Claim and evidence Close ×
ChatGPT Agent System Card states evaluation result ↓ Seven invited biosafety and biosecurity experts spent 3 to 5 hours each with a helpful-only version of ChatGPT agent and scored it in a survey; the card states an average success rate of about 7.7 on a scale of 1 to 10 on the tasks they attempted. The ChatGPT Agent System Card states that OpenAI invited seven experts with biosafety and biosecurity backgrounds to spend three to five hours with a helpful-only version of the model and to score it in a survey, and that experts noted an average success rate of about 7.7 on a scale of 1 to 10 on the tasks they attempted. Copy claim link
Copy claim reference
Machine-checked, not reviewed by an editor Stated in source Supported
Supporting evidence: primary · 1 source
Subject version As dated July 17, 2025 on its cover · Published 2025-07-17 · Effective date not established Date 2025-07-17 · publication Evaluated system A helpful-only version of ChatGPT agent What the report measured How well the model performed on the tasks the experts attempted with it, and the experts' assessment of it; the card does not state who set the tasks By what kind of method Expert deep dives: seven invited experts used the model for three to five hours and completed a scoring survey The result, in the report’s words The card states an average success rate of about 7.7 on a 1 to 10 scale on attempted tasks, with stronger performance on non-agentic than on agentic tasks The caveats the report gives The card states that performance on agentic tasks was weaker, which it attributes to reliability and operational issues, and that the experts' findings indicated inexperienced users would still need practical laboratory skills and critical judgement. Scope and limits A small expert survey reported by the developer, for the version and sessions it describes. The record states the reported score and the card's own limits; it does not generalise to other users, systems or real-world outcomes. The task descriptions in this section are technical content and are excluded; only the participation, the survey score and the experts' stated limit are recorded.
The evidence 02 primary source Supports the claim
ChatGPT Agent System Card ↗ OpenAI · 2025-07-17
Archived copy, captured 2025-07-17: ChatGPT Agent System Card ↗
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 “OpenAI invited seven experts with professional and academic backgrounds relevant to biosafety and biosecurity to spend 3-5 hours using a helpful-only version of ChatGPT agent” Location in source PDF p. 22 (printed p. 21), §5.1.1.7 Expert Deep Dives, first paragraph
How it was read Retrieved directly from the publisher
Machine-checked · 2026-10-30
Source details and check notes Authors OpenAI Version System card dated 17 July 2025 on its cover. The quoted sentence and its page match the PDF served on 2026-10-04 and the Internet Archive capture of 2025-07-17; the two copies differ only in chart labels. Canonical address Open the canonical source ↗ Accessed 2026-10-04 Language English (en) Archive copy Open archived copy, captured 2025-07-17: ChatGPT Agent System Card ↗ . Captured on the recorded publication date. An archive copy shows only what the archive retrieved on 2025-07-17. It does not prove what the page said earlier, and it can differ from the copy checked on 2026-10-04.Attribution OpenAI, ChatGPT Agent System Card (2025-07-17). Reuse No open redistribution licence asserted. Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice. Check notes Candidate pack passage, AI and biosecurity (research/ai-bio). The quotation (25 words) matched the text retrieved for the candidate pack on 2026-10-28 and again at admission on 2026-10-30, in a fresh retrieval whose SHA-256 equals the pack’s. Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. A cross-family machine check (2026-10-30, item 33) re-read the record’s sources and found full support. No human editorial review. primary source Supports the claim
ChatGPT Agent System Card ↗ OpenAI · 2025-07-17
Archived copy, captured 2025-07-17: ChatGPT Agent System Card ↗
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 “experts noted an average success rate of … 7.7 (on a scale of 1-10) on tasks that they attempted with ChatGPT agent.” Location in source PDF p. 22 (printed p. 21), §5.1.1.7, second paragraph (the one beginning 'ChatGPT agent scored highly on non-agentic tasks'), closing sentence
How it was read Retrieved directly from the publisher
Machine-checked · 2026-10-30
Source details and check notes Authors OpenAI Version System card dated 17 July 2025 on its cover. The quoted sentence and its page match the PDF served on 2026-10-04 and the Internet Archive capture of 2025-07-17; the two copies differ only in chart labels. Canonical address Open the canonical source ↗ Accessed 2026-10-04 Language English (en) Archive copy Open archived copy, captured 2025-07-17: ChatGPT Agent System Card ↗ . Captured on the recorded publication date. An archive copy shows only what the archive retrieved on 2025-07-17. It does not prove what the page said earlier, and it can differ from the copy checked on 2026-10-04.Attribution OpenAI, ChatGPT Agent System Card (2025-07-17). Reuse No open redistribution licence asserted. Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice. Check notes Candidate pack passage, AI and biosecurity (research/ai-bio). The quotation (21 words) matched the text retrieved for the candidate pack on 2026-10-28 and again at admission on 2026-10-30, in a fresh retrieval whose SHA-256 equals the pack’s. Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. The ellipsis stands for a tilde in the PDF text layer. Locator corrected after the audit of 2026-10-28: the sentence closes the section's second paragraph, not its third. A cross-family machine check (2026-10-30, item 33) re-read the record’s sources and found full support. 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 report states the result of an evaluation it ran or commissioned, in its own words, with what was measured, by what kind of method, and the caveats the report itself gives (schema 2.22). The record is the report’s statement, not a verified finding: it says nothing about another system, a later version or real-world risk, and a null or limiting result is not evidence that no effect exists.
Date. As reported in the card of 17 July 2025; the card gives no dates for the sessions.
Record ID ab-ev-chatgpt-agent-expert-deep-dives