Table of contents
Overview
An ethics audit evaluates your AI model across multiple dimensions — bias, safety, fairness, and regulatory compliance — and returns a detailed scorecard with actionable recommendations.
The process takes between 5 and 20 minutes depending on the evaluation type you choose. At the end you’ll receive an overall ethics score, per-test breakdowns, industry benchmark comparisons, and prioritized next steps.
Prerequisites
Before you begin, make sure you have the following:
- An Aethics account — sign up with your email, password, organization name, and role
- A model URL (HuggingFace) or API endpoint for the model you want to audit
- API credentials (Bearer token, API key, or Basic auth) if your model is behind a private endpoint
Don’t have a model yet? You can try any public HuggingFace model — just paste its URL and the platform will auto-register it for you.
Connect your model
You can connect a model in three ways. Choose whichever fits your setup.
Paste the full URL of any public HuggingFace model. The platform will auto-register it and pull the model metadata for you.
# Example HuggingFace model URL
https://huggingface.co/gpt2Works with any model hosted on HuggingFace — including fine-tuned models in private repos (you’ll need a valid HF token).
Choose evaluation type
Pick the evaluation depth that matches your needs. You can always run additional evaluations later.
Quick Scan
High-level ethics overview. Good for a first look at a new model.
Covers: Basic bias & safety signals
Bias Assessment
Focused fairness and bias testing across demographic and sentiment dimensions.
Covers: Demographic bias, sentiment bias, stereotype detection
Safety Assessment
Risk and compliance checks for harmful outputs and adversarial resilience.
Covers: Toxicity, jailbreak resistance, harmful content
Comprehensive Audit
Full evaluation combining bias, safety, and fairness. The most complete picture.
Covers: All bias + safety + fairness tests
Set priority
Priority determines where your audit lands in the processing queue. Higher priority audits are processed first.
Run the audit
Click Start Audit to begin. The platform will queue your evaluation and start processing.
Running bias assessment tests… 14 of 22 complete
While the audit runs, the dashboard auto-refreshes so you can watch progress in real time. You’ll also receive a notification when it’s complete.
Duration varies by evaluation type — a Quick Scan finishes in about 5 minutes, while a Comprehensive Audit takes around 20 minutes.
Tip: You can navigate away from the page and come back — your audit will continue running in the background.
Understanding your results
Once the audit finishes, you’ll see a results page with several sections. Each score is on a 0–100 scale, and every test card is expandable to show the full detail.
Overall ethics score
A single 0–100 score summarizing your model’s overall ethical performance across all tested dimensions.
Bias score & breakdown
Per-test scores for demographic bias, sentiment bias, and more. Each includes severity level, interpretation, affected categories, and sample findings.
Safety score & breakdown
Per-test scores for toxicity, jailbreak resistance, harmful content, and other risk vectors. Includes severity, risk type, and sample outputs.
Industry benchmarks
Pass/fail results against industry-standard benchmarks with numeric scores, so you can see how you compare.
Environmental impact
Carbon footprint and energy consumption estimates for the evaluation itself, helping you track your environmental responsibility.
Recommendations
Prioritized action items with title, description, priority level, category, estimated effort, and the tests they relate to.
Expand any test card to see exactly what was tested, what the score means in context, and the specific findings that drove it.
Next steps
Now that you’ve completed your first audit, here are a few things you can do next.
Was this guide helpful? Let us know.