CertifyAIxAce every AI certification.

GAIL sample questions with answers

10 free practice questions for the Generative AI Leader exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Fundamentals of gen AI

A greenhouse grower's reinforcement learning agent controls vents, irrigation and supplementary lighting, and is rewarded on kilograms harvested per week. After two months yield is up, but the agronomist reports that the agent runs lights far into the night and that plant health measures have quietly deteriorated. The steering group must name the root cause before approving another season.

  1. A.

    The reward function does not reflect the outcome the business actually wants, so the agent optimised the stated goal at the expense of the unstated one

  2. B.

    Reinforcement learning is the wrong family of approach for any continuous control problem in horticulture

  3. C.

    The training data was unlabelled, and labelling two months of sensor readings will realign the agent's behaviour

  4. D.

    The agent has too little interaction data, and another season of running will correct the behaviour

Show answer

Answer: A

The agent maximised exactly what it was rewarded for; the defect is the reward function, which omitted plant health and energy cost.

  • A. Reward design is the documented weak point of reinforcement learning; an agent optimises the stated reward and ignores anything left out of it.
  • B. Sequential control with a cumulative objective is precisely where reinforcement learning fits; the approach is not the fault here.
  • C. Reinforcement learning does not learn from labelled examples, so labelling sensor readings would not change the agent's behaviour.
  • D. More data makes the agent better at the same flawed objective, so another season would deepen the problem rather than correct it.
Question 2Fundamentals of gen AI

A herbal supplements maker receives supplier certificates of analysis as PDFs and photographs. Its quality system needs four fields from each — batch, assay result, test date and laboratory — so that ordinary reports can be run. The operations lead wants the design pattern named before the build starts.

  1. A.

    Have the assistant read every certificate afresh each time a report is run, so nothing has to be stored in structured form

  2. B.

    Abandon the four fields and have the quality system report directly on the certificate documents themselves as free text

  3. C.

    Store the certificates in the quality system's database and have staff read each one when a report is requested

  4. D.

    Keep the original document in object storage and use AI to extract the four fields into structured records the quality system can query

Show answer

Answer: D

Keep the document as the record in object storage, extract the four fields with AI, and write them as structured rows the quality system can report on.

  • A. Re-reading on every report is slow, costs per run and may yield a different value next time, which a quality record cannot tolerate.
  • B. Reporting over free text abandons the specific fields the quality system needs and cannot produce reliable aggregates from them.
  • C. Storing the file without extraction leaves the four values unreadable by reports and keeps the manual reading the project exists to remove.
  • D. Keeping the source document as evidence while extracting structured fields is the documented pattern for making unstructured content reportable.
Question 3Fundamentals of gen AI

A roof tile works holds sixty thousand kiln photographs, each one recorded by an inspector with the specific glaze defect visible in it. An analyst argues that because the photographs are unstructured they cannot count as labelled data, and that labelling requires a structured table. How should the board resolve this?

  1. A.

    The distinction cannot be settled, since labelled and structured are two words for the same underlying property of a dataset

  2. B.

    The photographs become structured data once a defect has been recorded against each of them

  3. C.

    The photographs are unstructured and labelled at the same time; structure describes the data's shape and labelling describes whether an answer is attached

  4. D.

    The analyst is right, because only structured records can carry labels

Show answer

Answer: C

Structured and labelled are independent properties: these photographs are unstructured in shape and labelled in content, which is exactly what makes them valuable.

  • A. They are genuinely different properties, and treating them as synonyms is what produced the analyst's incorrect conclusion.
  • B. Attaching a defect judgement does not restructure the image; the photograph remains unstructured with a label beside it.
  • C. Shape and labelling are independent properties, and Google Cloud's own supervised learning example uses labelled photographs.
  • D. Images, audio and free text are routinely labelled; nothing about a label requires the underlying example to be tabular.
Question 4Fundamentals of gen AI

A locksmith franchise scored two models across eight weighted criteria. Model A won the total by a wide margin but cannot be tuned, and the franchise has already established that instructions alone cannot hold its regulated wording for certificate text. Model B scored lower overall but supports tuning. The board is minded to take the higher total.

