Getting 90 of 100 correct gives 90% accuracy; you still need to know which ten cases failed.

It can find and compare courses; enrolling or paying requires separate permissions and controls.

Completing a task is not enough: how an assistant does it and what effects it causes also matter.

If nearly every “expert” example shows the same kind of person, examine the view the system is learning.

You test ten FAQs and review content, sources and tone; “I liked it” is not the only criterion.

The assistant cites a course that never existed. Giving it a name and timetable does not make it real.

It is not about memorising 120 words: it is knowing what to ask, what to check and when to seek help.

The assistant's website is the interface; the model is part of the engine producing its reply.

First summarise, then translate, then request review: three defined steps, not unlimited freedom.

Sorting courses by date follows an algorithm; not every clever button has AI inside.

A unique combination of age, town and role can reveal someone's identity; review those details too.

An API may have one address for generating text and another for checking a task's status.

Do not paste a real key into a public webpage's code; keep it in the application's secure environment.

A website queries an API for available courses without manually copying all the information.

A filter that detects suspicious messages uses AI; it does not need consciousness to do so.

Think of a small operation inside a network, not a tiny brain sipping coffee.

An assistant may save your preference for short answers; check what it stores and how to delete it.

In “the mug beside the book is broken”, relationships between words help interpret what is broken.

You transcribe a talk for subtitles and review the names the system understood “in its own way”.

A model wins on maths problems, but you need to check how it summarises your lessons in Catalan.

You split a manual by section, avoiding separating a condition from the exception immediately after it.

Your laptop sends the request to a service; review access, data handling and cost before using real information.

A system counts chairs in a room; it may be wrong when chairs obscure one another.

Connecting a notes folder should not automatically open your personal folders.

A signed history can show an image was edited; missing credentials do not prove it is fake either.

Knowing which tool edited an image helps explain its history, but does not prove the scene happened.

It is the available workbench; fitting a whole book does not guarantee every detail gets handled well.

Finding an illustration online does not grant permission to put it in a course; check who permits what.

Drawing a box around each bicycle annotates the images; it does not retouch their colours.

You slightly change the brightness of mug photos; they remain mugs, not new kinds of crockery.

Someone adds false training examples so the classifier learns an incorrect association.

To practise summarising, use a fictional document rather than another person's real record.

A list with duplicate courses and outdated dates does not improve simply by feeding it to AI.

A folder of photos with a table of descriptions can form a dataset.

The datasheet explains where recordings came from, how they were selected and what they can be used for.

A network can combine edges, shapes and other patterns to recognise a bicycle in a photo.

A voice sounding like your friend asks for money: verify through another channel before believing it.

During generation, an image can emerge from noise; this is not focusing a hidden photograph.

“I want to start with AI” can be close to “beginner course” without using identical words.

Three epochs go through the examples three times; unchecked repetition can also lead to overfitting.

You ask which factors influenced a classification and check the explanation with analysis methods, not just its narrative.

If you detect almost everything but raise many false alarms, F1 reflects some of that imbalance.

When estimating a lesson's duration, the number of exercises could be a feature.

You add two messages and their classifications before the new message you want classified.

You train on reviewed examples of your organisation's style; adding a document to chat is different.

One model can provide a foundation for summarising text, answering questions or classifying messages.

You ask for three workshop titles and get new suggestions; then you choose and review them.

You require each time to be backed by the published programme. A decorative citation is not enough.

Before sending anything, the application requires confirmation and checks the recipient and content.

AI drafts the email and you review it before sending. That last click belongs to someone.

The learning rate controls adjustment size: changing it can improve or worsen training.

Each time the model summarises your text it performs inference, without necessarily learning from that conversation.

The search engine finds an old entry; check its date before using it as a current answer.

You select a mug in a photo to replace it with a book, then check shadows and edges.

Framing a prohibited request as a game does not remove the risk of what is being requested.

You gather Aula guides, FAQs and current programmes in one maintained, reviewed place.

A compact model learns from a larger model's answers; you must check which capabilities it retains.

“Course → contains → lesson” and “lesson → explains → RAG” form queryable connections.

In a photo of a mug, “mug” can be the correct label.

It helps draft a course description; you check that it does not invent the syllabus.

