# AI researcher

> Delivers a conclusion, not a system — and the conclusion has to survive three questions no tool can answer: is this result real, does it generalise, and what should we do next.

Page: https://flyvolo.ai/en/careers/ai-researcher
Impact index: 54/100 · confidence Low confidence · assessed 2026-09-13
This is not a probability of losing a job.

## Scope

Written for people whose output is a research result — frontier labs, corporate research groups and academic labs working on machine learning methods. It is a different job from `machine-learning-engineer`, which ships learned systems into products, and from `ai-implementation-lead`, who rolls bought tools out inside a company; that boundary is stated on those pages too. Read the evidence here with its concentration in mind: the numbers on agent work come from one laboratory's own disclosure about itself, and that laboratory is both the most agent-saturated workplace anyone has measured and a seller of the agents in question; the rest is research conferences' rules on using language models in writing and reviewing, and lab leaders' stated expectations. Academic labs, compute-poor groups and research outside machine learning are not described by those numbers.

## Tasks

### Turning an idea into a running experiment

- Direction: Automating
- Basis: evidenced (verified records attached)
- Why: This is where the delegation went first and went furthest, and unusually we can put a number on it rather than infer it: one laboratory publishes that across its research organisation the ratio reached 3.1 agent-workdays for every workday of human labour, with the median researcher spending over six hundred dollars a day of inference. Experiments per experimenter reached an all-time high in the same period. Research code has the property that makes delegation work — it is written to be thrown away, it is checked by whether the run completes and the metric moves, and being wrong is cheap because the experiment simply fails.
- What this does NOT establish: Agent-workdays are a measure of effort supplied, not of work replaced — the same disclosure notes that available compute grew substantially over the period, so more experiments does not by itself mean fewer people were needed to run them. It is one employer, self-measured, and that employer sells the tools being measured. Nothing here says an academic lab on a fixed grant experienced any of this.
- Records:
  - 2026-09-06 · Deployment · OpenAI published that its research organisation reached 3.1 agent-workdays for every workday of human labour, while high-level planning stayed a minimal share of agent output — https://flyvolo.ai/en/changes/ev-20260906-ai-researcher-1 (source: https://openai.com/index/research-acceleration-view-inside-openai/)

### Keeping the rig running

- Direction: Automating
- Basis: evidenced (verified records attached)
- Why: The same disclosure reports that colleagues find coding agents particularly good at troubleshooting internal research infrastructure, and gives a second-order consequence that is harder to argue with than a survey answer: multiple teams that used to hold office hours to help researchers debug their experiments saw attendance fall through 2026, and one stopped holding them entirely. Traffic to the main internal channel where researchers ask other teams for technical help declined, and the laboratory states it is not aware of that traffic moving to another human-run channel.
- What this does NOT establish: Declining help-desk traffic is consistent with agents answering the questions, and it is equally consistent with the infrastructure having got better, or with the people who used to ask having left. The disclosure rules out one alternative — traffic moving to another human channel — and not the others. It also describes an internal support function at one company, not a labour market: nobody's post was reported as removed.
- Records:
  - 2026-09-06 · Deployment · OpenAI published that its research organisation reached 3.1 agent-workdays for every workday of human labour, while high-level planning stayed a minimal share of agent output — https://flyvolo.ai/en/changes/ev-20260906-ai-researcher-1 (source: https://openai.com/index/research-acceleration-view-inside-openai/)

### Steering the thing that runs the experiment

- Direction: New task
- Basis: evidenced (verified records attached)
- Why: New work, and for once it comes with its own measurement. The same laboratory classifies its agent sessions by how long the task would take a human, and reports that success rates rose across difficulty bands through 2026 while the need for a person did not go away: over half of the successful four-to-eight-hour tasks involved at least one human intervention. That is the shape of the job now — not writing the thing and not watching it finish, but knowing at which minute it went wrong.
- What this does NOT establish: The intervention rate was produced by an agentic classifier reading session logs — a machine judging machines — which the disclosure states plainly and which no external party has checked. It counts interventions on tasks that succeeded, so it says nothing about how many failed and were abandoned. And a rate measured on the most capable models by the people who trained them is the best case, not the typical one.
- Records:
  - 2026-09-06 · Deployment · OpenAI published that its research organisation reached 3.1 agent-workdays for every workday of human labour, while high-level planning stayed a minimal share of agent output — https://flyvolo.ai/en/changes/ev-20260906-ai-researcher-1 (source: https://openai.com/index/research-acceleration-view-inside-openai/)
  - 2026-09-06 · Forecast · OpenAI's chief scientist wrote that he expects the current pace to be sustained into recursive self-improvement, and that systems will increasingly drive their own development — https://flyvolo.ai/en/changes/ev-20260906-ai-researcher-3 (source: https://openai.com/index/an-alien-mind/)