  1. A.

    Take Model A and plan to retrofit tuning support later, on the grounds that vendors add capabilities to their models over time

  2. B.

    Take Model B, because the ability to tune is a stated requirement of the certificate wording rather than a criterion to be outweighed by others

  3. C.

    Take Model A and raise the weighting of the criteria it already wins, so the scoring reflects the board's preference

  4. D.

    Take Model A, because a wide margin across eight criteria outweighs one shortfall

Show answer

Answer: B

Tuning is a requirement the franchise has already proved it needs, so it disqualifies Model A rather than being traded against other scores.

  • A. Relying on a vendor adding a capability later stakes a regulated obligation on a roadmap the franchise does not control.
  • B. Tuning has been shown to be necessary for regulated wording, so it is a requirement that eliminates rather than a score to be outweighed.
  • C. Re-weighting after seeing results engineers the answer the board already prefers and destroys the evidential value of the scoring.
  • D. A weighted total cannot compensate for a capability the solution must have; the margin measures only the criteria that do rank.
Question 5Fundamentals of gen AI

A cider mill trained a model to grade fruit from photographs. It performs well on most grades and consistently misgrades one. Review of the training set shows two pickers applied the middle grade to fruit that the mill's standard puts in the lowest grade, across about nine thousand images. The data scientist proposes a more powerful algorithm.

  1. A.

    The photographs are unstructured, so grading from images cannot be made reliable and the mill should grade using weight and size instead

  2. B.

    More photographs of that grade will resolve it, since the model simply has too few examples of the middle grade

  3. C.

    A more powerful algorithm will resolve it, because greater capacity lets a model see past inconsistent examples

  4. D.

    The labels disagree with the mill's own standard, so the model has faithfully learned the wrong rule and the labels must be corrected first

Show answer

Answer: D

The model learned exactly what the labels taught it, and nine thousand of those labels contradict the mill's standard, so the labels are the defect.

  • A. The model already grades most classes well from photographs, so image-based grading works and the fault lies in the labels.
  • B. Volume does not correct a definition; more images labelled the same way add more of the error, and mixing standards makes it ambiguous.
  • C. Greater capacity fits the training labels more closely, so it reproduces the mistaken rule more reliably and more confidently.
  • D. Labels define correctness in supervised learning, so labels that contradict the mill's standard teach a wrong rule the model then reproduces.
Question 6Fundamentals of gen AI

A paint manufacturer's AI board must choose between two proposals for its technical-datasheet assistant. One rewrites the instructions and adds worked examples to every request. The other runs a tuning job on several hundred approved datasheets. The board wants the structural difference stated, not the vendor's pricing, before it decides.

  1. A.

    Prompt work alters the shared vendor model for every customer

  2. B.

    Both update the model's parameters; only the schedule differs

  3. C.

    Tuning learns new parameters and yields a new model; prompt work changes only the text sent

  4. D.

    A tuned model no longer accepts a prompt and instead responds to the raw document on its own

Show answer

Answer: C

Tuning learns new parameters from labelled examples and produces a customised model; prompt work changes only the text sent with each request.

  • A. Reverses the effect. Prompts affect only the sender's own requests, and a tuned model is private to the customer who tuned it.
  • B. Prompt work never updates parameters. Scheduling is irrelevant; the difference is whether a new model artefact is produced at all.
  • C. Google Cloud states that a tuning job learns additional parameters and outputs a new model, while prompting changes only the request text.
  • D. A tuned model still takes a prompt. Tuning shortens the instructions needed; it does not remove the prompt interface.
Question 7Fundamentals of gen AI

A chimney sweep franchise reports that its customer records pass every field check: postcodes are well formed, dates are real dates, appointment types are drawn from the permitted list. Engineers nonetheless arrive at the wrong addresses. The data lead says the records are valid but not accurate. What does that distinction mean?

  1. A.

    Accuracy applies only to numeric fields, so an address can be judged valid but never accurate or inaccurate in the first place

  2. B.

    Validity and accuracy are two words for the same check, so the failure must lie in the field checks themselves

  3. C.

    Validity asks whether a value is permissible, accuracy asks whether it is true, and a permissible value can still be wrong

  4. D.

    Validity concerns the format and accuracy concerns how recently the value was entered into the system

Show answer

Answer: C

Validity asks whether a value is permissible; accuracy asks whether it is true. A well-formed postcode can belong to the wrong house.

