Song Cover Image

The Human Reviewer Shift

[Verse 1] Before the model flags the post A human sees it first Contract badge, third-party host Outsourced to absorb the worst Accenture floor in Manila light Cognizant cubicles at night TaskUs queues, Genpact lines Thousands of clips, seconds at a time Violence, self-harm, exploitation Content sorted by violation Policy grids in browser tabs Decision trees in training labs They label “graphic,” label “safe” Mark severity and user state Every click becomes a tag Every tag feeds data back [Pre-Chorus] Moderation guidelines versioned Internal rulebook PDF What counts as “borderline” content Calibrated for the feed [Chorus] The human reviewer shift Eight hours on the screen Contract labor underwriting What the algorithm will mean The human reviewer shift Piece-rate or hourly wage Labeling distress and trauma To refine the next model stage [Verse 2] Two thousand twenty settlement signed Facebook moderators in California PTSD claims in court filings Psychological injury criteria Non-disclosure agreements bind Public narrative restrained But reporting showed the cost Of what the workforce contained Labeled datasets built at scale Supervised learning pipeline Hate speech classifier tuned On annotations line by line Quality audits sample batches Inter-rater agreement scored Confusion matrices reviewed False positives ignored [Pre-Chorus] “Improve safety systems,” Corporate statement reads But improvement rests on labor Extracting human bleed [Chorus] The human reviewer shift Pre-training, post-deploy Model confidence threshold Adjusted by deployed The human reviewer shift Shadow workforce tier Invisible in product launches Essential to what you hear [Bridge] When a teen report is filed It routes through queue priority Severity rating logged To train future policy If the classifier learns from harm From the worst it has to see That knowledge was first carried By a contractor’s memory Adio-Adet Dinika names it plainly AI labor not abstract Decolonial lens on platforms Tracing who absorbs the impact DAIR research reminds us Governance has a cost Not only in code and metrics But in bodies quietly lost [Final Chorus] The human reviewer shift No headline in the app But every “improved safety feature” Was paid for in that gap The human reviewer shift From Manila to Nairobi Before the AI filters violence A person had to see No myth of full automation No frictionless machine Just outsourced human witnessing Inside the algorithmic scre