saying a claim is unverified is more useful than hedging every sentence into vagueness, because it names what would settle the question.
Teaching a non-technical audience without overclaiming means resisting both the impressive demonstration and the dismissive framing, since each misleads in its own direction.
Keentune is not affiliated with or endorsed by the organizations whose documentation informs these maps.
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.