Your AI Is Still a Toddler (With a Thesaurus)

Go ahead. Ask your AI image generator to create a picture of scissors.

I'll wait.

Notice something? The handles don't work. The blades don't align. The pivot point defies physics. Your AI knows what scissors look like, but it has no idea what scissors are or what you do with them.

This isn't a quirky limitation. It's a window into what these systems actually are, and marketing leaders are making million-dollar decisions without understanding the fundamental gap.

Fei-Fei Li, often called the "godmother of AI," just published her framework for spatial intelligence. Her core argument: current AI systems don't understand the physical world. They're pattern-matching machines that can probabilistically estimate the next pixel or word, but they have no concept of how objects work, how space functions, or how things interact.

Scissors look simple. Two blades, two handles, a pivot point. But rendering functional scissors requires understanding physics, mechanics, and purpose. LLMs don't have that. They estimate pixels next to pixels, creating scissors-shaped objects that couldn't actually cut paper.

If you know me you are smiling that I am referring to Sudokus. I don't leave home without my Samurai Sudoku book - it is my meditation.

GPT-5 just became the first large language model that can reliably solve Sudoku puzzles. Not GPT-4. Not Claude 3. GPT-5.

Think about that for a moment. We've been using these systems for strategic analysis, customer segmentation, and campaign optimisation, but until now, they couldn't handle logic puzzles that 10-year-olds solve for fun.

The research paper on Sudoku capabilities references what they call "human break-in points", the natural reasoning shortcuts humans use instinctively but AI finds extraordinarily difficult. Things like: "If this number goes here, then that creates a conflict over there, so I'll try something else."

That's not computation. That's reasoning. And AI is just learning how.

I wrote about the hand problem in October. About how AI image generators struggled for years to render human hands correctly. Fingers, knuckles, thumbs, the way hands grip objects. Too complex, too many variables, too much physics.

The hands got better. The scissors reveal the same pattern. These aren't isolated bugs getting fixed. This is what probabilistic systems look like when confronted with tasks requiring actual understanding.

AI systems today are toddlers with extraordinary vocabularies.

They can write brilliant-sounding strategy documents. They can generate persuasive copy. They can analyse sentiment and optimise targeting. But they can't draw scissors properly, they just learned to solve Sudoku, and they fundamentally don't understand cause and effect.

That's not a criticism. It's a description of maturity level.

We're approaching what Gartner calls the "trough of disillusionment" in the AI hype cycle. That moment when:

This isn't AI failing. This is expectations calibrating to reality.

The technology is genuinely useful, but it's toddler-stage useful, not strategic-partner useful. It's extraordinarily good at specific tasks within narrow parameters. It's remarkably bad at anything requiring understanding, causation, or genuine reasoning.

And here's what matters: It will get better. Spatial intelligence research is advancing. Reasoning capabilities are improving. The toddler is growing up.

But right now, today, you're working with systems that can't draw scissors properly or solve logic puzzles. That's the reality. And the marketing leaders who succeed will be those who deploy AI capabilities appropriate to its actual maturity level, not its promised potential.

Not action items. Not vendor evaluation frameworks. Just realistic calibration:

AI is exceptional at:

AI struggles with:

This won't change quickly. Fei-Fei Li's spatial intelligence research is years away from practical implementation. Reasoning breakthroughs like GPT-5's Sudoku capability take generations of model development.

The marketing leaders winning with AI aren't those who deploy it most aggressively. They're those who understand its maturity level and use it accordingly: powerful augmentation for specific tasks, not autonomous strategic thinking.

The toddler has an extraordinary vocabulary. But it's still a toddler.

When Fei-Fei Li's spatial intelligence becomes practical, when reasoning systems move beyond Sudoku, when AI genuinely understands the physical world, that will be transformational.

But that's not today. And marketing budgets deployed today based on tomorrow's promised capabilities are budgets at risk.

The most strategic thing you can do right now? Understand that your AI can't draw scissors properly. And plan accordingly.

The question isn't whether AI is useful. It's whether you understand its actual capabilities well enough to deploy it strategically. Toddlers are extraordinary, but you don't put them in charge of strategy.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 11 November 2025 in the Brandflow newsletter on LinkedIn.