Useful, explainable and deliberately limited.

Relevant inputs

Suggestions respond to specific interests and questions rather than inferring a fixed personality from a few answers.

Sensitive characteristics stay out

Career suggestions focus on goals and interests, not health, ethnicity, religion or similar characteristics.

Test and improve

Quality, bias, safeguarding cases and complaints are reviewed systematically.

Risk checks across an answer

We assess more than the model: what input is requested, which source context is supplied, how uncertainty is expressed, which action the interface suggests and how an error can be reported. A capable model can still cause harm inside a poor interface.

• Tests in German and British English.

• Comparison across gender, background signals and education routes.

• Specific tests for minors and safeguarding situations.

• Review after model, prompt or source changes.

Complaint and correction

People must be able to report a problematic answer without publishing it. The team separates factual error, inappropriate suggestion, safety concern and product misunderstanding. The source, prompt, presentation or feature boundary is corrected according to the issue.

• Report with the minimum necessary context.

• No use of sensitive reports as marketing examples.

• Regression test before releasing a correction.

• Serious patterns can lead to withdrawal.