This training focuses on working with outputs that were not created by the person themselves — specifically, what to do with them when they read well and were generated by an AI model. Participants will strengthen two key skills: the ability to determine how much scrutiny each output requires and the ability to actually carry out that level of verification within the time available during a typical working week.
The main focus of the day is a shift from the role of the person producing the content to that of the person evaluating it and taking responsibility for what they pass on. This is not training about distrusting AI or about why it should not be used. It is simply not possible to verify everything, so the focus is on applying a level of scrutiny proportionate to the potential consequences of an error.
Today, a finished output can be generated in a matter of seconds. However, the responsibility for deciding whether it is fit to be used or shared remains entirely with the human. That is why the entire training is built around outputs that participants create during the training using tasks from their own work agenda.
Logical fallacies, media credibility assessment and fake news are covered in the general Critical Thinking course and are therefore not repeated here. The focus of this training is specifically on claims contained in AI-generated outputs.
- Develop the ability to determine how much scrutiny an output requires — and to carry it out, not merely plan it
- Develop the ability to identify unsupported claims even in text that reads well and sounds confident
- Increase confidence in verifying the evidence behind a claim outside the generated output — including situations where a source exists but does not actually support the claim
- Strengthen the habit of forming your own opinion before being influenced by the model’s first suggestion
- Encourage openly stating what has not been verified instead of silently passing the output on
What Did the Model Actually Give Me?
- What in a finished output is a verifiable claim, what is an assessment, and what is merely filler
- Which single claim would have the greatest impact if it turned out to be wrong — and why that claim should determine the order of verification
- Forming your own opinion before seeing the model’s first suggestion: when it pays off and what happens when this step is skipped
Why Errors in Good Outputs Are Hard to Spot
- Automation bias — why people tend to accept a system’s suggestion more readily than a colleague’s suggestion
- Overreliance — accepting an output even when there are sufficient grounds to reject it
- Anchoring effect — the model’s first suggestion becoming an anchor for everything that follows
- The confident tone of an output as information about its wording, not about the quality of its content
Three Levels of Output Validation
- Are the facts accurate? Do the cited sources exist? Does the argument hold together?
- Three symptoms of an unreliable reasoning chain: a leap in reasoning, evidence that does not support the claim, and certainty without evidence
- When a source exists but does not support the claim — three possible verdicts on the strength of the evidence
- A deliberately flawed output: participants first identify where the conclusion does not follow from the evidence and then learn what was actually wrong with it — including what was correct
How Much Scrutiny Should Each Output Receive?
- When does combining human and AI capabilities make sense, and where does the human add no value? Why more scrutiny is not always better
- Consequences as the measure: what happens if nobody notices the error, and who bears the consequences
- Four levels of scrutiny — from a quick review to “I would not send this without someone else checking it”
- How much time verification actually requires and what this means for recurring tasks
- How to consciously reduce the level of scrutiny when there is not enough time for full verification — and how to communicate this openly
The Decision — and What Comes Next
- Five possible outcomes of an assessment: use it, correct it, pass it on with a caveat, discard it, or escalate it
- Three moves when an output does not stand up to scrutiny: refine the prompt, ask for supporting evidence, go to the original source
- Why asking the model to explain its reasoning is not evidence — and what the human remains responsible for when passing an output on