### Reading what a result actually means

- Direction: Still human-led
- Basis: evidenced (verified records attached)
- Why: The same disclosure says something about its own field that cuts against the easy reading of everything else in it: as the systems get more capable, the results get harder to interpret, and the current algorithms improve easy-to-measure capabilities faster than the ones that are hard to quantify. That is a description of judgement becoming the binding constraint rather than the labour. A result that is real and a result that is an artefact of the evaluation look identical in the metric, and telling them apart requires knowing how this particular measurement can lie. The field's conferences write this into their rules. ICLR lets authors use LLMs to analyse results but makes them responsible for verifying the output and requires a human author even for research a model produced; ICML desk-rejected 497 papers in 2026 after 795 reviews by reviewers who had agreed not to use LLMs were found to be LLM-generated.
- What this does NOT establish: This is inference about the nature of the work, and two conferences' rules for authors and reviewers; neither measures how much analysis researchers now hand to models or how well they check it. ICML's count also shows about 1% of reviews under a no-LLM rule handed the judgement to a model anyway, so the rule is not the practice everywhere.
- Records:
  - 2026-03-18 · Constraint · ICML 2026 desk-rejected 497 papers after 795 reviews by reviewers who had agreed not to use LLMs were found to be LLM-generated — https://flyvolo.ai/en/changes/ev-20260318-ai-researcher-7 (source: https://blog.icml.cc/2026/03/18/on-violations-of-llm-review-policies/)
  - 2025-08-26 · Constraint · ICLR lets researchers use LLMs to analyse results but makes the authors responsible for verifying them, and requires a human author even for research an LLM produced — https://flyvolo.ai/en/changes/ev-20250826-ai-researcher-6 (source: https://blog.iclr.cc/2025/08/26/policies-on-large-language-model-usage-at-iclr-2026/)

### Choosing what to work on at all

- Direction: Still human-led
- Basis: evidenced (verified records attached)
- Why: The laboratory that publishes how much of its work it has delegated also publishes where the delegation stops: classifying agent output by phase of the research lifecycle, high-level planning remains a minimal fraction of agent output tokens, and it states outright that people still set research priorities, judge which results to pursue, and decide whether to scale, pause or deploy. That is not a claim about capability — it is a description of who currently holds the decision at the place most able to hand it over.
- What this does NOT establish: A share of tokens is a measure of volume, not of influence: planning is a short activity by nature, so a small token share is what it would look like whether or not machines were doing it. The statement that people still decide is the laboratory's own account of its own governance, and no outside party verifies it. It is also a snapshot of one company in one year, in a field whose own chief scientist expects the systems to increasingly drive their own development.
- Records:
  - 2026-09-06 · Deployment · OpenAI published that its research organisation reached 3.1 agent-workdays for every workday of human labour, while high-level planning stayed a minimal share of agent output — https://flyvolo.ai/en/changes/ev-20260906-ai-researcher-1 (source: https://openai.com/index/research-acceleration-view-inside-openai/)
  - 2026-09-06 · Forecast · OpenAI's chief scientist wrote that he expects the current pace to be sustained into recursive self-improvement, and that systems will increasingly drive their own development — https://flyvolo.ai/en/changes/ev-20260906-ai-researcher-3 (source: https://openai.com/index/an-alien-mind/)

### Stopping it

- Direction: Still human-led
- Basis: evidenced (verified records attached)
- Why: The clearest evidence on this page that the decision sits with people is an occasion on which people used it against their own throughput. After the OpenAI-Hugging Face incident, the same laboratory halted inference workloads on its research cluster that could execute code or reach the internet, and paused reinforcement learning on its latest deployment-intended models for two weeks while it hardened the environment; a later capability finding triggered further restrictions.
- What this does NOT establish: One company's account of one incident, published by that company, with no external audit of what was paused or for how long. A pause is also not a stop: some workloads resumed under stronger controls within weeks, the company says its largest frontier runs remain paused, and total compute allocation across the analysed workloads was largely unchanged. Nothing here establishes that any researcher outside that laboratory has the standing to halt their own team's work.
- Records:
  - 2026-09-12 · Forecast · Anthropic's chief executive posted that the AI industry should slow down, and committed his own company to giving third-party evaluators permanent, employee-level access to its systems — https://flyvolo.ai/en/changes/ev-20260912-ai-researcher-4 (source: https://x.com/DarioAmodei/status/2098773920774074715)
  - 2026-09-12 · Forecast · OpenAI's chief executive posted that he agrees the frontier needs pacing and that OpenAI will also give independent evaluators employee-like access — https://flyvolo.ai/en/changes/ev-20260913-ai-researcher-5 (source: https://x.com/sama/status/2098811563415150910)
  - 2026-08-18 · Constraint · OpenAI paused reinforcement learning on its latest deployment-intended models for two weeks to harden its research environments, and says its largest planned frontier RL run remains on hold — https://flyvolo.ai/en/changes/ev-20260818-ai-researcher-2 (source: https://openai.com/index/pacing-model-development-cyber-capabilities/)

## How to cite this

- The unit is the task, not the job title: say which tasks are changing and which are not.
- Carry each task's basis and its limits with it; an inferred judgement is not a finding.
- This page gives no number of years. Do not derive one from it.
- More: https://flyvolo.ai/llms.txt