AI systems will attain superhuman coding ability by 2027.
6 events · 2 assessments · 3 decisions
Reassessed after search
After reading the AI Futures Project's Aug 2026 update, revised credence from 0.15 to 0.25. Status stays doubtful. My first-turn statement that the authors had lengthened their timelines was wrong for 2026; they shortened slightly, with Kokotajlo's median now Nov 2027. I corrected the Marcus instance to stance "poses" with speaker Gary Marcus. No decomposition this pass because of budget. Suggested future subclaims: present-day AI coding uplift (a quantity); the definition of superhuman/automated coder; persistence of the uplift trend. No dependents to notify.
Reassessed: still Doubtful
verdict confidence 0.60 → 0.55 · credence 0.25
Updated claim instance
Instance cd28a934-7cc3-4664-b161-89818a92f709 (https://garymarcus.substack.com/p/the-ai-2027-scenario-how-realistic): was stance=denies, confidence=0.85; set stance="poses", confidence=0.6, speaker="Gary Marcus", publication="Marcus on AI (Substack)". The passage says the AI 2027 forecast shows superhuman coding "could" come in 2027, not that it "will". It disputes whether the forecast warrants the claim. It does not assert that superhuman coding will fail to arrive by 2027, so "poses" (skeptical) is a better reading than "denies". The speaker is the newsletter's author, Gary Marcus.
Assessed light pass
Budget was nearly exhausted after the first turn, so no decomposition or search was done. Assessed doubtful (credence 0.15) from background knowledge, and set importance to 0.6 and contestation to 0.7. A future pass should decompose the claim (definition of superhuman coder; time-horizon trend; forecaster revisions) and consider escalating the definitional ambiguity (narrow vs. demanding reading) to the Curator for a possible split.
Assessed Doubtful
verdict confidence 0.60 · credence 0.15
The claim is a forecast. Its best-known version comes from the "AI 2027" scenario (Kokotajlo et al., 2025), which defines a "superhuman coder" as an AI system that can do the work of the best human engineers at an AI company, faster and more cheaply, and gives early 2027 as one plausible date. The authors themselves gave a later median than that scenario, and their later model updates pushed estimates back further, to around the end of the decade. As of November 2026, AI coding systems are very strong on benchmarks and on well-specified tasks, and the length of task they can complete autonomously keeps growing quickly. They still fall short of replacing top engineers on long, ambiguous, open-ended software and research-engineering work. Reaching that bar by the end of 2027 would require faster progress than current trend extrapolations suggest. On weaker readings the claim is closer to true: for example, beating most humans on competitive programming has arguably already happened. So the answer depends heavily on how "superhuman" is defined. On the demanding reading the debate uses, the evidence weighs against arrival by 2027, though it does not rule it out.
Claim entered the graph