Generative AI is very good at one particular thing: taking a lot of text, understanding it, and handing you something useful back. Feed it a specification, a stack of technical documentation, an RFP, and out comes a summary, a report, sometimes working code.
It is also very good at producing an enormous amount of content. And there is a real temptation to do exactly that, simply because we can.
So when you throw a forty-page AI-generated report over the wall to a colleague, it is worth asking an uncomfortable question. What have you actually produced that is useful? Or have you made slop, and handed someone else the job of finding out?
The validation trap
In the old days, the volume of information coming at you was low enough that checking it was proportionate to the job. You did not spend one minute asking a question and then fourteen hours validating the answer.
That ratio has broken. You can write one line and get forty thousand words back.
If those forty thousand words are in your field, and you can validate them, critique them and push back on them, then crack on. That is AI doing exactly what it should.
But if you cannot validate it, something else happens. You start trusting the output because it is there and it reads plausibly, in the same way people follow a satnav into a field instead of reading the road signs in front of them.
Do that to produce something for someone else to review, and you have generated a large volume of work that nobody asked for. It looks impressive. You feel productive, because a long, detailed document now exists and you made it. But if you have no way of knowing whether any of it is true, all you have done is move the work onto somebody else.
Most people are not being malicious. They are doing it with good intentions: “I need to produce a report, I do not really understand the detail, my colleague can check it for me.”
Lazy, but well meant.
The output is a black box
Here is the deeper problem. The final output does not contain your intention. It contains whatever the model decided to produce. It does not carry the prompt you wrote, the sources you chose, or the thinking you did on the way.
Hand over the finished text and it behaves like a one-way filter. Picture a tall stack of source material pouring through: some of it discarded, some rearranged, some newly invented, all according to instructions nobody else can see. What comes out of the spout is a small subset of what went in, reshaped.
If all your reader gets is that final document, they are looking at a black box. There is no way to work backwards, so they are left guessing:
- What question were we actually trying to answer?
- Which sources went into this?
- What steers, biases or misreadings did the person prompting introduce?
Every one of those decisions is already baked in, permanently and invisibly. If your reader cannot inspect the process or get back to the source, the output is unverifiable. Unverifiable work is worth very little.
While you are in the room, you can carry it. You can explain why the AI went with green and yellow rather than red and black, and defend the choice. But documents outlive meetings. Once it is circulating without you, one of two things happens. People work out there is no value in checking it, and quietly ignore it. Or, worse, they trust it, and assume an expertise went into it that never did.
Share the prompt, not the output
The piece of paper with the report on it is not the valuable part. The thinking behind it is.
So if you are not the domain expert, stop shipping the artefact. Tell people what you were trying to give them, hand over your context, and let them drive the AI themselves.
When you share the sources, the context and the prompt, two things change. Everyone can see the foundation, so they know exactly what information is in play. And everyone can steer it themselves: a designer asks design questions, a developer asks developer questions, and the person who should be setting the direction is the one setting it.
Tools like NotebookLM make this practical. Put your private source material in, generate your report or your slides, then share the notebook rather than the export. Whoever receives it can see the sources you used, the prompt you wrote and the context you added. If it is not what they need, they keep your sources and write a different prompt.
Three rules
To stop teams drowning in this stuff, we stick to three.
If you can verify it, crack on. If you have the expertise to judge both the inputs and the outputs, you know what good looks like, and the work is for you, then generate away.
If you are not the expert, do not dump. Do not send an unverified multi-page document to someone just because you could produce it in seconds. You are spending their time to save your own.
Show your working. Share the prompt and the context instead. Let the expert run the query, change the direction and find the value.
The output was never the artefact. It was always the thinking.