Generative AI and the Case of the Imperfect Apple Pie Recipe

Generative AI like ChatGPT and Gemini continues its rapid evolution. Recent rumblings suggest these AI models possess enhanced text interpretation capabilities. Naturally, I became curious. Could they truly transform a cherished, yet flawed, artifact of the past? I decided to put them to the test.

My mother’s apple pie recipe, passed down from her mother, exists as a handwritten note. It’s stained, faded, and a little cryptic, scrawled on a notepad from a doctor’s office – a testament to a pre-digital era. It’s also the key to a truly spectacular pie.
>

With Thanksgiving on the horizon, the thought of losing this recipe nagged at me. So, an idea sparked: what if AI could create a “better” version, preserving its essence for posterity? I envisioned a delightful infographic, complete with accurate instructions and charming illustrations.

I started with Gemini. I uploaded a photo of the handwritten recipe and prompted the AI to turn it into an entertaining infographic with clear cartoon-style art, ignoring any extraneous information on the paper.

The initial result? Visually appealing, undoubtedly. Usable? Not so much.
>

Right away, bizarre errors cropped up. “9/4 – 1 cup All tbrp All purrose” appeared beside a picture of sugar. Misspellings abounded – “tablesoposs,” “cixttre,” “Suples,” and the perplexing “Dash nutamon.” My handwriting isn’t stellar, sure. Still, I figured Gemini would leverage its supposed understanding of apple pie creation to fill in the gaps. Apparently, I overestimated it. The AI recommended “Sprinkle Spitter” on top of the crust, whatever that means.

It’s worth noting that simply translating text-to-image wasn’t the goal. The aim was comprehension, inference, and ultimately, improvement. This challenge exposed a critical gap in the AI’s abilities.

I gently critiqued Gemini’s attempt, pointing out the errors, the superfluous words, and the puzzling cheese grater (completely unnecessary for apple pie). I asked it to try again.

Gemini apologized and dutifully revised the infographic. Yet, the second iteration fared little better. The misspellings persisted, “Dash nutamon” remained stubbornly in place.

A different strategy seemed required. I became more explicit, demanding only “real words” and emphasizing the need to apply pie-making knowledge to ensure coherence. The pie crust instructions also appeared to be a point of confusion, so I separated those out.

The third attempt showed slight progress. The cheese grater remained, but the nonsensical phrases began to dwindle.

It’s tempting to write this off as a failure of AI image generation, but the deeper issue is AI’s ability to reason. Even with clear prompts, the AI struggled to bridge the gap between raw data (my messy handwriting) and practical knowledge (how to bake a pie).

Next, I turned to ChatGPT 5.1. Its output offered marginal improvement. The art proved cute and functional, but the process lacked key details. It didn’t seem to understand what a rolling pin does, and the apple slice precariously perched atop the pie defied logic. Mustard on the pie? I think not.

Claude Sonnet 4.5 took a different approach. It generated the most accurate recreation of the recipe. However, it ditched the charming artwork entirely. The result felt less like an infographic and more like a bare-bones set of instructions. While accurate, it lacked the personalized touch I had hoped for.

This whole exercise underlines a crucial reality: despite their impressive capabilities, these AI models can falter on seemingly simple tasks. They may transcribe and copy, but their ability to understand, interpret, and improve remains a work in progress.

Imagine if I’d trusted these AI renditions and discarded the original recipe. It doesn’t bear thinking about! It highlights the importance of skepticism when adopting new technologies, especially when they interact with something deeply personal.

My handwritten recipe, with all its flaws, holds a certain magic. It’s a tangible link to my mother and grandmother. It’s a reminder that perfection isn’t always necessary, especially when it comes to apple pie. This experiment confirmed my suspicions. For now, at least, the AI can’t replicate that. The AI can assist. It can process. But it can’t truly create in the same way. That requires something more – a touch of human experience, perhaps. It requires a connection to the past, a willingness to embrace imperfection, and maybe just a dash of “nutamon,” whatever that may be. And these are all qualities that the AI, at present, cannot match.

I will continue to refine the prompts and iterate on the process, because i think the future holds more promise than the current reality exhibits.

Read Original

Keywords: Generative AI, Apple Pie Recipe, ChatGPT, Gemini, AI recipe generation, Handwritten recipe, AI limitations, AI and cooking