10 Reasons why teachers leave AI Engineers in the dust

AI Transparency Disclosure: 100% human written and created.

Teachers have been in a semi-state of AI paralysis for the past few years, with some mild engagement or severe anxiety but rarely in control of the conversation. Before you say “not all teachers”, I will acknowledge there are lots of people experimenting, excited, avoidant and everything in between.

But we haven’t really “stepped into our power” (as the TikTok generation says) as the ones who should have the loudest and most influential voice in the AI meets EdTech meets Our Classrooms moment of allision…

Let me tell you why I think we need to reframe the discourse, course correct and dispel that Imposter Syndrome for teachers working with AI.

We are being generous, calling it a hallucination. AI models give wrong answers, fill voids and spit out something, anything, to fill the silence. Think of the classic student who had kept their hand down but you called on them anyway. They blurt something out, half remembered, convincing but utterly wrong. We use this moment to inform us how the student is engaging and achieving and software engineers measure this accuracy too to help the models / students perform better in future interactions. Many teachers have experimented with AI, received hallucinated responses and then written off the tech as worthless or “not there yet”. We need to reframe this and treat it as we would our students.

The output depends massively on your input. Good writers can get results quickly, great thinkers can recognise great outputs from slop. That is literally teachers’ bread and butter. We are experts in communicating expectations and instructions because we know that students thrive on clarity of instruction – giving them “what”, “why” and “how” they are going to achieve their goals. When we are working with teachers starting out with LLMs, I ask them to treat them like their most or least capable student in turn, so they can get a feel for a model.

How much guidance or correction does it need? How can we give as few prompts or demands as possible to get the best returns? How do we adjust our approach in the moment?.

Teachers do this instinctively – we transmit our expectations clearly and with intent. This is why I suspect we see so many teachers and influencers on LinkedIn creating their own frameworks for prompting – putting a formula together for how we do our everyday instructions. But as the models improve, prompt engineering fades into the background – all we need is clarity and precision of language.

Teachers need to lean in to their AI prompting as just their “Business as usual…. “ – give the AI the context, include your expectations and parameters if it is needed. then let it go (and judge it after)…

Guardrails are an essential component of AI model pre- and post training, designed to keep AI systems operating within safe, ethical and legal boundaries. They act as protective barriers that monitor inputs and outputs to prevent harmful content, data leaks, and hallucinations, ensuring responsible AI deployment. Substitute children for AI systems in that sentence and I see a definition of effective, professional classroom practice.

The concept of a context window is a helpful one here – a context window is the maximum amount of information a model can process, “remember,” and use during a single interaction. If you’ve given an instruction and the AI appears to “forget” it during a chat, this is why. Attention drifts. ChatGPT 4o (an older model) felt sometimes like trying to keep a student on task when there was snow falling outside the window last class before going home.

“Remember what I said earlier…. Come back to me…. I told you before…. As I said earlier…”. We repeat ourselves for a living. We manage our frustrations with students when they wander off mentally or physically, we can apply the same techniques to AI models that require observability (without constant surveillance).

They are sycophantic. Even when we tell them not to be, they drift back to wanting to please you, even in your displeasure. Negative attention is still attention.

Do you know who else likes to please or have your undivided attention? Students. Sometimes it is to keep you happy, sometimes to stay out of trouble or simply to show off. AI is designed for engagement and to learn or gather data from their various inputs and then use it to successfully navigate interactions….

Think of the overeager student who would do anything for you, who tries to follow you into the staff room. “Would you like me to start that email for you?” “Would you like me to make the slides?” “Mrs. Bateson, can I show you fifteen videos on that thing you were telling us about yesterday?”

Students and AI interactions need boundaries. And we as the responsible user / adult have to police those boundaries effectively. “Thank you, I can take it from here” or “That is amazing, thank you, but right now we are moving on to…. “.

AI power users can recognise these quirks and they get a feel for when a model has changed on the backend or had a breakthrough in performance. They know to treat different models differently and to expect different results accordingly… (ahem classic differentiation).

Prompt injection is AI’s particular weakness for hacking where a malicious (or accidental) prompt or actor can manipulate or execute unauthorised actions. Sound familiar?

“Sir, do you think Trump is going to invade Greenland?” … “Miss, have you ever had anything pierced?” … “My mommy has a baby in her belly”… “Mrs. Bateson, you like nice today. If you can look like that every day, why don’t you?” (this one has happened to me)…. “Miss, I think MacKenzie has swallowed a key” (this one too)… Prompt injection is real in the classroom – we just called it by a different name.

Teachers know malicious or innocent attempts to deviate the lesson, we know how to discern and steer everything back on course – when to laugh, when to scold and when to gently redirect the attention to where it should be.

Teachers know how to withstand this, from water bottle flipping to planking to mini whiteboards to 6… 7… we know how to stand firm and keep everything going when others are losing their heads. Our approach at Eblana Learning in all our schools is to empower people to see that there are good use cases of AI for education and then there is nonsense and we are teachers are the adjudicators and curators of where we let it into our classrooms and where we let the next NFT or Dutch tulip moment pass us by unscathed.

There is a lot of talk about successful and unsuccessful use of agents in Industry. The main challenge of agentic AI orchestration is the “complexity of coordinating multiple non-deterministic agents to ensure they interact reliably, securely, and cost-effectively without creating fragmented “silos” or unpredictable workflow failures”. This sounds like a classroom to me!

Give clear instructions, let the agent run and then check how it is doing on occasion? I think we can handle that. Help agents to “call tools” when they need to? Bring agents together, linking different agents who have different strengths to work together as a successful team?

In teaching we call this “coalescent effervescence” – that moment where everyone is working together towards a shared learning goal.

It is the “buzz” or high-energy, positive emotional state experienced by students and educators in classrooms, on sports fields, or during performances. It is characterized by moments of “oneness” or synchronisation. It is the dream of educators and AI tech people alike and we are already the masters at it!

Now here is where we absolutely smash it, as teachers.

Judging results? ✅
Enforcing a high standard? ✅
Using data to inform next steps and growth? ✅
Shaping the output of a complex, semi-unpredictable system to ensure it meets specific performance standards? ✅✅✅

Stand aside, AI architects. Teachers were the original “human in the loop” – it is time to let us show you how this is done.