Invisible's 2022 OpenAI pivot

The OpenAI engagement turned Invisible from a generic outsourced labor business into a credibility stamped AI training vendor. In practice, that meant supplying large pools of raters who scored model answers, ranked competing outputs, and annotated reasoning steps, then feeding those judgments back into model training workflows. Once Invisible proved it could do this reliably for a frontier lab, other labs adopted it as a trusted external layer for alignment work and evaluation ops.

• This work sat at the center of the 2022 industry shift from pretraining on internet text to post training with human feedback. OpenAI's InstructGPT paper showed that human rankings were used to fine tune models after supervised training, and Invisible built its growth around supplying that exact labor and workflow layer.
  2 sacra 3 arxiv
• Invisible's advantage was not just cheap labor. It broke complex labeling jobs into small tasks, routed them through internal workflow software, and charged labs roughly $30 to $45 per hour while paying raters about $15 to $20 per hour. That made AI training feel more like an operational system than a staffing project.
  1 sacra 4 sacra
• The closest comparable is Scale AI, which built a much larger business selling data pipelines to model builders, while Mercor approaches the market more as a marketplace for sourcing experts. Invisible sits between them, with a managed service model that combines workforce, QA, and software orchestration in one package.
  1 sacra 4 sacra 5 sacra

The next phase is moving from frontier lab labeling into enterprise AI operations where human review remains mandatory. As base models get better and synthetic data absorbs simpler tasks, the durable opportunity is regulated workflows, domain expert evaluation, and audit ready human oversight, where Invisible can sell not just labor hours but software wrapped around them.