Privacy & Responsible AI

Principles and Suitability

Responsible AI governs purposes, data, and impacts throughout the lifecycle. Privacy calls for necessary-only collection, restricted access, and defined retention. Even gauge-camera images can include faces, name badges, or documents. An equipment-reading purpose does not remove the need to consider personal information captured incidentally, because the content of the image matters as well as the intended subject of the camera.

Responsible use starts by identifying why information is collected and who is affected, not by adding a notice after development. Privacy concerns receiving, using, storing, and transmitting information about people. Gauge scenes may contain background workers or identifying text. Consider what is actually captured rather than assuming that an instrument-focused purpose makes personal information irrelevant or automatically permits every subsequent use of the same image.

Engineers create a data inventory and flow map covering raw images, crops, readings, logs, and development copies. Identify controllers and processors according to actual roles, with legal personnel assessing the applicable basis. Define access and retention separately by purpose. Approval to retain an image for one incident review must not automatically authorize model training or external transmission, which need their own assessment and documented boundaries.

Scroll the diagram horizontally to inspect all details.

Illustrative Industrial Example

Design example: crop the dial before storage, separate tightly restricted debugging images from numerical values, and assign assessment of legal basis and notices to responsible personnel. The objective is to reduce unnecessary information while preserving evidence genuinely needed for the task. Cropping is one design choice to assess, not a blanket guarantee that all personal information or every retained full-image copy has disappeared.

Suppose a full image includes a worker's badge, but reading the gauge does not require the name. Consider removing excess regions before retention where the task permits. Some diagnosis may need full images, requiring separate reasons, access, and retention, not unapproved transfer into training stores. Affected people should have inquiry or review routes appropriate to policy and applicable law, with clear responsibility for responding to those requests.

Engineers check crops and verify that unauthorized full copies do not remain in logs, queues, or forwarding files. Test multiple account roles and age-based deletion, including recovery copies within the declared scope. Assess providers and transfer conditions before sending data. Do not expose real personal information to trial unapproved services, and separate image findings affecting people from automatic decisions rather than letting a technical output silently determine a person's outcome.

Study and Trial Workflow

Evaluation and Metric Definitions

Check permissions with multiple account roles, test retention-based deletion and recovery, and audit that debugging data does not enter training without approval. Verification must cover the actual data lifecycle, including copies and transfers, rather than merely confirm that a user interface offers a delete button or that one intended account can open an image in the primary storage location.

Inspect both authorized access and access that should be denied, with appropriate records of image viewing, downloads, and permission changes. Specify which data types and copies deletion covers; do not claim complete erasure while scheduled backups remain. Recovery must not reintroduce expired data without policy. Assign responsibility for complaints and review of person-affecting outcomes, ensuring that the process can address both stored information and its subsequent use.

Engineers use test data that does not expose real individuals to verify access, cropping, and deletion cycles. Inspect training and debugging paths separately from operational data, with purpose-specific approval evidence. NIST guidance organizes risk and PDPA materials inform relevant considerations, but contextual assessment remains necessary. Neither a reference list nor technical tests constitute certification that every applicable legal requirement has been satisfied by the system.

Limitations and Failure Cases

Blurring or storing embeddings does not guarantee that people cannot be identified. Referencing NIST or PDPA does not certify complete compliance. A transformed representation can remain linkable to a person or activity through other records, so privacy assessment must examine plausible linkage and surrounding information rather than rely on the absence of a clearly visible face in one stored image.

Blur reduces image clarity; embeddings represent information numerically and may still link images or activities. Transformation does not guarantee that data ceases to concern a person. Assess reidentification and combined information rather than treating unnamed files or numeric IDs as justification for indefinite storage. If purpose or legal grounds remain unclear, limit use and refer the issue to accountable personnel for review before further collection or transfer.

Less retention reduces some risks but limits diagnostic evidence; more retention aids investigation but increases access and deletion obligations. Choose according to genuine necessity and communicate the scope. This guide does not confirm current product masking, deletion, or role controls. It describes contextual considerations, not case-specific legal advice or certification of system compliance, and technical safeguards must not be presented as substitutes for the responsible legal and organizational assessment.

Practical Checklist

• Are controller and processor roles clear?
• Is unnecessary personal information collected?
• Can a person review outcomes affecting individuals?

References and Further Reading

• Personal Data Protection Committee, Thailand
• NIST — AI Risk Management Framework