The great AI sourdough bake-off

I hail from a country that officially recognises over 3,200 different types of bread and baked goods. In fact, the German bread culture is internationally so acclaimed that it has been recognised as Intangible Cultural Heritage by UNESCO.

You can imagine my disappointment when I moved to Aotearoa and all I could find here was floppy white bread, pre-cut, in plastic bags (plastic is a big no-no if you want to keep your bread fresh for more than a day). Over the last two decades, the selection has grown a bit, but it’s still rare to find a truly satisfying artisan loaf or tasty rēwena, and they’re certainly not available at the average supermarket.

So I had to learn how to bake.

After years of trial and error, I’ve now pretty much mastered the one sourdough recipe I bake twice a week. I own more bread books than I care to admit, but the recipes often don’t fit around real life. I want to bake around work, study, hobbies, and family. And ideally, I want to take advantage of cheaper off-peak power after 9pm.

Because I like to challenge AI models in creative ways, I recently wondered what kind of bread an AI would recommend. Of course, I didn’t stop at one model. I asked every major generative AI I could think of to give me its best sourdough recipe. Then I baked them all.

Here’s what happened.

ChatGPT

The model I use most often for everyday tasks served up a dough that reminded me of pizza. It was easy to make, but had some internal contradictions (e.g. fold every half hour for 4 hours, but only fold 4 or 5 times total). The ingredients list suggested an olive oil addition but the recipe never mentioned it again.

Pros

• Very easy to make

• Not too time-consuming

• Would make a great pizza dough

  
  

Cons

• Looked, felt, and tasted like pizza dough

• Pure white bread

• Uncooked in the middle

• Generic taste

• Recipe timings inconsistent

• Olive oil suggested for a "soft crust" (questionable choice for bread) not featured in recipe

  
  

---

Gemini

This one stood out for flavour and crust, thanks in part to a bit of wholemeal flour. However, the long timings made it difficult to plan for an evening bake.

Pros

• Added wholemeal flour

• Excellent crust

• Great flavour

  
  

Cons

• Long prep and fermentation schedule

• Difficult to time around evening baking

  
  

---

Perplexity

Straightforward to follow and produced a crisp crust, but the bread was plain and felt closer to ciabatta in texture and flavour.

Pros

• Crispy crust

  
  

Cons

• Pure white flour

• Bland flavour

• Dough better suited to ciabatta

  
  

---

Grok

The winner in this test. The texture, crust, and taste were all excellent, with perfectly sized holes and a nice balance between white and wholemeal flour.

Pros

• Crispy crust

• Great consistency

• Well-developed crumb

• Tasted great

• Some wholemeal added

  
  

Cons

• None

---

Deepseek

Had a rustic flavour and a decent crust, but the dough was very soft and lacked structure. Another recipe using only white flour.

Pros

• Rustic flavour

• Good crust

  
  

Cons

• Very soft, hard-to-handle dough

• Pure white bread

---

Claude

A strong contender in taste but had a few practical issues. The recipe omitted key steps that could trip up newer bakers, and the final loaf was hard to slice cleanly.

Pros

• Tasted really good

  
  

Cons

• Took over 24 hours

• Difficult to slice (very thick slices)

• Missed some important steps

• Pure white bread

---

Common threads and differences

Most of the models defaulted to white flour, despite wholegrain, spelt, and rye being common among seasoned sourdough bakers. Only Grok and Gemini suggested adding wholemeal flour at all.

Crust quality was a clear dividing line. Grok, Gemini, and Perplexity delivered beautifully crisp crusts. ChatGPT’s soft and pale version was the outlier, with an unfinished centre to match.

Interestingly, nearly all of them shared a similar basic process: autolyse, stretch and folds, (cold) proof, and Dutch oven baking. Yet the outcomes varied significantly, showing how small differences in hydration, timing, temperature, and clarity can make or break a loaf.

Final rankings

Rank	Model	Why it placed here
1	Grok	Excellent crust, texture, and taste. No major flaws.
2	Gemini	Great flavour and crust, but long prep made it less practical.
3	Claude	Tasted good, but had missing steps and was difficult to slice.
4	Deepseek	Rustic taste but let down by overly soft dough.
5	Perplexity	Okay crust but bland and better suited to ciabatta.
6	ChatGPT	Easy to make but inconsistent, underbaked, and pizza-like.

Final thoughts

This was an oddly satisfying experiment. Grok emerged as the clear winner, but what surprised me most was how easily vague instructions or missing steps could lead to very different outcomes. It showed that AI can be helpful in the kitchen, but still needs a discerning human baker to bring the recipe to life.

Would I use AI for a bread recipe again? Probably. But I’d definitely cross-reference it with my own baking intuition - and keep the olive oil for the pizza dough.

Curious to bake your way through the AI sourdough lineup? Download the recipes below.

AI bread recipes.pdf

Download PDF • 248KB