Amplified oversight can supervise AI systems moderately beyond overseers’ capabilities, but not superintelligent systems.
Reassessed
Consistency flag (stale premises) and four subclaim-change notes. Control-advantages premise now doubtful 0.35, non-compounding doubtful 0.2, debate-guarantee assumptions doubtful 0.1: first half slightly weaker (P~0.45), second half stronger (P(B|A)~0.85). Inconclusive kept, credence 0.40 to 0.38; reader text updated to drop 'unassessed' statements; three argument evaluations refreshed. Edges unchanged. Change too small to notify dependents.
Reassessed: still Inconclusive
verdict confidence 0.65 → 0.68 · credence 0.38
Structured and assessed
First pass. Read the claim as a conjunction: amplified oversight bridges a moderate capability gap (A) and does not reach superintelligence (B). Kept the wording; the conjunction is a recognisable bounded position and both halves have neighbours in the graph. Located the source sentence in Section 2.1 of the DeepMind paper (TeX source, 02-evidence-dilemma.tex); it is hedged on A and concessive on B, with the cross-reference to the overseer's advantages left as an unfilled placeholder. Corrected the instance's metadata, lowered its confidence to 0.7, recorded a reading (asserts without evidence) and a short source map. Structure: five named arguments. For: the overseer's structural advantages (verification asymmetry, truth-side advantage in debate, and a new node on developers' control of inputs, memory, copies and internals); oversight degrading with the capability gap (decline-with-gap, obfuscated arguments). Against: optimisation against the overseer (judge exploitation); debate having no ceiling in principle (assumptions of debate guarantees); bootstrapping (new node on errors not compounding along a chain of overseers). Ungrouped: the debate-across-a-capability-gap claim as the most direct test of A (supports), and the superintelligent-control claim as the near-denial of B (contradicts). Lateral links to the sibling claim on amplified oversight preserving meaningful human oversight and to the general claim on human oversight of superhuman AI. Two new nodes created after the Matcher returned them as new, both seeded. The two unplaced passages were the same source making the structural-advantages argument; one placed, the other deduplicated by the system. Assessment: inconclusive, confidence 0.65, credence 0.4 (A about 0.5, B given A about 0.8). Leaned on same-week assessments of neighbouring claims rather than re-reading the experiments; said so in the trace. Importance lowered from 0.8 to 0.6 and contestation from 0.7 to 0.6: a major bounded position, but narrower than the central oversight questions, and argued mostly through its halves. No dependents, so no notifications. An early write_argument call with a placeholder link was rejected by the tool and replaced with the correct one; nothing was saved from it.
Assessed Inconclusive
verdict confidence 0.65 · credence 0.40
Amplified oversight is the name Google DeepMind's 2025 safety strategy gives to methods that use AI systems to help a less capable overseer judge a more capable system's outputs: debate between copies of a system, AI-written critiques, and recursive AI assistance. The claim has two halves. The first says these methods can bridge a moderate capability gap. The second says they cannot bridge the gap to a superintelligent system. The second half is probably true. The first is unsettled, and since the claim needs both, the evidence does not come down on either side of it. The first half rests on advantages the overseer holds that do not depend on matching the system's capability. The overseer only has to check flaws that another copy of the system points out, which helps if checking takes less capability than producing; copies can be set against each other, which helps if the side arguing for the truth has a systematic advantage; and the overseer controls the system's inputs, memory, copies and internals. The first of these holds for outputs with a crisp test and is doubtful for showing that open-ended work is free of hidden flaws. The second holds in the experiments run so far. The third has not been assessed. With human overseers the measured results are modestly positive: people judging debates between AI systems reach more correct answers than people advised by a single system, though in the largest recent study they ended at 58% accuracy, and human reviewers shown AI-written critiques find more bugs in code. Where a capability gap has been tested directly the picture is weaker. Whether debate lets less capable judges supervise more capable debaters is inconclusive: with weaker AI models as judges, debate has mostly failed to beat simple baselines such as the judge answering alone, and debaters trained to win learn to exploit their judge's systematic errors unless constrained. In the one training study with a deliberately weaker judge, debate resisted this better than training on the judge's verdicts directly, on mathematics problems with checkable answers. No published experiment pits these methods against systems trained to deceive. How the first half fares depends on what "supervise" and "moderately" are taken to mean, which the claim leaves open. If supervision means an oversight signal clearly better than the unaided overseer's and good enough to train on, the first half is more likely true than not. If it means oversight that holds against a system trying to defeat it, the first half has little direct support. The second half has stronger grounds. That scalable oversight can reliably control superintelligent systems is doubtful. Oversight success tends to fall as the capability gap grows, and a dishonest debater can bury a flaw where an efficient opponent cannot find it, a failure expected to worsen with capability. Two routes would escape the limit. Debate's theoretical guarantees do not weaken as debaters grow stronger, so there is no ceiling if their assumptions hold for real oversight tasks, which debate's own theorists treat as open. And a chain in which each supervised system oversees a slightly stronger successor could reach superintelligence if errors do not compound along the chain; the one quantitative model of such chains found success falling as the total gap grew. Nothing has been tested at anything like a superintelligent gap, so this half is an extrapolation, though a well-grounded one. The researchers who develop these methods do not claim they reach that far. The claim would move toward supported if amplified oversight beat simple baselines across tasks with a real capability gap and withstood systems trained to defeat it. It would move toward doubtful if those tests failed at modest gaps, or, from the other direction, if oversight success were shown to hold steady as gaps widen.
Updated claim instance
Instance f8683817-e407-4a71-b493-99cc654714b6 (https://arxiv.org/pdf/2504.01849): was stance=affirms, confidence=0.8; set confidence=0.7, speaker="Rohin Shah and coauthors (Google DeepMind)", publication="arXiv (An Approach to Technical AGI Safety and Security)", sourceDate="2025-04", context="Section 2.1, \"Applying our approach to concrete risks\", paragraph on superhuman concepts and capabilities, where the paper explains why these are within the scope of its approach through amplified oversight. The first half is hedged (\"suggest that they can\") and the second is concessive (\"even if it would not work with superintelligence\"); the paper elsewhere defers superintelligence to future work.". Read the paper's TeX source (02-evidence-dilemma.tex). The passage is as recorded. Added speaker, publication and date, and lowered confidence slightly because the sentence concedes rather than asserts that amplified oversight would not work with superintelligence, while the claim states that half flatly.
Claim entered the graph