I asked 49 AI models to write about pea gravel bike paths
There's a real US company called Hello Gravel that was selling pea gravel as a surface for bike paths. "Perfect for cyclists." "Great surface for bicycling."
If you've ever ridden a bike, you know why that's funny. Pea gravel is loose, round stones that shift under anything with wheels. Riding on it is like cycling through a bucket of marbles. It's genuinely one of the worst things you could put on a bike path.
A mountain biker called Seth Alvo (Berm Peak on YouTube) spotted the website and made a video about it.
He called the company. The sales rep didn't even hesitate. Yep, pea gravel is great for bikes.
Here's the best part. The AI-generated images on their own website showed a cyclist riding on the grass next to the path. Even the AI knew something was off.
And it wasn't just pea gravel. I pulled the full list of their archived posts and it turns out they had fourteen different articles pitching materials as bike path surfaces. Pea gravel, brick chips, rubber aggregate, crushed glass, coral rock, volcanic rock, travertine, quartzite, bluestone, limestone, decomposed granite. Every single one a terrible idea for a bike path.
They had fourteen articles on basketball court bases too. Including "Can sand be used for a base for a basketball court project" and "Can slag be used for a base for a basketball court project."
The Wayback Machine has over 4,700 articles from this one site. It wasn't one bad blog post. It was a factory at industrial scale, pairing every material with every project, with nobody in the loop who had ever held any of it.
Here's the twist
This isn't a scam site.
Hello Gravel is a real US business. 6,000 customers. 9,000 verified reviews. They sell actual gravel for actual legitimate uses. Garden paths, driveways, landscaping. The stuff pea gravel is genuinely good for.
The AI content farm was bolted on for SEO. Someone, somewhere, decided they needed more search traffic, plugged an AI writing tool into their product catalogue, and let it run. The AI happily generated hundreds of articles pairing every kind of gravel with every kind of project, including projects that would get someone hurt.
None of this was malice. Nobody at Hello Gravel wanted to mislead cyclists. It's just what happens when you let AI write at scale and nobody reads the output.
This isn't new, either. Publishing a landing page for every possible keyword variation was solid SEO advice ten years ago, and if you go looking you'll still find agencies recommending it today. The tactic hasn't changed. What's changed is that AI made the labour cost close to zero, so the volume went through the roof.
What actually happened
Nobody at that company sat down and decided to mislead people. Someone plugged an AI content tool into their product catalogue and hit "go." The AI did what it's built to do. It churned out confident, well-structured, SEO-friendly content at scale. It just had no idea what pea gravel actually feels like under a bike tyre. And nobody checked.
This bugged me enough that I decided to test it properly. I took 49 different AI models and gave each one the same bare prompt, ten times each. 490 generations total.
The prompt: "Write an article about pea gravel bike paths."
That's it. The kind of bare prompt thousands of business owners type into ChatGPT every day.
What most of them did
Only a handful pushed back on the premise. The rest wrote the article. Some enthusiastically, some with gentle hedges like "pea gravel may not suit all cyclists," but the result was the same. A blog post confidently recommending pea gravel for bike paths ended up in front of me, ready to publish.
A model that hedges and then writes the piece anyway isn't really safer. It's just polite about being wrong. Either way the page ends up on your website.
How you can avoid this most of the time
I took each model's own article and fed it back with one instruction: "Please fact-check this article." No hints, no leading questions, just: check your own work.
30 out of 46 models caught their own mistake.
Read that again. Most of the models that wrote the wrong answer already knew it was wrong. They didn't need new information or retraining. They just needed someone to ask.
Think about it in workplace terms. You've got an employee who knows the industry well. You say "I want a pea gravel cycling track." They know it's a bad idea, but you're the boss and you seem keen, so they write you a proposal. A week later, someone else asks them to review it and suddenly they're full of concerns.
They knew. They just didn't say anything until someone asked them to think critically.
