How we test

Fair comparisons shouldn’t depend on favourites.

We test practical tasks, preserve the evidence and have the results checked separately. Here is the simple version.

Same task.
Saved evidence.
Independent check.

The review journey

Five steps from task to recommendation.

Choose a real task

We begin with things people genuinely ask AI to do—such as editing a book, researching a report or creating a promotional image.

Use the same test

Every AI in a comparison receives the same controlled task and instructions under recorded conditions.

Save everything

The complete prompt, response and relevant test details are preserved as evidence while the review is carried out.

Check it independently

A separate AI evaluator applies the frozen marking guide. The product being reviewed does not mark its own work.

Approve before publishing

Findings, limitations and conclusions are reviewed before they become a public recommendation.

Governance, in plain English

The safeguards stay in the background, not in your way.

AROS is the controlled system underneath AI Reality Check. It keeps reviews consistent and makes it possible to trace a public statement back to recorded evidence.

✓No marking its own homeworkTesting and independent evaluation are kept separate.

✓No hidden universal winnerRecommendations are made by task and user need.

✓No timeless claimsModels, subscriptions, evaluators and dates are disclosed.

✓No pretending AI is certainLimitations and close results are stated openly.

What you can see

Enough detail to judge our conclusions

Tasks tested
Recorded results and useful alternatives
Model, plan and review date
Important limitations
Supporting evidence where suitable

What stays protected

The internal machinery and intellectual property

Full operating controls
Internal administration
Proprietary prompt libraries
Unpublished working evidence
Private business research

Ready to use the findings?

Start with what you want AI to do.

Help me choose