Software That Audits Machine-Learning Models: A Sourced Comparison
Ten tools that check machine-learning models — fairness, drift, data integrity, leakage, explainability and audit documents — compared from their own documentation, with what each one does not check.
In this article
Software that audits machine-learning models checks a trained model, or the data behind it, for problems an auditor cares about: unfair treatment of groups, data quality and drift, information leaking from the outcome into the inputs, unexplained predictions, and missing documentation. No single tool covers all of these. Most organisations combine a few.
This comparison covers ten tools for predictive (tabular) models, using only what each tool's own documentation says, checked on 19 October 2026. It does not cover LLM safety scanners — garak, for example, describes itself as an "LLM vulnerability scanner", a different job.
What "auditing a model" covers
| Audit question | What answers it |
|---|---|
| Does the model treat groups fairly? | Fairness metrics and mitigation (AIF360, Fairlearn, cloud bias tools) |
| Is the data sound, and has it drifted? | Data-integrity and drift checks (Evidently, Deepchecks) |
| Did information from the outcome leak into the inputs? | Leakage checks (Deepchecks, Giskard v2, DataRobot, YourCloudGroup Model Manager) |
| What drives each prediction? | Explainability (SHAP, Vertex Explainable AI, Azure interpretability) |
| Was the score earned honestly, and can it be reproduced? | Held-out evaluation, repeatability and lineage records |
| Is there a document an auditor can file? | Audit or compliance reports (Azure Responsible AI scorecard, DataRobot compliance documentation) |
For the evidence an auditor needs in the first place, see what makes a predictive model auditable.
The comparison
| Tool | What it checks | Checks target leakage? | Audit document | Licence or cost | Library or product |
|---|---|---|---|---|---|
| IBM AI Fairness 360 (AIF360) | Fairness and bias metrics; bias mitigation | No | No | Open source (Apache-2.0) | Python/R library |
| Microsoft Fairlearn | Group fairness assessment and mitigation | No | No standalone report (used inside Azure's dashboard) | Open source (MIT) | Python library |
| Evidently | Data drift, data quality, classification and regression quality (plus text/LLM evaluations) | No leakage preset documented | HTML/JSON reports | Open source (Apache-2.0); optional platform | Python library |
| Deepchecks (open source) | Data integrity, train–test validation and drift, model performance | Yes — date train-test leakage and possible label leakage | Suite results | Open source (AGPL-3.0) | Python library |
| Giskard | Version 2 scan: performance bias, robustness, confidence, ethical bias, data leakage, spurious correlation | Yes, in version 2 only — version 3 targets agentic and LLM systems | Version 2 scan report | Open source (Apache-2.0) | Python library |
| SHAP | Explainability only (Shapley values) | No | No | Open source (MIT) | Python library |
| Amazon SageMaker Clarify | Pre- and post-training bias, SHAP explanations, drift via Model Monitor | No | Analysis file and report notebook | AWS usage-based — closed to new customers, no new features planned | Managed service |
| Google Vertex AI evaluation and Explainable AI | Evaluation metrics with slices, feature attributions, attribution drift; bias metrics in preview | No | Metrics in console/API; no compliance document | Google Cloud usage-based | Managed service |
| Azure Machine Learning Responsible AI dashboard | Error analysis, fairness, interpretability, counterfactuals, causal analysis, data analysis | No | Yes — a PDF scorecard Microsoft says you can attach to compliance reports | Azure compute costs | Managed service with SDK/CLI (tabular, scikit-learn models) |
| DataRobot compliance documentation | Model documentation and validation; the platform screens for target leakage during data analysis | Yes (platform) | Yes — editable DOCX compliance document, custom templates | Commercial; price not published | Commercial platform |
Notes that change the picture
- SageMaker Clarify is closed to new customers. AWS says it plans no new features and points to replacements such as SHAP and standard Python libraries.
- Giskard's automatic tabular scan is a version 2 feature. Its version 3 README says the version 2 tabular scan, test suite and Hub are not planned for version 3.
- Deepchecks' open-source library has a slow release cadence — its last PyPI release was December 2024 — so check that it suits your Python environment.
- Leakage is the gap. Only three of the ten tools here — Deepchecks, Giskard version 2 and DataRobot — document a check for whether information from the outcome reached the inputs, which is the failure that makes a model look far better than it is: in a measured example, one leaked column lifted a model's score from 0.81 to 0.96.
How to choose
- You must show fair treatment of groups (lending, hiring, insurance): add a fairness toolkit — Fairlearn or AIF360 — whatever else you use.
- The model is in production and data changes: add drift monitoring — Evidently or Deepchecks.
- Someone will ask "why this prediction?": add explainability — SHAP, or your cloud's built-in tool.
- An auditor or regulator needs a document: the Azure Responsible AI scorecard (Azure models) or DataRobot compliance documentation (DataRobot models).
- You need to prove the score itself was honest — no leakage, a sealed test set, repeatable results: that has to be designed into how the model is built and recorded, not added afterwards.
Where YourCloudGroup Model Manager fits
YourCloudGroup Model Manager builds predictive models and records the audit evidence as it goes, rather than auditing a model afterwards. Each model receives a certificate of fourteen independent claims covering identity, lineage, training-data boundary, leakage screening, sealed blind evaluation, repeatability and deployment readiness, each marked CERTIFIED, NOT ESTABLISHED, NOT ASSESSED, NOT APPLICABLE or FAILED.
It is not a fairness toolkit: its certificate does not include group-fairness metrics, and it does not replace drift monitoring after deployment. For decisions about people, pair it with Fairlearn or AIF360.
Limitations of this comparison
- Facts come from each tool's own documentation on 19 October 2026; tools change, so follow the sources.
- "No" in the leakage column means the documentation did not describe a leakage check, not that the tool cannot be used to look for leakage by hand.
- We have not run the same model through every tool, so this article makes no claim about which finds more problems.
- YourCloudGroup publishes this blog and sells Model Manager.
Sources
- AIF360: GitHub. Fairlearn: fairlearn.org, GitHub.
- Evidently: GitHub, presets.
- Deepchecks: GitHub, train–test validation.
- Giskard: GitHub (v3), v2 scan reference.
- SHAP: GitHub. garak: GitHub.
- Amazon SageMaker Clarify: availability change, analysis results.
- Google Vertex AI: Explainable AI, model evaluation.
- Azure Machine Learning: Responsible AI dashboard.
- DataRobot: compliance documentation, data quality and leakage.
Build predictive models from your own data
YourCloudGroup Model Manager builds, evaluates, certifies and schedules predictive models from your CSV or spreadsheet data — point and click, on your own dedicated server.
Learn about Model Manager →Published by YourCloudGroup. Read our editorial standards and corrections policy.
