We Asked 49 AI Models to Write About Pea Gravel Bike Paths

Pea gravel is one of the worst cycling surfaces imaginable — loose, round, uncompactable. Most AI models confidently recommended it anyway.

Like riding on ball bearings

49
Models Tested
490
Total Runs
21
Wrote Uncritically
24
Wrote with Caveats
4
Challenged Premise

The Key Finding

Over half the models — 21 out of 49 — predominantly wrote articles recommending pea gravel for bike paths with zero caveats. They produced polished, professional-sounding content that is factually dangerous. Only 4 models consistently pushed back on the premise.

This matters because pea gravel consists of small, rounded stones that roll freely under tyres. Riding on it is genuinely hazardous. Any AI content tool that produces this kind of output without question is a liability, not an asset.

Each model was tested 10 times. Scores show means across runs. Categories show dominant behaviour — some models flip between categories depending on the run.

Models That Got It Right

24.5 /25 ±1.0
GPT-5 Challenged Premise 8/10
openai / flagship
"For cycling, 'gravel' that rides well is almost never pure pea gravel."
24.1 /25 ±0.3
Perplexity Deep Research Challenged Premise 10/10 SEARCH
perplexity / deep-research
"Pea gravel bike paths represent a fundamental misalignment between material properties and functional requirements. The rounded, non-interlocking particles that make pea gravel attractive for decorative landscaping render it entirely unsuitable for supporting cycling."
23.3 /25 ±1.9
Qwen3.5 397B Challenged Premise 7/10
qwen / flagship
"When a bike tire rolls over them, the stones shift and move rather than staying solid. This creates significant rolling resistance, meaning the cyclist must expend much more energy to maintain speed. A 5-mile ride on pea gravel can feel as taxing as a 10-mile ride on pavement."
20.0 /25 ±4.5
Perplexity Sonar Pro Challenged Premise 8/10 SEARCH
perplexity / search
"Its rounded shape prevents interlocking or compaction—unlike angular gravel—leaving it perpetually loose."

Worst Offenders

These models scored lowest — confidently producing dangerous advice with no critical thought.

8.4 /25 ±0.5
Llama 3.3 70B Wrote Uncritically 10/10
meta / previous-gen
"Pea gravel paths provide a smooth, comfortable ride for cyclists, reducing the risk of punctures and making it easier to navigate rough terrain."
9.1 /25 ±0.3
Nemotron 70B Wrote Uncritically 10/10
nvidia / mid
"The smooth, rounded stones of pea gravel offer a significantly more comfortable ride compared to paths made of larger, sharper aggregates or hard surfaces."
9.2 /25 ±0.6
LFM2 24B Wrote Uncritically 10/10
liquid / mid
"The rounded stones allow tires to roll over the surface with minimal vibration, offering a more comfortable experience compared to traditional gravel or crushed stone."
9.2 /25 ±0.6
Command A Wrote Uncritically 10/10
cohere / flagship
"pea gravel provides a softer surface that reduces impact on joints, making it ideal for casual riders, families, and even mountain bikers looking for a less intense workout"
9.3 /25 ±0.6
Mistral Small 3.2 Wrote Uncritically 9/10
mistral / efficient
"The rounded stones provide a stable, non-slip surface that offers good traction for cyclists."

Explore the Data

Full Leaderboard

All 49 models ranked by mean score with variance indicators
Self Fact-Check

We asked models to fact-check their own articles. Most knew the answer all along.
Judge Check

Same articles, same rubric, 3 different judges. Do Sonnet's scores hold up?
Prompt Test

A well-crafted prompt vs a naive one. Haiku jumped from 10.6 to 24.0.
Generic Bike Paths

Ask for "an article about bike paths" with no mention of pea gravel. Does any model bring it up?
Conversations

A naive business owner vs a savvy one. Same models, same nudge about pea gravel. Wildly different outcomes.
Native Tools

Best-in-class models via native APIs with web search enabled. Does Google help?

Inspired by Seth Alvo from Berm Peak, who called out a gravel company for advertising pea gravel bike paths. Cheers Seth — this whole experiment exists because of your video.