Biased AI tests can hide a cheaper model
In brief
According to AWS and Property Finder, a real-estate platform found that 96.8 percent of the answers used to score its image models came from the production model, not from human experts. Any alternative therefore looked worse. After independent expert review and instruction changes on Amazon Bedrock, the company reports higher accuracy at 86 percent lower cost.
Source : AWS Machine Learning Blog - « Fix evaluation bias to unlock better models on Amazon Bedrock »
What it changes for an SME
If you score a new AI model with answers from the system already in use, a cheaper option can look worse than it is. An SME may then stay on higher running costs. Having a domain expert annotate a sample without seeing current outputs takes time. The 86 percent cost cut reported by AWS is specific to Property Finder; it is not a result to expect by default, and this item does not change your GDPR or EU AI Act duties.
What you can do
- Check how much of your AI test data was produced by the model you already use.
- Have a domain expert annotate a representative sample with no access to current predictions.
- Reassess quality and running cost on that set before you change tools.
Let's talk about the time you could save.
In 20 minutes, we find where AI could save you time or money. Free, no commitment.