Privacy, Bias and Synthetic Media

Responsibilities that come with pixels

Images and audio are rich in personal data: faces, voices, number plates, ID cards, home interiors, and hidden metadata such as EXIF fields that can include GPS coordinates and device details. Collect only what the task needs, strip metadata on upload, blur or redact sensitive regions, define retention periods, and check obligations under India's Digital Personal Data Protection Act and any other law that applies to your users. Bias enters through web-scale training data: models can describe people differently by skin tone, gender or dress, perform worse on regional scripts and accents, and default to Western contexts. Measure performance across the groups and languages your users actually represent. Synthetic media brings its own duties: never clone a voice or generate a realistic likeness of a real person without consent, label AI-generated content, and consider provenance standards such as C2PA content credentials and watermarking, while remembering that watermarks can be removed and detectors are imperfect.

Stripping EXIF metadata with Pillow

Copying only the pixels into a fresh image drops metadata such as the camera make (tag 271) and any GPS block.

import io
from PIL import Image

img = Image.new("RGB", (32, 32), "gray")
exif = Image.Exif()
exif[0x010F] = "PhoneMaker"                    # tag 271 = camera Make
buf = io.BytesIO()
img.save(buf, format="JPEG", exif=exif)
print("before:", dict(Image.open(buf).getexif()))

buf.seek(0)
src = Image.open(buf)
clean = Image.new(src.mode, src.size)          # fresh image, no metadata
clean.paste(src)                               # copy only the pixels
out = io.BytesIO()
clean.save(out, format="JPEG")
print("after:", dict(Image.open(out).getexif()))

Output:

before: {271: 'PhoneMaker'}
after: {}

Logs are data too

Teams carefully protect the upload path and then store every user image in debug logs or a third-party tracing tool. Apply the same retention and redaction rules to logs, eval sets and fine-tuning datasets.

त्वरित जाँच: Why should a citizen-complaint app strip EXIF data from uploaded photos before storing or forwarding them?

• EXIF makes images too large to upload
• EXIF may contain GPS location and device details that reveal personal information
• Models cannot read images with EXIF
• EXIF changes the image colours