You ask AI to help prepare a note for a meeting with colleagues. Its response contains information you recognise, an explanation of why the group supposedly behaves a certain way, and an organisational suggestion. All appear in the same paragraph, with the same confident tone. Before using the text, distinguish what is documented, what is interpreted and what is merely proposed.
For psychologists and psychotherapists, this distinction also matters in study materials and professional communication. Here it does not involve assessing people or clinical cases. We use a completely fictional message and response to make the type of claim in each sentence visible. The suggested categories are reading notes, not an automated system for certifying truth.
Start with the available information
Imagine you wrote this fictional message: “For Thursday’s meeting with colleagues, the room is available from 6 pm. Bring one page with a question about the article. We have not yet chosen who will introduce the discussion.” This is all the material provided in the example. We do not know participants’ motivations or the meeting’s confirmed start time.
Before reading any AI proposal, you can identify three explicit details: room availability, the requested material and an introduction still to be assigned. The first does not establish a 6 pm appointment. The second requests a question about the article without specifying a complete summary. The third describes an open organisational decision.
The NIST Generative AI Profile distinguishes facts, opinions and inferences in its information integrity section. This general guidance does not validate our exercise. Our first step is to keep the original message beside the response so that the distinction can be checked.
Recognise a documented statement
A sentence such as “the room is available from 6 pm” can be traced to the message. You can label it “explicit information in the source” and identify the corresponding passage. This label describes the relationship between the texts. It does not prove that the room is actually booked, that availability is still current or that no further conditions apply.
Organising a real meeting would require checking through the responsible people and channels. Our exercise stays with the fictional text. The distinction prevents agreement with a document from becoming a general guarantee of accuracy. A document can itself contain an error or information that is no longer current.
Also check whether a response inserts a detail into an otherwise faithful sentence. “The room is available from 6 pm for two hours” preserves the start but introduces an absent duration. Split the sentence and separate the documented information from the addition. Recognising a few words from the message is not enough to verify the entire statement.
Make the interpretive step visible
Now imagine this problematic response, also invented: “The meeting will begin at 6 pm. The group is inexperienced and therefore needs a facilitator. Each participant must prepare a full summary of the article.” The first sentence concludes more than the message says. The second assigns a level of experience and a need. The third changes the requested task.
These are not three identical errors. The first confuses room availability with the start of an activity. The second explains supposed group characteristics without supporting information. The third replaces an instruction. Separating these steps helps identify the check or correction needed, rather than approving or rejecting the entire paragraph as a single unit.
An interpretation is not automatically useless. It can open a question if presented that way and supported by an explicit reason. Here you can ask whether facilitation should be agreed without claiming that the group is inexperienced. Adding “perhaps” before a personal attribution is not enough to give it a basis. Often the better choice is to remove the attribution and return to the organisational decision.
Separate proposals from commitments
“We could agree beforehand who introduces the discussion” is a proposal about the open decision. It does not describe something already agreed. If you include it in your note, keep wording that shows its status and make clear who would decide. AI does not turn an idea into a commitment by the group.
Similarly, “Luca will introduce the discussion” adds both a name and a decision absent from the message. There is no Luca in our example. Even a reasonable suggestion can change a communication when presented as an established statement without support. The question is not only whether the idea might be good, but whether the text describes its status correctly.
The same attention applies to “must”, “it is necessary” and “it has been decided”. Ask which passage supports the obligation or agreement. If none does, rewrite it as a proposal or leave the choice open. Do not invent reliability percentages to compensate for missing information. A number added by a model does not document a decision.
Preserve the type of claim when rewriting
A revised note might read: “The message states that the room is available from 6 pm and requests one page with a question about the article. It does not specify the meeting’s start time. The person introducing the discussion has not yet been chosen; I suggest agreeing this before Thursday.” The final clause is an original proposal, clearly separate from the information being reported.
The CDC guidance on the state of the science recommends making knowns and uncertainties visible. It concerns public health communication, not certification of AI responses. Here our editorial choice is to make the boundary between information, an open question and a proposal readable, without adding explanations about people.
You can ask AI to annotate the sentences initially, but check every label against the original. If a sentence combines information and interpretation, divide it. If you cannot support a step, remove the claim or formulate the question needed. A useful result is text whose reasoning you can explain, rather than a sequence of reassuring labels.
Reuse only what you can explain
Before using the note, read it without your internal annotations. Would a colleague understand which details come from the message, what remains undefined and which idea you are proposing? If uncertainty appears only in your private notes, bring it into the wording intended for the reader. Keep the context needed to avoid a stronger meaning than the evidence supports.
You can practise with a short invented message or non-clinical material of your own. To explore an AI workspace for working on drafts, you can try PsyLab for free. Checking the source, distinguishing information from interpretation and taking responsibility for what you communicate remain your tasks.