
How to read this finding
The researcher estimate is not directly comparable with the narrower 2021 alignment count. Fine-tuning compute excludes the cost of pretraining, and improved instruction following does not establish comprehensive alignment.
It estimated around 300 full-time researchers, while noting that the field remained small relative to research on capabilities.

The researcher estimate is not directly comparable with the narrower 2021 alignment count. Fine-tuning compute excludes the cost of pretraining, and improved instruction following does not establish comprehensive alignment.
Evidence you can use
Historical snapshot: October 2022. Dates and populations are specified per row.
| Measure | Reported value | Definition and source |
|---|---|---|
| Full-time AI safety researchers | Approximately 300 | Estimate reported in 2022; scope differs from the 2021 selected-organization count.2022 report, slide 98 (PDF page 98) |
| InstructGPT fine-tuning compute | Less than 2% of GPT-3 pretraining | Relative compute figure reported for the fine-tuning stage.2022 report, slide 100 (PDF page 100) |
| Human feedback effort | 20,000 hours | Effort reported alongside the InstructGPT discussion.2022 report, slide 100 (PDF page 100) |
The researcher estimate is not directly comparable with the narrower 2021 alignment count. Fine-tuning compute excludes the cost of pretraining, and improved instruction following does not establish comprehensive alignment.
Historical snapshot published October 11, 2022. This web edition was prepared on 2026-10-11 from the online deck and original launch posts. Findings and forecasts retain their original time frame.
Benaich, Nathan, and Ian Hogarth. “How large was AI safety research in the 2022 report?” State of AI Report 2022. Historical report snapshot; web edition prepared 2026-10-11.