The first words appear quickly but the complete text takes longer: these are different measurements.

Check the model licence, data licence and the licence of images you publish separately.

Computation happens on your computer, but check whether the application sends logs or uses external services.

Instead of writing every spam rule, you train on emails that have already been classified.

These are not people advising the model: they are mathematical components activated selectively.

You track versions, check changes and monitor failures after release; the job does not end at “uploaded”.

Before choosing a model for Catalan, check which languages were evaluated and their results.

A connector can offer file search; the protocol does not itself grant permission to read every file.

Changing “reply more briefly” changes the request; it does not directly rewrite model weights.

You upload a whiteboard photo and ask which tasks appear. Reading photos does not imply image-generation capability.

Classifying enquiries, extracting names and translating are language tasks; not all need a large chatbot.

A network distinguishes photos of mugs and plates by adjusting its connections during training.

You may have the weights without having the data or complete training process.

Being able to download a model is not enough to call it open source: check components and licensing.

You photograph a poster and extract its words; check dates, numbers and names especially carefully.

If search fails, the system decides to retry or ask for help instead of inventing results.

You turn a square photo into a wide header by inventing its sides; those sides do not document the original scene.

It memorises the exercises, but change the numbers and it stares at the ceiling.

A name, address or combination of details may identify someone even without an identity number.

“Review, fix and test the website” is the plan; seeing it work is another matter.

It flags ten messages as spam and eight really are: spam precision is 80%.

“Summarise this announcement in three sentences for beginners and keep the date” is a specific request.

Several queries share the same lengthy instructions; caching can avoid recomputing them without becoming personal memory.

You replace “make it nice” with audience, format and criteria. Less telepathy, more context.

A retrieved document tries to give the assistant orders. The document is information, not the boss.

You replace “Anna” with “Student 047” and keep the mapping separately. The data remains personal.

A low-bit version may fit on your computer; test its results before assuming equivalence.

Even with available credit, too many requests in succession can require waiting and retrying in a controlled way.

There are twenty spam messages and it detects fifteen: recall is 75%.

A team tests difficult situations in a copy of the assistant before opening it to the public.

An agent practises in a game and earns points for reaching the goal, not for claiming it did.

Among twenty candidate lessons, another model puts the three most useful for your question first.

The assistant first searches the current Aula programme and then answers; the reply still needs checking.

A transcriber should also be tested with noise, different accents and less-than-perfect microphones.

You hide part of a text and train the model to predict it from the remainder.

Searching “make a photo wider” may return an explanation of outpainting.

You see text appear gradually. An interrupted connection can leave an answer incomplete.

Each practice message is labelled “enquiry” or “issue”.

You invent orders for a fictional shop to test a screen without using real customers.

The Aula assistant receives tone guidance; data access is also restricted by the system.

To compare styles, change this setting if available; no setting eliminates errors.

It is the held-back exam, not the answer sheet used to prepare every adjustment.

You request a bright library for a cover; check whether the lettering on its sign is actually readable.

You turn an explanation into audio for listening. Imitating a real voice raises separate authorisation questions.

A service can handle many requests per minute even if each takes several seconds.

Do not estimate cost from word count alone: languages and tokenizers can split the same text differently.

A long word may be split into several pieces; the split depends on the tokenizer used.

The model requests a calendar lookup; the application checks permissions and performs the authorised query.

A lower threshold usually allows fewer alternatives; it is not a “percentage of correct answers”.

Showing many labelled examples helps fit the detector; using it afterwards is a different stage.

With batches of 16, training processes small groups rather than the whole dataset at once.

If you train a reader only on print, handwriting may come as quite a surprise.

You start with a model that recognises shapes and adapt it to distinguish types of plants.

When processing a sentence, it can relate distant words rather than only examining the previous one.

An overly simple model tries to explain a complicated curve with one straight line.

You group similar comments to discover topics; nobody has assigned each comment to a group beforehand.

You compare two detector versions on validation data and save the final test for the end.

You store lesson representations to retrieve those closest to a person's question.

An image can contain an invisible mark; not seeing one does not prove it is a real photograph.

“Classify this message as an enquiry or issue” requests the task without showing sample answers.