MMarketing Against The Grain
← All frameworks
Strategy

Pre-Compression AI Media Flagging

Inspect uploads before compression, then label detected AI media for users.

Difficulty
Advanced
Time to result
~months to results
Steps
5
Confidence
90%

This platform-control model places AI-media detection at the only stage where the network reliably possesses the original upload. Instead of trying to analyze an image after it has been compressed for storage and delivery, the platform intercepts the file during upload and runs authenticity checks first. The resulting classification can then travel with the processed asset and appear as a visible label to users. The mechanism addresses both a technical problem—the loss of forensic signals during compression—and a communication problem—the absence of clear provenance cues in feeds. A production implementation would also need confidence thresholds, audit records, appeal handling, and safeguards against presenting uncertain detector output as incontrovertible fact.

Origin

Extracted from Marketing Against The Grain during a discussion of how social networks could flag AI-generated imagery.

Core principles

  • 01Run authenticity checks before lossy processing.
  • 02Preserve detection signals at the point of upload.
  • 03Communicate detection results directly to users.
  • 04Make provenance controls a platform-level responsibility.

How to run it

  1. 1

    Capture the Original Upload

    Route newly uploaded images and videos through a provenance-checking stage before any resizing or compression occurs.

    Pro tip Keep detection logically separate from the later media-optimization pipeline.

    Watch out Running only against the published derivative forfeits valuable evidence.

  2. 2

    Analyze Authenticity

    Apply available generation detectors, metadata checks, and structural tests to the original file. Produce a classification with an explicit confidence level.

    Pro tip Combine multiple independent detection signals.

    Watch out No single detector should be treated as infallible.

  3. 3

    Persist the Result

    Associate the classification and supporting evidence with every derivative produced from the upload.

    Pro tip Use an immutable audit record for consequential labels.

    Watch out Do not discard uncertainty information when simplifying the user-facing result.

  4. 4

    Label the Media

    Show users a clear indicator when the evidence supports an AI-generated classification. Distinguish confirmed provenance from probabilistic detection.

    Pro tip Make the label understandable without requiring technical knowledge.

    Watch out An overly definitive label can amplify detector errors.

  5. 5

    Review Disputes

    Provide a process for creators to challenge an incorrect classification and submit provenance evidence.

    Pro tip Use disputed cases to evaluate detector performance.

    Watch out An opaque process will undermine trust even when most classifications are correct.

In the wild

Social-Network Upload Check

A user uploads an uncompressed synthetic event photograph. The platform analyzes the original file, records several generation signals, and then creates compressed feed versions. Each derivative retains the classification and displays an AI-generated label, even though the later compressed files no longer contain enough detail for reliable independent detection.

Users receive provenance context without depending on post-compression analysis.

Common mistakes

Detecting After Compression

Post-processing may remove the very artifacts needed to make the classification reliable.

Hiding Detector Uncertainty

A binary public label without confidence handling can convert probabilistic errors into reputational harm.

Is it for you?

Best for

It is best for social networks and publishing platforms that control the original media-upload pipeline.

Not ideal for

It is not ideal for third parties that only receive already-compressed copies of published media.

From the transcript

What they would have to do, Kieran is have an application that when you're uploading it, checks it before the file is compressed, checks it.…

11:30

They would have a little flag on them and say, This is AI generated.

11:00

From the episode

Ai Images: Flux, Spotting Fakes & 4 Marketing Use Cases