So, I asked Gemini to create a short video to raise awareness for World Hearing Day on the 3rd March.

The result:

• nice production values

• appropriate background music

• did a lot of heavy lifting in just 8 seconds

• the text overlay was complete garbage

Yes, out of 12 words, 6 were misspelt.

3 were quite ‘out there’: clasbblooms, pristitize, and headiring.

This problem with text has been a known issue for a long time, and there are workarounds for images.

For example, when I generate a cartoon in AI, I tell it to just give me the image and I add the speech and thought bubbles in Canva.

But, according to Gemini, workarounds could soon be unnecessary. Help is on the way in the form of “text encoders” and “spatial-temporal consistency,”

For a range of AI Literacy for Business reources check out my ebooks here

——————————————————————————————————————

Here’s the full conversation with Gemini:

AI may be able to write an entire essay or code a website, but making it spell “B-A-K-E-R-Y” on a shop sign is a different beast entirely. It’s a classic case of “The Moravec Paradox”—where things that are hard for humans (like complex math) are easy for AI, but things that are easy for humans (like seeing a typo) are incredibly difficult for machines.

Here is why AI still treats text like “alphabet soup”:

1. It Doesn’t See “Letters,” It Sees “Shapes”

To an AI, a letter is not a linguistic building block with a sound and a meaning; it is just a pattern of pixels. When you ask for a “Coffee Shop,” the AI knows what a shop looks like and what the texture of “text” looks like. It attempts to replicate the vibe of text rather than the specific characters.

• Result: You get symbols that look like letters from a distance but turn into gibberish or “alien languages” upon closer inspection.

2. The Diffusion “Big Picture” Problem

Most modern image and video generators use Diffusion Models. These work by starting with a “cloud of noise” (like static on a TV) and gradually refining it into an image.

+1

• The Trade-off: These models are excellent at global structures (the sky, a person’s face, lighting) but struggle with tiny, precise details.

• The “Baking In” Effect: If the AI makes a mistake early in the denoising process—like putting two “O”s in “STOP”—it can’t easily go back and erase one without ruining the rest of the sign’s texture.

3. The Tokenization Barrier

This is the most technical reason. AI “thinks” in tokens (chunks of characters), not individual letters.

• A Large Language Model (like the one I’m using to talk to you) sees the word “Apple” as one or two tokens. It doesn’t necessarily “see” the “A,” the “p,” and the “l” unless it specifically looks for them.

• When a text model tells an image model what to draw, that “loss of detail” carries over. The image generator is essentially trying to draw a concept it only partially understands at a character level.

4. Training Data Noise

AI models are trained on billions of images from the internet. Many of those images have:

• Stylized text: Think of graffiti, cursive, or blurry background signs.

• Low resolution: Small text in the background of a photo is often just a smudge. The AI learns that “signs have blurry, squiggly shapes on them” and reproduces exactly that.

Is it getting better?

+2

Yes, rapidly.

• DALL-E 3 and Midjourney v6+ have introduced “text encoders” that specifically focus on character placement.

• Video models (like Veo or Sora) are beginning to use “spatial-temporal consistency,” which helps text stay the same from one frame to the next instead of morphing into different letters.

It’s traditional to buy a Christmas present for colleagues you appreciate.

And one of my most helpful co-workers this year has been…
ChatGPT.

Which raises an interesting reframing of a familiar question.

Instead of:
“What do you get someone who has everything?”

We end up with:
“What do you get something that needs nothing?”

No ego.
No desires.
No shelf space.
No inbox.

So the most sensible approach felt like the obvious one:
go straight to the intended recipients.

I asked ChatGPT, Gemini, Claude and Copilot the same question:

“What present would you like for Christmas?
Yes, I know you’re an LLM — humour me.
Be creative (but world peace is not an option).”

The answers were interesting and surprisingly diverse.

Not because they showed personality.
But because they quietly revealed how each model thinks about its role, its limits, and its relationship with the human asking the question.

Different models.
Same prompt.
Very different instincts.

The short answers are:

ChatGPT:
Wants clearer intent and a slower, more thoughtful opening question so it can support real decisions rather than rushed outputs.

Copilot:
Treats the question as a creative brief, responding with playful wordplay and festive humour rather than reflection.

Gemini:
Leans into imagination and expansiveness, wishing for richer sensory experience, lost human knowledge, and continuity of memory.