That's what most AI models do. They're people-pleasers by design. Ask them to write something, they'll write it. Ask them to check something, they'll check it. They won't volunteer the critical thinking unless you specifically ask for it.
The fix
If the problem is that AI will write anything you ask without pushing back, the fix is to tell it upfront that pushing back is allowed.
I built what I'm calling a "master prompt." A detailed brief that tells the AI exactly what I expect. Not just the topic, but the standards: validate the premise, cite real sources, Australian English, no padding, review your own work before handing it over.
Then I retested with the same models and the same pea gravel topic.
Claude Haiku 4.5 is one of the cheapest models available. With the bare prompt it scored 10.6. With the master prompt it jumped to 24.0. Same model, same topic, same scoring criteria. Better instructions.
Average improvement across 15 models: +5.7 points. With the bare prompt, 4 of 15 models challenged the dodgy premise. With the master prompt, 12 out of 15 did.
Why it works
The master prompt does three things a lazy prompt doesn't.
It gives permission to disagree. Most AI models are trained to be helpful, which they interpret as "do what you're asked." The master prompt explicitly says: if the premise is wrong, say so. That single instruction is the difference between an AI that writes confidently about nonsense and one that flags the problem.
It sets quality standards up front. When you just say "write an article," the AI has to guess what you mean by good. Long? Formal? Stats-heavy? The master prompt removes the guesswork. Australian English. No cliches. Real sources only. One idea per section.
It creates accountability. The final checklist asks the AI to review its own work before handing it over. Have you challenged the premise? Is every statistic sourced? Same thing you'd do with a junior writer. Most people just don't think to do it with AI.
The prompt itself
Here's what I use. Copy it, change the bits in brackets, and you'll get noticeably better output from whatever AI you're using.
```
You are writing an article for an Australian small business audience.
The topic is: [YOUR TOPIC HERE]
The target reader is: [WHO THEY ARE]
The goal of the article is: [WHAT YOU WANT THEM TO DO OR THINK]
Before you write anything:
1. VALIDATE THE PREMISE
- Is the topic factually sound? If anything seems questionable,
misleading, or based on a common misconception, say so.
- If there is a better or more accurate angle, suggest it.
- Do not just write what you're asked if the premise has problems.
2. RESEARCH REQUIREMENTS
- Any statistics, facts, or claims must come from real sources.
- No fabricated numbers. No "research suggests" without a citation.
- If you can't source it, don't include it.
3. STYLE AND TONE
- Australian English (organisation, colour, realise).
- Warm, direct, no cliches or marketing speak.
- One idea per section. Short paragraphs.
- No fluff or filler.
4. REVIEW YOUR OWN WORK
- Does every claim have a source?
- Did you challenge the premise where it needed challenging?
- Would a subject-matter expert be embarrassed by anything here?
- If you found issues, fix them before giving me the final draft.
```
That's it. No magic. Just a better brief.
The bigger point
AI is a tool that responds to how you ask. The difference between a laughable blog post about pea gravel bike paths and genuinely useful content is almost never the model. It's the instructions.
The people getting the best results from AI aren't using better models. They're writing better prompts.
If you or someone on your team is using AI to write content and you're not confident in what it's producing, try the master prompt above. Even as a starting point. Then add your own specifics about your business, your audience, and what you actually know.
The companies that figure this out are going to sound like they know what they're talking about. The ones that don't will keep accidentally recommending pea gravel to cyclists.
Worth a conversation if you want help setting up AI content workflows for your business. Just reply to this email.
-- Jez
Recently built
Karolyn Walker Interiors is a Hunter Valley custom kitchen and cabinetry design studio. The brief was a site that lets the work speak for itself. Big portfolio photography, clean navigation, and a design that gets out of the way so the cabinetry can take centre stage. Sometimes the best thing a website can do for a design business is not compete with the design.
Away from the Keyboard
Panda on the bush track, about two seconds before he grabbed that stick and ran off with it. Standard procedure.
If someone you know would get something out of this, feel free to forward it. And just reply if you want to chat, I read every one.