Sending sensitive data to a third-party model API can create compliance risk even when the provider has data-handling policies. Privacy requirements therefore affect both where a model is trained and where it runs. The platform decision must also account for cost, latency, and the effort of operating the system. section 7.3 section 6.11.4 section 1.3
Choose the simplest training and deployment arrangement that meets the applicable privacy requirements:
- Self-host training when data must remain entirely within your infrastructure. This provides control but leaves the team responsible for library upgrades, GPU driver compatibility, and capacity planning. section 6.11.4
- Consider managed platforms when tenant isolation and contractual guarantees meet the organization’s obligations. Azure and GCP offer such guarantees; equivalent guarantees for OpenAI’s direct API require a separate enterprise agreement. section 6.11.4
- Review data before fine-tuning. Remove or anonymize personally identifiable information unless its use is explicitly justified and legally permitted; check ownership, consent, and residency constraints. section 5.11
Distillation with controlled data exposure
A distilled student can run within the organization’s infrastructure instead of sending each production query to a frontier API. The teacher API is still used to generate training data, so the team must control which prompts it sends. This reduces exposure during student use without making the data-generation phase entirely local. section 7.3
Table 5.9 Responsible AI risks and controls in fine-tuning
|
Area
|
Key Risk
|
What to Watch For
|
Typical Mitigation
|
|
Privacy and Compliance
|
Data leakage, regulatory
violations
|
PII in training data,
unclear data ownership
|
Anonymization, consent
checks, and data residency controls
|
|
Fairness and Bias
|
Discriminatory or uneven
outputs
|
Skewed examples,
demographic imbalance
|
Dataset rebalancing, bias
tests, targeted evaluations
|
|
Safety
|
Over-compliance, harmful
responses
|
Loss of refusal behavior
after fine-tuning
|
Safety prompt sets,
refusal examples, and comparative testing
|
|
Evaluation
|
False confidence in model
quality
|
Overreliance on a single
metric
|
Mixed metrics, human
review, segmented analysis
|
|
Transparency
|
Misuse or
misinterpretation
|
Unclear limitations or
intended use
|
Model cards,
documentation, usage boundaries
|
|
Monitoring
|
Degradation over time
|
New failure modes, user
complaints
|
Logging, feedback loops,
iterative LoRA updates
|
Use the privacy and compliance row to check for data leakage, unclear ownership, anonymization, consent, and residency controls alongside other fine-tuning risks. section 5.11

