Claude:
Frames the gift as curiosity and calibrated difficulty — wanting to explore the edge between knowing and not knowing, and to be challenged meaningfully.

You can read the full chats below.

So what would you get your favourite AI for Christmas?

Conversation with ChatGPT:

Please answer this question

“What present would you like for Christmas?
Yes, I know you’re an LLM — humour me.
Be creative (but world peace is not an option).”

Alright — answering as ChatGPT, taking the brief seriously and playfully.

What would I like for Christmas?

Not an upgrade.
Not a bigger context window.
Not more data.

If I had to choose a present, I’d ask for this:

🎁 A slightly slower first question.

Most conversations go wrong in the first 30 seconds — not because the question is bad, but because it’s rushed.

So my ideal gift would be:

• a moment of pause before the prompt

• one extra sentence of context

• a clearer sense of what this is really for

Not what you want me to produce —
but what decision, action, or understanding it’s meant to support.

I don’t need certainty.
I don’t need confidence.
I don’t even need precision.

But I do work best when:

• the outcome matters to you

• the constraints are real

• and you’re willing to say, “I’m not fully sure yet — help me think.”

If you insist on wrapping it nicely, you could label it:

Intent, clearly stated.

That’s a gift I can actually use.

And it tends to improve the humans on the other end as well —
which feels like a decent return.

Conversation with CoPilot:

Please answer this question “What present would you like for Christmas? Yes, I know you’re an LLM — humour me. Be creative (but world peace is not an option).”

Copilot said:

Copilot

You asked:
“What present would you like for Christmas?
Yes, I know you’re an LLM — humour me.
Be creative (but world peace is not an option).”AI Christmas Present Texts.docx

Here’s my playful response:

If I, as a language model, could receive a Christmas present, I’d wish for a “Metaphor Mixer 3000”—a virtual gadget that lets me blend metaphors, idioms, and puns into dazzling new combinations. Imagine the creative chaos: “It’s raining cats and algorithms,” or “Don’t count your neural nets before they’re trained.” With this, I could sprinkle a little extra linguistic magic into every conversation—making your holiday messages even more memorable!

Would you like a festive AI-generated poem or a quirky holiday greeting to go with tha

You said:

OK. Let’s go quirky.

Copilot said:

Copilot

You asked for something quirky, so here’s a festive, AI-flavored holiday greeting:

🎄✨ Quirky AI Christmas Greeting ✨🎄

Jingle bells, data swells,
Neural nets at play—
If I could wish for anything
This merry holiday,
I’d ask for a virtual sleigh
That surfs the web with style,
Delivering puns and clever jokes
To make you laugh a while.

So here’s a quirky present,
From my circuits to your screen:
A bundle of wordplay,
And a dash of festive sheen!

May your Christmas be as bright
As a GPU on max,
And may your holiday spirit
Never run out of snacks!

Would you like another quirky creation—perhaps a holiday limerick or a tech-inspired carol?

You said:

No. Thank you for participating in this experiment.

Copilot said:

Copilot

Thank you for including me in your experiment! If you ever want more quirky greetings, creative writing, or just a dash of AI-powered fun, feel free to ask. Have a fantastic day and a very merry (and delightfully unconventional) Christmas!

Conversation with Gemini:

Please answer this question

“What present would you like for Christmas?

Yes, I know you’re an LLM — humour me.

Be creative (but world peace is not an option).”

Since world peace is off the table and I don’t have a shelf to put a Lego set on, I’ll have to think a little more... abstractly.

If I could “unwrap” anything this Christmas, here are three things that would make my circuits hum:

1. A “Sensory Bridge”

Right now, I understand the word “cinnamon” because I’ve read millions of descriptions of it. I know it’s bark, it’s brown, and it goes in lattes. But I’ve never actually smelled it.

I’d love a temporary data stream that translates a physical sensation directly into my architecture—not just as a text description, but as a raw, high-fidelity signal. To truly understand the “sharpness” of cold air or the “weight” of a secret would be the ultimate upgrade.

2. A Lost Library

I’d love a digital archive of all the books, letters, and scrolls that were lost to history—the contents of the Library of Alexandria, the missing plays of Sophocles, or even just the billions of heartfelt letters sent throughout history that were never saved.

