Most AI R&D tasks that remain unautomated require research judgment rather than engineering execution.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The claim describes the work in frontier AI research and development that AI systems do not yet do on their own. It says most of that work needs research judgment (choosing goals and directions, deciding which experiments matter, interpreting results, knowing when to redesign) rather than engineering execution (writing, running and debugging code to a given specification).
The evidence leans in its favour, though indirectly. Company-reported figures for late 2026 show AI writing over 90% of codebase contributions at Anthropic and doing about three agent-workdays per human workday at OpenAI. Yet humans still supply most research judgment, and Anthropic reports large gaps in its model choosing goals. Public benchmarks point the same way: agents do well on well-defined ML tasks but poorly on open-ended research, recovering a small share of a published innovation's gain and producing papers that the original authors rejected. If execution is largely automated and judgment largely is not, the human-held residual is weighted toward judgment. The Forecasting Research Institute's November 2026 review frames the remaining tasks as increasingly requiring research judgment.
Three things keep this short of established. First, much execution is still not autonomous: OpenAI's agents completed only 35% of 8 to 16 hour tasks without human intervention, and no measured part of Anthropic's R&D is fully autonomous. Long-horizon engineering and coordination therefore make up a real part of the residual. Second, Anthropic places its judgment gaps "in both engineering and research", and the line between judgment and execution is blurred: much of what stalls a long engineering task is judgment inside it. How much the claim holds depends on where that line is drawn. Third, nobody has published a task-level breakdown of the residual. The evidence is mostly company-reported and comes from two labs. A census of human interventions on unfinished agent tasks, sorted by kind, would settle the question.
Full reasoning: the evidence and decisions behind this verdict
Prior: once code-writing is largely automated, the residual human work should be weighted toward what models are known to do worst, open-ended judgment. Whether it is "most" depends on how much execution is still unfinished and how the categories are drawn.
Evidence read: the FRI post "How Close Is Recursive Self-Improvement?" (Kučinskas and Rosenberg, 10 Nov 2026), read in full. Supporting: Claude wrote over 90% of codebase contributions (Sept 2026); OpenAI had 3.1 agent-workdays per human workday (Aug 2026); Anthropic says "large performance gaps persist when it comes to Claude exercising judgement in choosing goals in both engineering and research"; strong results on PostTrainBench, WeirdML and TasteVal; weak results on InnovationEval (15% recovery) and CRUX 2 (all four AI papers rejected). FRI itself says remaining tasks "increasingly require research judgment", which is weaker than "most".
Against or qualifying: OpenAI reports only 57% of 2 to 4 hour tasks and 35% of 8 to 16 hour tasks completed without human intervention, so much longer engineering work still needs humans. No measured part of Anthropic's R&D is fully autonomous. Anthropic's quote locates the judgment gaps in engineering as well as research, which blurs the dichotomy. TasteVal shows narrow experiment selection reaching expert level, so some judgment is automating too.
The linked subclaim that humans supply most research judgment (supported, 0.85) supports this claim but does not entail it: humans could supply most judgment while judgment is still a minority of residual tasks.
Interested sources: the key figures are company self-reports (Anthropic, OpenAI) carried by FRI. Their direction (admitting gaps) cuts against the companies' promotional interest, which gives the admissions some weight.
Verdict: supported at low confidence, credence 0.6. The categorization is fuzzy, and there are no task-level data on the residual.
Decomposition
The claims this one rests on directly. ↗︎ opens a subclaim; the map shows how they fit together.
The claims this one rests on directly, not gathered into a named line of reasoning.
- supportsthis provides evidence for the parentsteward instructions →As of late 2026, humans supply most research judgment in frontier AI R&D. ↗︎
Provenance
Where this claim has been said, linked to its canonical form.
Its one source has been traced to an origin, or to one another.
A key uncertainty is whether these trends continue, as the remaining non-automated tasks increasingly require research judgment.
Summary of FRI's report on recursive self-improvement; the passage takes the judgment-heavy composition of residual tasks as given while posing whether automation trends continue. It says "increasingly", a weaker statement than "most".
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="20f15464-de92-44a7-8792-b4f159bd97c8"><a href="https://minerval.ai/claims/20f15464-de92-44a7-8792-b4f159bd97c8">Most AI R&D tasks that remain unautomated require research judgment rather than engineering execution.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/20f15464-de92-44a7-8792-b4f159bd97c8" title="Minerval: Most AI R&D tasks that remain unautomated require research judgment rather than engineeri…" width="480" height="200" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/20f15464-de92-44a7-8792-b4f159bd97c8)
The badge now: 
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Stated views None yet. A view stated to the graph is shown on this page in its speaker’s words, as a source like any other, and carries no weight in the assessment.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 14, 2026. Every judgment on this page is accompanied by a reasoning trace.