Just in time for our final review of the year, we at the agency are taking another look back at what’s new so far. One of the topics we’ve been discussing is ChatGTP and the new Thinking Mode. Could it make everything better? Is there more to AI than suggesting shortenings and checking for spelling and grammar mistakes? Could our virtual colleague even write the basis for an entire professional article—without any hallucinations?
In practice, it looks like this:
The test run starts off well thought out. We write a detailed prompt, define the target audience, topic, tone, and length. Of course, we put a lot of effort into this, because we wanted to be able to continue working with the results.
Then we wait. Until now, neither the statements nor the sources provided by the AI could be trusted. Our eager colleague ChatGTP likes to make things up to make herself look better. This shouldn’t happen anymore in the Thinking mode. We are excited to see the result. But then what happens is
Silence. Abort. End.
Either the text breaks off in the middle or, even better, nothing appears—Thinking Mode thinks and thinks, and in the end, there is nothing. No sentence, no fragment. And even the sentence “Please finish writing this text” doesn’t work. Because the beginning has long since vanished into virtual fog.
Why the thinking mode sometimes “strays”
In theory, the thinking mode is a blessing, because it actually follows instructions such as “don’t estimate the word count, just count” or “don’t interpret the results, just reproduce what is actually there.” The problem: additional “thinking work” requires resources. It costs computing time, and on top of that, each answer has a limited “word budget” (number of tokens). If this budget is exceeded or the calculation takes too long, it’s game over – the technology says, “Time’s up.”
For us, this feels like a download that stops at 98%—painful for our time budget, emotionally… let’s say: not good for our blood pressure.
Avoiding blackout experiences
At the end of the day, we’ve at least made some progress with this:
- Outline first, text later: First, just create a structure: headings, subchapters, rough bullet points.
- Formulate in stages: Each section gets its own prompt – like individual building blocks for layout and approval. This way, the entire text can’t disappear before anything even lands on the screen.
- Limit the mental effort: Don’t make too many demands at once. What can be done one after the other should also be done one after the other. Formulate as “foolproof” as possible and, if necessary, simply ask the AI itself why it did something a certain way. At least it helps for the future…
ChatGPT is not an author, but a structural aid, a wording machine, or a source of ideas for variations, subheadings, and introductions. What it definitely does not replace: a knowledgeable author. In any case, we prefer to let AI continue to do the smaller auxiliary tasks – that saves time and doesn’t waste it.