Leaked documents show that Microsoft pays $12 to $14 per hour to the contractors who transcribe up to 200 audio clips per hour to train Cortana
Joseph Cox / VICE :
Context & Ripple Effects
This lands one week after Joseph Cox revealed that Microsoft contractors were listening to Skype calls run through the translation service, and it supplies the missing economic layer: the humans reviewing those recordings are paid $12–$14 an hour and expected to process up to 200 clips per hour. A week later Cox documented the same pipeline touching accidentally invoked Xbox recordings, many from children, making this a sustained series rather than a one-off leak.
The leak also quantifies what Bloomberg's later cross-industry investigation into Amazon, Apple, Google, and Facebook transcription practices generalized: every major voice assistant runs on low-cost contract labor that companies rarely disclose. The $12–$14 rate against a 200-clips-per-hour quota puts a concrete price tag on the 'training data' behind Cortana, which Microsoft was simultaneously marketing through products like the $350 Surface Headphones.
First-order effects
- Microsoft's Cortana training operation is exposed as dependent on piecework-style human transcription at roughly $12–$14 an hour, undercutting the perception that voice assistants are purely automated software and putting Microsoft's contractor disclosures under immediate scrutiny.
- The contractors themselves gain leverage from the leak: their pay rates and workload quotas are now public, complicating Microsoft's ability to keep terms confidential while the Skype and Xbox stories keep the program in the news.
Second-order effects
- Amazon, Apple, Google, and Facebook — all shown by Bloomberg's investigation to rely on similar contractor transcription — face pressure to disclose their own rates and review practices before leaks do it for them, since Microsoft has now set the reference point for what 'human review' actually costs.
- Privacy expectations harden around voice products: with consumers learning that assistants mean human ears at industrial clip rates, Microsoft and rivals must weigh consent disclosures against the data volume their speech models need.
Third-order effects
- If the pattern holds, AI training labor stays structurally invisible in product economics — a dynamic visible again years later when Microsoft's GitHub Copilot was reported losing $20+ per user per month — meaning assistant businesses carry both cheap human annotation upstream and heavy inference costs downstream.
- Regulatory and consumer scrutiny converges on human-in-the-loop review as a disclosure category of its own, forcing voice-assistant makers to treat contractor programs as a reputational surface equal to the privacy of the recordings themselves.
The trend: Voice assistants across Big Tech are being re-priced as hybrid human-machine systems, with low-paid contract transcription becoming both a cost line and a privacy liability that companies can no longer keep out of view.