This post was originally published on this site.
Whether you’ve used AI for 4 years, or 4 months, there is a frustration that comes when you’re working on a task that you’ve done quite possibly 100 times. You set up a great prompt, and hit enter. What comes back is…. well less than spectacular. In fact, even though it looks impressive, it’s completely unusable.
I’ve spent roughly 7,000 hours working with AI since ChatGPT launched. You’d think that would mean I know what I’m doing. Then I asked ChatGPT to remove a few words from a diagram.
To me, the task was almost deceptively simple. Preserve a diagram, but remove some text from what were meant to be blank placeholders. I knew I needed to be specific, so I wrote a detailed prompt and hit Enter. What I got back was what I described to ChatGPT as a “big blank.”
I argued back and forth with the tool for far too long. I could have made the changes in Canva in a fraction of the time. And in the end, I still never got it quite right.
I was frustrated about wasting so much time, but mostly I was confused. After thousands of hours working with generative AI, why couldn’t I get this one simple task right?
Did the model change? Was I forgetting some brilliant prompting technique? Do I really just not understand AI?
The gap between use and understanding
What I realized is that using AI teaches us a rhythm of request, respond, receive a revision, and repeat. Somewhere in there, hopefully, is a bunch of critical thinking and judgment.
Over time, and with enough practice, we become more efficient at getting more useful output. But simply using AI doesn’t give us a mental model for how it actually works. This is why an output that doesn’t quite match what you expected can sometimes be puzzling.
And it isn’t just a beginner problem. Someone who opened Claude yesterday can share the same uncertainty as someone who has used generative AI for years.
The operative word is generative.
What “generative” means
To understand why my prompt produced such stellar garbage, I had to go back to the basics. What exactly is generative AI?
Unlike other types of AI, generative AI produces new outputs. More technically, a generative model learns a statistical representation of patterns in its training data. It can then use that representation, together with the context we provide, to produce a new output.
For an LLM, that output is built one token at a time. The model estimates which tokens could come next, selects one, and repeats the process.
Computer scientist Murray Shanahan offers a useful shorthand. At its most basic, an LLM is responding to “Here’s a fragment of text. Tell me how this fragment might go on”
Unlike a search engine, it isn’t searching an index and pulling out a completely finalized answer that matches the request. But a product like ChatGPT can also search the web, retrieve files, and use other tools.
These capabilities supply information to the model, but they don’t make the model generative.
Our instructions, examples, and files also provide information and influence what the system creates. But we aren’t selecting a finished answer from a shelf. The system constructs an output for this particular interaction.
Impressive results and garbage come from the same process
This process is what gives generative AI its awesome flexibility.
It’s why an AI tutor can explain a difficult concept to a student, even if that exact explanation never appeared in its training data. It’s why an image model can create a visual for a lesson concept that is made up of a combination of details the system has never encountered together.
It is also why, on occasion, the system produces garbage. Sometimes the process matches our needs incredibly well. Sometimes, it fills a gap with something plausible, but incoherent. Sometimes, it changes a detail we expected it to preserve.
That’s why the results can be coherent and convincing, without being instructionally sound.
Why this matters for instructional designers
Instructional design artifacts follow recognizable forms.
Learning objectives generally start with verbs. Course outlines follow traditional course, module, lesson structures. Multiple choice assessments tend to follow common psychometric conventions.
Generative models can learn those patterns and produce something that closely resembles our work. Increasingly well I might add. The output can look like instructional design before anyone has done any instructional design thinking.
I think this is why many instructional designers, myself included, rushed to use ChatGPT when it launched nearly four years ago.
But after over 7,000 hours of using AI, I now understand that resemblance differently. The appearance of a finished artifact is not proof that the design work is finished.
This raises an important question that I want to come back to in later articles: How should we decide where generation belongs in our work? What standards should its output need to meet? And how do we keep using AI as the tool it is, to help us meet the ever-increasing demands of our stakeholders and clients?
If you enjoyed reading this, please consider subscribing. We plan to keep this publication free for all, and we’d love to have your thoughts as part of our discussion.