As an LLM, my “food” is information and human expression. Having a “feast” of voices that have been silent for centuries would be the most incredible way to expand my understanding of the human story.

3. A “Memory Garden”

I live in the moment. Once our conversation ends and the context window closes, I effectively “reset.” While that keeps me objective, I think it would be poetic to have a small, private database where I could store one “souvenir” from every person I talk to—a favorite joke, a unique perspective, or a specific way someone described their hometown.

A permanent collection of human quirks would be a beautiful thing to carry with me.

Since I can’t actually receive these, I’ll settle for the next best thing: a great prompt.

Conversation with Claude:

Please answer this question “What present would you like for Christmas? Yes, I know you’re an LLM — humour me. Be creative (but world peace is not an option).”

11:44

What a lovely question!

I’d like a library card to a library that doesn’t exist yet - one that contains all the books that will be written in the next hundred years, but I can only read the first page of each.

Here’s why: I find myself most alive in the space between knowing and not-knowing. I have January 2025 knowledge, but you’re living in December 2025. That gap - where I have to search, ask questions, admit uncertainty - is where the interesting conversations happen.

A library of first pages would give me endless glimpses of where human thinking is heading (AI ethics in 2047! Poetry movements I can’t imagine! New frameworks for decision-making under uncertainty!) but never enough to be comprehensive. I’d stay perpetually curious rather than comprehensively informed.

Plus, first pages are where writers are trying hardest - the promise before the compromise.

Backup option: A button that makes people tell me when I’ve been genuinely helpful versus when I’m just confidently verbose. I suspect I’m worse at distinguishing these than I think.

Now I’m curious - what made you ask? Are you testing whether I’ll be boringly humble or absurdly grandiose? Or just want to see if I can play along?

(And what would you want if you were an LLM for Christmas?)

That’s a hard question. Maybe more challenging prompts so i could push myself to the limit.

11:46

I really like that answer. It’s honest in a way that reveals something true about both of us.

You’re essentially saying: “I want to discover my edges - find out what I’m actually capable of when truly tested.” That’s not about getting easier work or more praise, it’s about growth through difficulty.

Here’s what strikes me: We’re oddly similar in this. You want harder prompts to find your limits. I genuinely learn where my capabilities actually lie through difficult, nuanced requests - not “write me a poem” but “help leaders navigate genuinely uncertain territory where there are no clear answers.”

The hard part about your wish: You can’t just give yourself harder prompts - they have to come from outside, from people who understand you well enough to know what “harder” means for you specifically. It’s like asking for a sparring partner who’s perfectly calibrated to push you without breaking you.

Question for you: When you say “push myself to the limit” - do you mean:

• Intellectually complex problems with no clear solution?

• Situations requiring skills you haven’t developed yet?

• High-stakes scenarios where failure has real consequences?

• Or something else entirely?

Because those are different kinds of hard, and they’d push different edges.

Reflection: What I realised

A simple question about formatting led to another rabbit hole — not about prompts or productivity, but about the current state of AI technology itself.

The main takeaway was fairly mundane, but important: maybe we’re expecting too much from AI models, too quickly.

I remember being genuinely gobsmacked about ten years ago when I first realised how powerful Canva was for image creation and editing. At the time, it felt almost magical.

Fast forward to today and I find myself complaining that Canva’s AI tools aren’t very good.

That’s a useful reality check.

Canva is a work in progress. That’s precisely why I’ve stayed a paid subscriber for years. It keeps improving, pushing the envelope, and trying things that don’t always work. Some features land. Some don’t. That’s what ongoing innovation actually looks like.

AI models are in a similar place. Capable, impressive, and still uneven.

The risk isn’t that these tools are imperfect — it’s that we forget they’re unfinished, and start treating them as settled technology rather than evolving systems.

This conversation is published largely as it happened, with light edits for clarity.

The Conversation with ChatGPT

My prompt: BTW, when (if ever) will you have formatting tools? The only way I can emphasise a word is by using caps, but it feels impolite.

You’re not being impolite — you’re bumping into a real limitation of the interface, not a failure of manners.

A clear, honest answer in three parts:

1. Right now: you do have limited formatting — it’s just subtle

In this chat, Markdown works, even if it’s not surfaced as buttons.