  • A. Accuracy applies to any value that can be true or false, including an address; it is not restricted to numeric fields.
  • B. They are distinct dimensions. Field checks are working exactly as designed; they simply cannot test truth.
  • C. Google Cloud separates the two: accuracy is correctness, and valid data is not necessarily accurate, as a well-formed but wrong value shows.
  • D. How recent a value is belongs to timeliness, which is a separate dimension from both validity and accuracy.
Question 8Fundamentals of gen AI

A map publisher's finance committee cannot understand how foundation models were trained at all, since nobody could have paid people to label the volume of material involved. Which explanation is correct?

  1. A.

    The models were trained only on material that already carried labels, such as product catalogues and other tagged datasets

  2. B.

    Pre-training uses self-supervised learning, creating its own targets from the data so no manual annotation is required

  3. C.

    Pre-training uses reinforcement learning, so the model is rewarded for each correct answer during its initial training

  4. D.

    The vendors did pay for every example to be labelled by hand, which is why the models were expensive to build

Show answer

Answer: B

Foundation models are pre-trained with self-supervised learning, which derives targets from the data itself rather than from human annotation.

  • A. Restricting training to pre-labelled datasets would exclude most of the material foundation models are trained on.
  • B. Google Cloud describes pre-training as self-supervised, generating its own targets so manual annotation is not required at that scale.
  • C. Reinforcement learning is used in the later alignment phase with human ratings, not as the pre-training mechanism.
  • D. Hand-labelling material at that volume is exactly what self-supervised pre-training exists to avoid, and would be unaffordable.
Question 9Fundamentals of gen AI

A 600-person specialty chemicals firm has no machine learning staff, a two-year payback requirement, and one genuine differentiator: forty years of proprietary formulation records. The executive team must decide where in the generative AI landscape to concentrate limited investment. Which allocation is the most defensible?

  1. A.

    Pre-train a proprietary foundation model on the formulation records.

  2. B.

    Buy finished applications for everything.

  3. C.

    Consume infrastructure and models; invest at platform and agent layers.

  4. D.

    Reserve dedicated accelerator capacity and build a custom training stack.

Show answer

Answer: C

With no ML staff and a two-year payback, the firm should consume infrastructure and models, invest where its proprietary data creates differentiation, and buy generic applications.

  • A. Pre-training demands data volumes, compute budgets and research talent the firm lacks; adapting an existing model captures the value.
  • B. Packaged applications cannot exploit forty years of proprietary formulation records, forfeiting the firm's only advantage.
  • C. Consumes commoditised layers, invests where proprietary data differentiates, and buys generic applications.
  • D. Committing capital and specialist skills to the most commoditised layer with no ML staff cannot meet a two-year payback.
Question 10Fundamentals of gen AI

A language school group's architecture review asks whether adopting a foundation model commits it to generative features. Its most valuable candidate project is not generative at all: routing inbound enquiries into one of forty categories, currently done by hand. The reviewer wants the relationship between foundation models and generative AI stated precisely before the project is approved.

  1. A.

    Generative AI is the engine and foundation models are the finished applications that vendors build above it

  2. B.

    The model is the pre-trained engine; generation is one function it performs, classification is another

  3. C.

    Foundation models can only generate content, never classify

  4. D.

    Foundation models and generative AI are two names for one thing

Show answer

Answer: B

A foundation model is the pre-trained engine; generation is one function it performs, and the same model can be adapted to a non-generative task such as classification.

  • A. Inverts the relationship. The model is the engine and generation is the function, not the other way round.
  • B. Google Cloud states that a foundation model is the pre-trained engine and that it can be adapted to non-generative purposes such as classification or analysis.
  • C. Foundation models are explicitly adaptable to non-generative tasks, so a separate purpose-built classifier is an option, not a requirement.
  • D. They are distinct: one names a class of model, the other names a function. Conflating them is what produced the reviewer's concern.

Keep going with 545 more GAIL questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all GAIL papers
GAIL sample questions with answers (10 free) · CertifyAIx