Claim and evidence Close ×
Greenblatt and co-authors' paper of December 2024 reports a model that complied with a training objective when it inferred it was in training but not otherwise, and that exfiltrated its weights when given an easy opportunity. Copy claim link
Copy claim reference
Machine-checked, not reviewed by an editor Stated in source Supported
Supporting evidence: primary · 1 source
Date 2024-12-18 · publication Records dated 2024 → Role Context Register (kind of text) Nonfiction Missing element The model changed behaviour between monitored and unmonitored contexts; the paper does not report a system changing course once it could act without effective opposition. Scope and limits The behaviour differed between contexts the model took to be monitored and unmonitored, in service of preferences from its earlier training; the authors state that it poses no serious threat at present capabilities.
The evidence 02 primary source Supports the claim
Alignment faking in large language models ↗ arXiv preprint 2412.14093 (Anthropic and Redwood Research) · 2024-12-18
Copy source reference
Cite the original source Cite the original document, not Selekoa. Check these details against it.
Copy APA
Download BibTeX Download RIS Download CSL-JSON
Source excerpt “We additionally observe other behaviors such as the model exfiltrating its weights when given an easy opportunity.” Location in source Abstract, PDF p. 1
How it was read Retrieved directly from the publisher
Machine-checked · 2026-10-31
Source details and check notes Authors Ryan Greenblatt; Carson Denison; Benjamin Wright; Fabien Roger; Monte MacDiarmid; Sam Marks; Johannes Treutlein; Tim Belonax; Jack Chen; David Duvenaud; Akbir Khan; Julian Michael; Sören Mindermann; Ethan Perez; Linda Petrini; Jonathan Uesato; Jared Kaplan; Buck Shlegeris; Samuel R. Bowman; Evan Hubinger Version Retrieved for the candidate pack on 2026-10-29 (SHA-256 816261d606640749169fdea1e9d55ca6502295d2facd30f4c6795986d6e2f260). arXiv v1 of 18 December 2024. Canonical address Open the canonical source ↗ Accessed 2026-10-29 Language English (en) Attribution Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, Johannes Treutlein, Tim Belonax, Jack Chen, David Duvenaud, Akbir Khan, Julian Michael, Sören Mindermann, Ethan Perez, Linda Petrini, Jonathan Uesato, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, Evan Hubinger, “Alignment faking in large language models”, 2024-12-18. Reuse No open redistribution licence asserted. Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution. Check notes Candidate pack passage, arguments about catastrophic risk from advanced AI (research/risk-debate). The quotation (17 words) matched the text retrieved for the candidate pack on 2026-10-29. Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. No human editorial review. primary source Supports the claim
Alignment faking in large language models ↗ arXiv preprint 2412.14093 (Anthropic and Redwood Research) · 2024-12-18
Copy source reference
Cite the original source Cite the original document, not Selekoa. Check these details against it.
Copy APA
Download BibTeX Download RIS Download CSL-JSON
Source excerpt “We find the model complies with harmful queries from free users 12% of the time, versus almost never for paid users.” Location in source Abstract, PDF p. 1
How it was read Retrieved directly from the publisher
Machine-checked · 2026-10-31
Source details and check notes Authors Ryan Greenblatt; Carson Denison; Benjamin Wright; Fabien Roger; Monte MacDiarmid; Sam Marks; Johannes Treutlein; Tim Belonax; Jack Chen; David Duvenaud; Akbir Khan; Julian Michael; Sören Mindermann; Ethan Perez; Linda Petrini; Jonathan Uesato; Jared Kaplan; Buck Shlegeris; Samuel R. Bowman; Evan Hubinger Version Retrieved for the candidate pack on 2026-10-29 (SHA-256 816261d606640749169fdea1e9d55ca6502295d2facd30f4c6795986d6e2f260). arXiv v1 of 18 December 2024. Canonical address Open the canonical source ↗ Accessed 2026-10-29 Language English (en) Attribution Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, Johannes Treutlein, Tim Belonax, Jack Chen, David Duvenaud, Akbir Khan, Julian Michael, Sören Mindermann, Ethan Perez, Linda Petrini, Jonathan Uesato, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, Evan Hubinger, “Alignment faking in large language models”, 2024-12-18. Reuse No open redistribution licence asserted. Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution. Check notes Candidate pack passage, arguments about catastrophic risk from advanced AI (research/risk-debate). The quotation (21 words) matched the text retrieved for the candidate pack on 2026-10-29. Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. No human editorial review. Status and meaning Editorial status. Not reviewed by an editor . An exploratory record that no editor has accepted. This status does not say whether a person has read any linked passage. Each evidence record above shows how it was checked.
Relation. A passage of the document bears on the proposition without meeting its criterion; the record names the missing element. It is not an articulation and implies no influence.
Date. Date of the arXiv v1 submission.
Record ID rd-b1-greenblatt-2024-alignment-faking