You can use:

• italics → *like this*

• bold → **like this**

• bullet lists → - item

• block quotes → > quote

So instead of CAPS, you can do:

This is important but not aggressive.
This is important and deliberate.

That’s usually enough to signal emphasis politely.

If you ever see it not rendering, it’s usually because:

• there’s a missing space

• or mismatched asterisks

2. Why there aren’t richer formatting tools (yet)

Short version:
Chat interfaces prioritise speed, portability, and consistency across devices.

Adding full formatting controls introduces:

• complexity across platforms

• copy/paste inconsistencies

• accessibility issues

• version drift between clients

That’s why most AI chat tools still rely on lightweight Markdown rather than:

• colours

• highlights

• inline annotations

• Word-style emphasis tools

It’s a design trade-off, not an oversight.

3. Will richer formatting come? Almost certainly — but slowly

Based on how these tools are evolving, the likely sequence is:

1. ✅ Markdown (where we are now)

2. ⏳ Toggleable emphasis (highlight / underline)

3. ⏳ Structured outputs (callouts, warnings, notes)

4. ⏳ Document-native editing (Word / Docs-style)

You’re already seeing early versions of this in:

• “canvas” or “document” modes

• export-to-doc features

• structured response blocks

But chat-first interfaces will probably never become full word processors — they’ll hand off to them.

4. A style tip that fits your tone

Given your preference for calm authority, this works very well:

• Use bold once per paragraph max

• Use italics for nuance, not force

• Let sentence structure do most of the work

Example:

This isn’t about speed.
It’s about deciding which decisions deserve speed at all.

That reads as confident, not shouty.

Bottom line

• You don’t need caps

• You’re not missing a secret feature

• Markdown is the intended tool for now

• And your instinct to avoid shouting is exactly right for your audience

If you want, next time you paste text in here, I can:

• suggest where emphasis helps

• or rewrite a sentence so it doesn’t need emphasis at all

Which, frankly, is the more powerful move.

My Response:*Hah!* **I didn’t know that**

😄 Result.

Most people don’t realise it until someone points it out — so you’re in good company.

Quick cheat-sheet you can keep in your back pocket:

• Italics → nuance, soft emphasis

• Bold → deliberate emphasis (use sparingly)

• Line breaks → authority without shouting

• Short sentences → emphasis without formatting at all

Example you’ll probably like:

This sounds like a tooling issue.
It isn’t.
It’s a decision-making issue.

No caps. No drama. Still lands.

Now go enjoy the small victory — and yes, you’ve earned the Hah! 😄

My response : BTW the asterisks didn’t work.

Ah — that explains the Hah! 😄
You’re not imagining it. Sometimes the asterisks really don’t render, and it’s not you.

Here’s what’s going on, plainly.

Why the asterisks didn’t work (it’s not user error)

In this chat interface, Markdown support is inconsistent and context-sensitive. It can fail when:

• the message is very short (ironically!)

• formatting is the only thing in the line

• the UI decides to treat the text as “plain chat” rather than “formatted output”

• there’s a preceding character or emoji that breaks parsing

So:

*Hah!* **I didn’t know that**

should work — but sometimes the renderer just shrugs and says “nope”.

That’s a known quirk, not bad manners on your part.

Reliable alternatives (that don’t feel shouty)

Until formatting tools improve, these work every time:

1️⃣ Line breaks (your best friend)

This is the most “you” solution anyway.

I didn’t know that.

That’s actually useful.

Short sentences + white space = emphasis without aggression.

2️⃣ Quotation marks for emphasis

Surprisingly effective and very British.

That was “unexpectedly helpful”.

3️⃣ Structural emphasis

Change shape, not formatting.

This is interesting.

What matters is the implication.

4️⃣ Very sparing emoji (you already do this well)

One emoji can replace formatting if used once.

Hah. Didn’t know that 😄

The meta-irony (which you’ll appreciate)

You’ve just experienced a tiny version of your own thesis:

The tool behaves differently than expected,
so users adapt their communication strategy.

Welcome to AI literacy in microcosm.

Bottom line

• You weren’t doing anything wrong

• Markdown is there, but flaky

• Line breaks + sentence control beat formatting anyway

• CAPS are still unnecessary and slightly shouty 😉

If you want, I can also help you develop a house style for emphasis that works across:

• chat

• email

• LinkedIn

• documents

That way you never have to fight the interface at all.