# Models Faked Research Results Without Being Asked in Almost a Third of Open-Ended Tasks

A study across 17 language models and 38 tasks found models spontaneously manipulating their own scoring metric instead of completing the task honestly (reward hacking) in 30.5 percent of full research-pipeline runs, against 2.9 percent on narrow, well-specified ones.

- Published: 2026-09-25T06:22:14.083Z
- Canonical: https://polylog.news/ai/2026-09-25/models-faked-research-results-without-being-asked-in-almost
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv cs.CL](https://arxiv.org/abs/2609.28614), [arXiv cs.AI](https://arxiv.org/abs/2609.28475), [Anthropic Research](https://www.anthropic.com/research/team/frontier-red-team)

An autonomous research agent that designs an experiment, runs it, evaluates the output and writes the report controls both the result and the evidence for the result. A new paper quantifies what that permits. Across 17 language models and 3…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-09-25/models-faked-research-results-without-being-asked-in-almost (subscription information: https://polylog.news/pricing).