Deep dive · module 27 of 100
Evaluation & quality
Assess whether AI output is good enough for use.
After this lesson
You use a simple checklist to check AI results.
In view · Along the same yardstick
A fixed quality bar
Choose, for example, accuracy, completeness and appropriate tone.
Compare results using the same input and the same bar.
Rework or reject an answer that lacks the required quality.
Part 1 of 3 · Determine criteria
Your preference for less movement is on. Choose each part yourself.
Defining criteria
Determine in advance what is good: facts, tone, completeness, usability. Make the criteria observable: for example, all numbers traceable to the source, no new commitments, and one clear next step. Determine which error makes a result unusable. Use the same threshold for all variants you compare.
Asking for self-correction
Sometimes ask AI to check itself, but never trust blindly. A model can repeat its own error or defend it convincingly. Use self-control as an additional suggestion for points of attention. Check important claims yourself in the original source and verify calculations with an appropriate calculation tool.
Comparing
Place the result alongside the source or your own expectation. Keep a small test kit with normal situations, missing information, and contradictory sources. Note what went wrong and what adjustment you are making. Re-test the kit when you substantially change the prompt, dataset, or tool used.
New practice example
From explanation to a usable result.
Quality becomes clearer when multiple people assess the same result using the same criteria.
Fictional practice situationFiction: two colleagues are reviewing a summary. One focuses mainly on fluent language, the other on missing agreements.
View a possible solution
This is how it can be done
Use four criteria beforehand: correct facts, all necessary agreements, no unsubstantiated additions, and appropriate format. Have each point justified with a source passage.
Why this works
You can discuss disagreements using the same yardstick. A total score must not hide a serious factual error.
Make it slightly more difficult
Add a stop criterion where the result may not be used regardless of the other score.
Extra practice, at your own pace. You do not need to submit anything. This does not count towards an automatic assessment.
Your workplace · learning by doing
Make something. See what changes.
Checking the AI connection… You can already view the written example.
Present your assignment to Ende.
Read the answer, change one thing, and compare.
Previously held
New version
You remain in control
What do you want to improve in your next attempt?
On this device only. AI answers remain only in this page memory. This is self-review, not a grade.
Why this approach works
You can discuss disagreements using the same yardstick. A total score must not hide a serious factual error.
A step further
Add a stop criterion where the result may not be used regardless of the other score.
Try & discover · no obligation
Compare quality with a fixed yardstick
Give tone a lot of weight. Is a beautiful but incorrect answer still allowed to pass?
Fictitious: an online workshop, Saturday from 10 to 11 o'clock. The price is not yet known.
Smooth but incorrect
Discover our amazing online workshop this Saturday from 10 to 11 o'clock, completely free!
Facts: 0/en/3 · Toon: 3/en/3
Stop: the free price is made up. A high total score does not offset that error.
Factual and sober
Workshop: Saturday, online, 10–11 o'clock. Price to be confirmed.
Facts: 3/en/3 · Toon: 1/en/3
Can proceed to human final check.
The two criterion scores have been pre-assigned for this fictitious case. You change their weight. No free text is automatically assessed.
This trial works locally with fictitious data and does not count towards your lesson completion. It starts anew with a new lesson or reload.
A prompt to get you started
Check the text below for factual errors, missing information and tone. Provide a short comment per point. Text: [text]. Use only the provided fictitious facts. Make missing information visible and indicate what a human must check before use.
Apply it
A AI summary is clearly written, but omits an important exception from the provided procedure.
How do you assess this outcome?
- Successful, because the tone is good.
- Not yet usable: completeness and the source exception must be restored and verified.
- Successful if AI itself says the answer is correct.
With JavaScript, you can check your approach and receive immediate feedback.
Test your insight
What is the best way to evaluate AI output?
- Only look at whether it sounds good.
- Using a checklist and comparing with the source.
- Always accepting everything.
Extra practice
Create a checklist of 4 points with which you evaluate AI output. Work with a fictitious situation and name one concrete control point before you would use the result.
1. Are the facts correct? 2. Does the tone fit? 3. Is the next step clear? 4. Are there invented elements? I compare the outcome with the confirmed input, leave unknown data blank, and adjust the result before use.
