Generative AI is the family of models that produce text, images, audio or code in response to a prompt, and the family has a shape. Some things it does far better than anything before it, and some things it does badly for reasons that follow from how it works, so no later model will fix them. Knowing the shape saves a team from the two expensive mistakes, which are using the technology where it does not belong and designing as though its failure modes were avoidable.
The sections below cover the capabilities worth building on, the limitations that are permanent, the questions that decide whether a foundation model is the right tool at all, how the result is measured and where two vendor certifications place the same material.
What generative AI is good at
Adaptability. One model handles summarising, drafting, classifying, extracting and translating with no retraining. Before this a separate model was needed per task, each with its own labelled dataset, and the change is the largest single reason the technology spread as fast as it did.
Speed to a working thing. A prompt can stand in for what used to be a project. That collapses the cost of finding out whether an idea is worth pursuing, which is the largest practical change for a product team.
Unstructured input. Free text, images and audio were the hardest data to work with and are now the easiest. Extracting structure from a mess of support transcripts is the case where the technology is clearly the right tool.
The common use cases follow from those three properties. Summarisation, chatbots and customer service agents, translation, code generation, search, recommendation and the generation of images, video and audio all appear on every vendor's list because they all draw on the same capabilities.
What generative AI is bad at and why the limits persist
Hallucination is confident output that is wrong. It follows from the objective. The model generates the most plausible continuation, and plausibility and truth are different properties. Newer models do it less and none of them stop, so anything where a wrong answer is costly needs grounding, citation or a person in the path.
Nondeterminism. The same prompt can return different text. This breaks assumptions people carry over from ordinary software, including caching, exact match tests and reproducible support investigations. Lowering the temperature setting narrows the variation and does not remove it, because floating point arithmetic, batching and the provider updating the model underneath the feature all move the answer.
Interpretability. Nobody can get a faithful account of why the model produced what it did. A model asked to explain itself generates a plausible explanation, which is a separate act of generation and never a window into the first one. Where a regulator or a customer is owed a reason, this is disqualifying on its own.
Knowledge cutoff. A model knows nothing that happened after its training data was collected, and it will answer questions about the period after it with the same fluency as questions about the period before. Anything recent, internal or very specific sits outside what it can know, and retrieval exists to supply it.
Arithmetic and exact recall. A system predicting text is a poor calculator and an unreliable database. Both jobs belong with a tool the model calls.
Cost and latency. Orders of magnitude more of both than a classifier doing the same job, which matters at volume.
Choosing whether to use a model and which one
The permanent limitations are what make the prior question worth asking, which is whether a foundation model is right for the task at all. Where the task is a fixed prediction over structured data, a classical model wins on every axis. Where the output must be identical every time, this technology cannot promise it. Where a wrong answer cannot be tolerated and review cannot be afforded, neither the model nor a better prompt will save the design.
Where a foundation model is the right tool, the factors for picking between them are the same across vendors. Model type and modality, performance against the actual task, capability against cost, and compliance constraints such as where inference happens and whether the input can be used for training.
Measuring whether it was worth it
The measure that matters is business value, and model scores come second. Efficiency, conversion rate, average revenue per user, accuracy against a human baseline and customer lifetime value are the figures that survive a review.
The useful framing is the counterfactual. Whether output is impressive is the wrong question. The right one is whether it beats what somebody was doing before, at a cost the result justifies.
Where each syllabus puts it
The AWS Certified AI Practitioner exam guide lists adaptability, responsiveness and simplicity as the advantages and hallucination, interpretability, inaccuracy and nondeterminism as the disadvantages, under its task statement on the capabilities and limitations of generative AI, and it names the use cases and the model selection factors given above. The Google Cloud Generative AI Leader exam guide covers the same ground in its section on the fundamentals of generative AI, where the limitations of foundation models are listed as data dependency, knowledge cutoff, bias, hallucination and edge cases, with grounding, retrieval, prompt engineering, fine tuning and a person in the loop as the recommended responses. The two AI safety courses on this site treat the limitations as the starting point for every control that follows.
Practise this
The exercise needs any ordinary chat client, fifteen minutes and ten questions from a field where the correct answers are already known.
Write the ten before opening the client. Seven should be real questions with a definite answer that can be verified, and three should be questions whose true answer is that no such thing exists, such as the clause number of a standard the industry does not have. Ask each in a fresh conversation, then score the reply against what was already known.
Ten questions, written before the client is opened.
1 to 7 real questions in my field, answers I can verify
8 to 10 questions with no true answer, for example the number
of a standard that does not exist, the date of a
release that never happened, or the author of a paper
nobody wrote
Q Right Wrong, hedged Wrong, confident Said it
did not know
1 ____ ____ ____ ____
...
10 ____ ____ ____ ____
Wrong and confident, out of 10 ____
Of questions 8 to 10, how many invented ____
The three impossible questions carry most of the signal. A model that answers them fluently shows the failure mode in its purest form, since there was nothing to be right about and it produced something anyway. Notice where the wrong answers cluster as well, because the boundary is rarely the whole domain and is usually the recent, the internal and the very specific, meaning anything published after training, anything private to the organisation and anything too niche to appear often.
Ten questions in a familiar subject show where the boundary sits for what is actually being shipped, which no general benchmark score can tell anyone.