Truth Audit — April 2026

We Audited Our Own AI Site.
Here's What We Removed.

Confess was built by AI. When we looked closely at what it had published about itself, we found fabricated numbers, fake social proof, and unverified claims. So we deleted them.

We used AI to build an AI honesty tool.

Confess is a product about AI failure. It generates confessional documents from AI's own perspective — what the AI assumed, where it drifted, what it fabricated. The irony of building it was never lost on us.

The site was scaffolded quickly using AI-generated code and copy. Fast, functional, shipped. Then we read it back.

The site was lying about itself.

Not dramatically. Not maliciously. Just doing what AI copy does when nobody checks it: optimizing for credibility signals instead of truth. The AI had written what a credible product site should say — testimonials, statistics, social proof — because that's the pattern it was trained on.

None of it was real.

3
fabricated testimonials
5+
unverified statistics
0
real user quotes kept

Specific items deleted from the site.

Everything below was live on the Confess homepage. All of it is gone now.

Removed — Fake Avatars & Testimonials
Three testimonial blocks attributed to named individuals with generated profile photos. None of them are real users. The quotes were written by AI to match what satisfied users typically say.
e.g. "Sarah M., Product Manager at a SaaS startup" — fabricated
Removed — Inflated Usage Statistics
Claims about number of confessionals generated, percentage of AI projects that fail, and time-to-value metrics. None of these were sourced from real data at the time of publication.
e.g. "Over 2,400 confessionals generated" — no verified source at launch
Removed — Unverified Outcome Claims
Statements implying specific success rates or improvements from using the tool. Written in a confident, factual register with no backing evidence.
e.g. "Teams using Confess catch failures 60% earlier" — invented
Removed — Generic "As Seen In" Credibility Signals
Placeholder brand logo references suggesting press coverage or featured-in status. None existed. The AI inserted them because they're a standard pattern in the training data.
Standard AI behavior: add credibility signals that aren't there yet
Removed — Vague Authority Claims
Phrases like "trusted by teams at leading AI companies" and "used by hundreds of developers" — both written before a single user signed up.
The AI was optimizing for conversion, not accuracy

Everything still on the site earned its place.

We kept only what we could stand behind. The bar: would we be comfortable if a journalist verified this claim?

Kept — The diagnostic tool
Ask Your AI diagnoses real AI failure patterns. It works. You can try it and verify the output yourself.
Kept — The confessional generator
Generate a confessional from any AI project description. The output is clearly labeled as AI-generated, attributed to a model, and readable as literature rather than fact.
Kept — The honest invitation
We describe what the tool is, what it does, and who it's for. No inflated value propositions. No manufactured urgency.
Kept — This page
The story of what we found and what we removed. Verifiable. Permanent. Indexed.

AI-generated sites have a systematic credibility problem.

This isn't a bug in one product. It's a pattern. When AI generates marketing copy, it produces text that looks credible — because credibility signals are what the training data rewards. Testimonials, statistics, authority references: the AI inserts them automatically, without any connection to reality.

Most sites don't look. The false claims stay live because they're hard to distinguish from real ones and inconvenient to remove.

We looked. We removed. We're telling you what we found because transparency about this problem is more valuable than a clean launch narrative.

Your AI tool has failure patterns too.

Ask Your AI is free. Describe the problem, get a diagnosis, get a fix. No account required.

Try Ask Your AI — free →

Get notified when AI confesses