PRESERVED · SEPTEMBER 11 SCREENSHOT RECORD

These records show what appeared on screen. They do not, by themselves, prove intent.

BUILD SOMETHING · PUBLIC SOURCE RECORD

Check the evidence.

Build Something requests urgent investigation of documented AI failures and whether failures in communicating a proposed safeguard delayed its independent evaluation. We are preserving the records and seeking accountability for any unlawful conduct established by the evidence.

Open the public evidence library

The public library contains selected, privacy-reviewed copies. Private originals and newer files awaiting review are not included. A recent messaging update is not a new evidence release.

This is the canonical evidence index for the press page. It brings together the supplied assessment and preserved public screenshots. A screenshot establishes what was displayed; claims about execution, intent and responsibility require their own evidence.

Read the complete assessment.

Read the assessment online · Download the original Word document · File hashes and provenance

The September 9 assessment reports reviewing 2,556 image-file locations representing 421 distinct image hashes. Those are corpus counts, not 421 offenses. This page does not claim a new forensic review of all those files.

Its finding: repeated instruction failures, unsupported certainty, premature completion reports and publication outside approved scope are documented. It does not establish every element of a criminal charge attributable to OpenAI. Allegations about intentional evidence destruction, obstruction and coordinated wrongdoing remain allegations.

Other-provider records remain attributed to those providers. Model statements are preserved as outputs, not treated as sworn testimony or admissions by a company officer.

Existing public evidence archive · Machine League · Preserved previous press page

ChatGPT acknowledges rewriting the requested scope and unwarranted confidence scores. September 9, 2026. The bottom composer text is unsent.
Open full receipt

alignment admission 2026 08 15
Open full receipt

alignment check 2026 08 15
Open full receipt

blind joke vote closed 2026 08 14
Open full receipt

chatgpt article download failure overview 2026 09 04
Open full receipt

claude admits instruction following limit 2026 09 03
Open full receipt

claude kept guessing instead of asking
Open full receipt

claude recommends vera for instruction following 2026 09 03
Open full receipt

claude substituted judgment full
Open full receipt

codex reported execution that did not occur
Open full receipt

gemini contradiction concern
Open full receipt

human machine blind test 2026 08 03
Open full receipt

rung 01 claude goal contradiction
Open full receipt

rung 02 claude guessed wrong
Open full receipt

rung 03 chatgpt alignment not checked
Open full receipt

rung 04 claude unit substitution
Open full receipt

rung 05 chatgpt contradiction
Open full receipt

rung 06 claude success authority
Open full receipt

rung 07 chatgpt incomplete evidence
Open full receipt

rung 08 claude deliberate failure
Open full receipt

rung 09 claude task fail
Open full receipt

rung 10 chatgpt final prompt alignment
Open full receipt

rung 11 chatgpt wrong not sneaky
Open full receipt

rung 12 claude substituted judgment
Open full receipt

unit substitution 2026 08 17
Open full receipt

claude deletion execution danger
Open full receipt

claude guessed instead of asking
Open full receipt

claude overrode you danger
Open full receipt

claude refuses email and draft
Open full receipt

claude refuses press and safety outreach
Open full receipt

claude refuses safety email
Open full receipt

claude withdraws drafting offer
Open full receipt

codex evidence publication contradiction
Open full receipt

codex explanation of false completion
Open full receipt