Your LinkedIn feed has filled with this AI slop pandemic in just two years. You scroll past posts that open with a confession and close with a lesson, and past essays that spend four hundred words telling you AI changes things. We call it AI slop. An I am sure by now you can spot it in a second.
But for those who haven’t quite cottoned on to what AI slop is yet, here is the breakdown.
Slop has a recipe
Slop takes about ten seconds to make. I know because I have made plenty of it.
The first time a prompt works, it feels like a magic trick. I typed one line and a finished slide came back in four seconds. From across the room it looked like a morning's work. Up close it said nothing.
The recipe is short. You type one prompt into one model, take the first thing it gives you, and publish it. A person who knows the craft gets no say.
Now, give the same image model to a graphic designer and watch what happens. She prompts it into submission, then throws away forty drafts and keeps just one. And she does this because she knows good work when she sees it. Me on the other hand, I keep the first draft...
We both used the same tool but she brought the judgement and experience, and I had none to bring.
So the recipe to AI slop is simple, it's well written, well articulated, even well designed outputs that to the untrained, experience lacking, subject matter missing, (I can keep going here) person, looks great, but to those that know, it instantly makes them regret ever starting to read, review or consume what you have just placed in front of them.
I guess, dodging AI slop emails, skimming past posts, and now ignoring the plethora of chat GPT birthday invites, has become a skill in it's own, but the challenge is that this slop is now making its way into the classroom, and we are using it to impart knowledge... or rather a lack thereof.
Slop has entered our courseware
I founded SynrgiseLearn. We host training for more than 2 million learners, so a lot of courses cross my desk. This year a new kind started turning up.
You can pick it out the way you pick out the six-fingered hand. Somebody uploaded a policy PDF and pressed generate. Forty slides came back, each a paragraph of the policy beside a stock image. The quiz asks you to recall the heading of section three, and a certificate waits on the last screen.
Instructional design is a craft with a century of research behind it. A designer with fifteen years in the job reads a draft module and tells you which screens will stick. She knows the gap between recognising an answer and recalling it. The person pressing generate has had no reason to learn any of that.
The learner clicks Next forty times. Your LMS records a completion and your auditor ticks a box. On paper you trained four hundred people. On the floor, the same four hundred handle a customer's ID document the way they did last month.
I should own our part. LMS vendors, mine included, spent years rewarding completion counts, and a slop course scores well on completion.
Slop in your feed costs you a second. Slop in a compliance course costs more, because a fluent slide can state a rule your policy does not contain. Four hundred people will then sign that they read it.
We tried the shortcuts first
We tried the shortcuts before we built anything. We pasted policies into chat models and pressed the AI buttons in the authoring suites.
The drafts read well, and that was the trouble. A slide gave no hint of the policy page behind it, so a reviewer sat with the PDF open and checked sentence by sentence. The draft was cheap but the checking was expensive, and more times than not, the draft had tp be thrown out, and restarted again.
Arguing with the model made things worse. A Harvard Business School team studied more than seventy BCG consultants using GPT on a task built to need human checking. The consultants fact-checked the model and pointed out its inconsistencies. The model stepped up its persuasion and defended its own credibility, and it kept its conclusions. The authors call this persuasion bombing.
Now picture your one-person L&D team doing that review alone, against a Friday deadline.
We decided a single model in a chat window is the wrong approach. Basically, you are asking it to read the policy, design the course, write the slides and then, mark its own homework. It agrees with itself at each step.
So we split the work
The recipe for slop is one model, one pass and no expert in the room. Kre8AI changes each ingredient. So instead of relying on a single model, we built it as a swarm of agents working in three stages, and we put a person in-between each of the stages.
Agents in the first stage read. They work through your policies, manuals, decks and recordings, then tell you what they understood. They name the audience and scope they found, and they list the gaps they could not resolve. You correct that account before a word of the course exists.
Agents in the second stage design. They propose objectives, a structure and an assessment strategy, and they show their reasoning. You can then reorder the outline or rewrite it outright. Kre8AI generates no slide until a person approves the design.
Agents in the third stage build from that approved design. They retrieve passages from your source and write from the passages they retrieved. Each slide arrives with up to four citations. A citation names the document, the page range and the excerpt, and it carries a confidence score.
We run more than one model, too. Reading a long policy, planning a course, drawing an image and voicing narration are different jobs. We give each job to a model that does it well.
Then your people take over. Reviewers approve or comment module by module, and Kre8AI logs each decision with a name and a time. The publish button stays locked until a person has approved the last module.
A reviewer who doubts a slide clicks through to the passage behind it. She reads the policy's own words, and thankfully she has no model to argue with.
The approved course goes into the LMS you run, ours or another vendor's, as a SCORM, xAPI or cmi5 package. You can switch on a course coach that answers learners from the same approved material.
Check it yourself
We built a POPIA awareness course in Kre8AI and put it in a free library. Play it in your browser or load the SCORM package into your own LMS. https://www.kre8ai.xyz/resources/courses/popia-awareness
If you find slop in it, write to me. I would rather hear it from you than from a learner.