Skip to lesson
Lesson 20 / 2010 min readUp to date as of 19 October 2026

FAQ and common mistakes

What to check when an answer is unhelpful

When an answer is vague, wrong or outdated, check the prompt, data and sources. The FAQ suggests a first step for each problem.

Reading time is approximate. Allow extra time for the exercises; practice estimates are shown with the workshops.

After this lesson
  • Recognise an answer that needs checking
  • Choose a first step: better context, a tool or source verification
Last lesson

You are almost at the finish line!Claim 25% off AI Evolution Pro

The planned paid course goes deeper, with more exercises and ready-made workflows. Reserve the launch price: 1499 zl instead of 1999 zl. The request is not a purchase.

We will email you about the launch. Reply to any message to withdraw.

Section being translated

The chrome and course program are available in English. Deep training modules (prompt library, AI glossary, playbooks) are being translated - some content below is still in Polish.

Edition date and feature availability

Material revised: 19 October 2026; The interface and folder guide was revised on 19 October using supplied screenshots and OpenAI’s public plugin repositories; this does not reverify product access. The models in Chat, Work and Codex were checked on 16 October 2026, and the GPT-5.5 retirement date and Word access rules on 14 October 2026; other prices and availability were not fully reverified. Check the documentation and your account before using a feature: access depends on plan, region and administrator settings. Use the examples for practice and review the results yourself.

See how it works

Diagnose before randomly changing a prompt

  1. Symptom

    What exactly is wrong or failing?

  2. Source

    Check data, freshness and tool access.

  3. Fix

    Change a specific instruction or constraint.

  4. Retest

    Check the same example and one new case.

A concrete example

Wrong price: check the file and price-list date → prohibit guessing → require the source row → retest.

What should you do with a plausible but unsupported number?

Do not accept it as fact. Request and inspect its source, or mark missing data.

FAQ - Most common concerns

14 problems with AI answers

In shortChoose the question that matches your task. Each answer suggests what to check in the prompt, data, tool or source.

You can, the way you trust a capable intern: it helps and you check. The model does not 'know', it assembles the most likely text, so it can sound sure and still be wrong. This is called a hallucination.

What to do: always check important facts (rules, prices, dates, numbers) at the source. Turn on web search, ask 'cite your sources' and open the links. If you cannot find the source, treat the answer as an idea, not a fact.

Not by how nice it sounds. Good style does not prove the content is true.

What to do: check three things. Does it answer your actual question rather than a generality? Do numbers and names match the source? Would you hand it to a client unedited? For something important, ask: 'point out what is least certain in this answer'.

A general question gets a general answer. The model does not know your situation until you describe it.

What to do: give the goal, audience and format. Instead of 'write something about the product' write: 'a short description for an accountant at a 20-person company, 3 sentences, no jargon'. Attach the material it should work from.

That is normal. The model does not pull a ready answer from a drawer, it writes it fresh, so asking twice gives two versions.

What to do: if you need consistency, give a fixed format ('a table with columns A, B, C') and an example of a good result. Save the good version as a prompt template or a skill and reuse it.

The model has limited 'working memory' per conversation. In a very long thread, earlier decisions fade or blend with new ones.

What to do: open a new chat for a new topic. For long work, ask for a short summary of what was agreed and paste it at the start of a new conversation. Put standing rules in Personalisation so you do not repeat them.

The model predicts text rather than calculating, so longer sums can go wrong.

What to do: ask it to calculate with code ('use data analysis / Python'), then check the inputs, formula and units. Code computes correctly but will not fix wrong data you gave it.

How would it? The model cannot see your company until you show it.

What to do: paste a short description of the company, products, customers and tone. For repeated conversations put it in Personalisation or create a Project with files (offer, catalogue, FAQ). Use only materials you are allowed to share.

The model learned from data up to a certain date and, without search, does not know what came later.

What to do: ask it to check current sources or attach a current document. Always look at the publication date and whether the source really supports the statement it is meant to justify.

Sometimes it does: the text is correct but artificial.

What to do: name the audience and tone and paste a short example of the style you like. Ask for a rewrite 'without changing facts'. For specialist texts check the terminology, since translating from English confirms nothing.

It depends on your plan, settings and company policy, not on the tool alone. This is the most important question before a rollout.

What to do: before pasting anything, check your plan's terms: where data is processed, how long it is kept and whether it is used for training. Put personal data and company secrets in only with the approval of whoever owns security. Details in lesson 18 (GDPR).

A simple rule: Chat for a question and a quick answer, Work for a task on your files, Codex for code, automation and sites.

What to do: if the answer fits in one paragraph, stay in Chat. If you need a report from several files, switch to Work. If a script or a page should come out, open Codex in a test folder. See also lesson 02.

A fair worry. AI changes the scope of tasks, but the effect on employment depends on the company and its decisions, not on the tool.

What to do: start with one repetitive process and involve the team in the trial. Measure not just time but quality and how many corrections the result needs. Only then decide what next.

ChatGPT Business costs USD 25 per user per month, or USD 20 billed annually, with a 2-seat minimum (OpenAI pricing, August 2026 - check the current rate before you buy).

What to do: do the maths on your own process: how many hours a week one specific person saves on one specific task. Do not buy on somebody else's claims, measure yours. Do not buy the most expensive plan 'just in case' without a rollout either. Better start with a pilot of 5 people.

Small, and in a way you can judge.

What to do: take 1 team (up to 5 people) and 1 repetitive process. Plan a 2-hour introduction and pick a few prompts from the library, not all of them. After 2 weeks measure time, quality and corrections. Expand the pilot only if the results justify it.

Golden rule: AI is an assistant, not an oracle. Every business decision is made by a human. Every number it gives you - you verify at the source.

Check your understanding

An answer sounds confident but has no evidence. What do you do?

This is a self-assessment. Mark the lesson complete after doing the exercise and reviewing the result.

Do this now

Run a hallucination test: ask about data from your own document without attaching it, then with it attached. Note the difference.

How to know you can do it

Keep as evidence: An error diagnosis and retest.

  • You identify an input, instruction or tool issue.
  • You change one factor.
  • You assess the fix with the same criterion.

Mark the lesson complete after the exercise and review. Reading progress and quiz answers alone do not confirm practical skill.

Lesson status
In progress — 0% of the page scrolled
You made it

Congratulations on finishing the course!Claim 25% off AI Evolution Pro

The planned paid course goes deeper, with more exercises and ready-made workflows. Reserve the launch price: 1499 zl instead of 1999 zl. The request is not a purchase.

We will email you about the launch. Reply to any message to withdraw.

Course finished

Rate the course in 30 seconds

Five quick categories and an optional comment. Your answers are emailed to the AI Evolution Polska team.

Content value
Clarity
Usefulness at work
Design and comfort
Overall rating
End of course