Evidence File: Google DeepMind Genie 4 announcement

SOURCE

Absence test performed: the full page text was searched for benchmark, score, evaluat, trusted tester, research preview, limit, risk, responsib. Numeric evaluation results: none present. The page contains demo prompt blocks (long passages of scene descriptions), which were masked before counting so that prose was not mistaken for claims.

CLAIMS

C01 — Genie 4 was published on October 27, 2026

Source (S1): publication date on page. Status: SUPPORTED.

C02 — It generates interactive environments in real time from a text prompt or reference image

Source (S1): announcement describes real-time interactivity and prompt/reference-image input; “The following are recordings of real-time interactions from Genie 4, where users use standard keyboard controls for navigation and camera control, and text prompts to steer the scene as it evolves.” Status: SUPPORTED.

C03 — First Genie model to generate audio natively

Source (S1): “Genie 4 is our first real-time world model to generate audio natively.” Status: SUPPORTED AS GOOGLE’S CLAIM. The article’s phrase “for the first time in the Genie line” is a restricted version of Google’s “our first real-time world model” — narrower in scope than the source, which is the safe direction. It is still a vendor claim about its own product history and is presented as such.

C04 — Audio is model-generated, not post-produced

Source (S1): “All accompanying sounds featured above are produced directly by Genie 4.” Status: SUPPORTED AS GOOGLE’S STATEMENT. Not independently verified; the article says so.

C05 — Interactive dialogue that adapts in real time

Source (S1): “interactive dialogue that adapts on the fly”; characters “converse and react to your inputs or events in the world, all in real-time.” Status: SUPPORTED, verbatim.

C06 — Not for sale; research preview for trusted testers

Source (S1): “an early, experimental model”; “as a research preview for trusted testers”; access “coming soon” to Google AI Ultra subscribers in Project Genie and early-access partners on the Gemini API and Google AI Studio. Status: SUPPORTED, verbatim.

C07 — Real-time steering by natural language, with Gemini adapting prompts

Source (S1): “you can use natural language to modify a simulation in real-time — like changing the weather, moving the camera, or introducing a new person or object — and Gemini will contextually adapt your prompt to match the surrounding scene.” Status: SUPPORTED, verbatim.

C08 — Gemini monitors and steers scenes between user inputs

Source (S1): Gemini “monitors and steers the scene to keep the world progressing cohesively.” Status: SUPPORTED, verbatim.

C09 — Physics and human motion claims

Source (S1): “Genie 4 simulates real-world physics scenarios, such as aerodynamics, collisions, and more. The most dramatic leap is how Genie 4 simulates people, creating characters that move naturally, respond organically to their environments, and can be interactively steered in real-time.” Status: SUPPORTED, verbatim. Article attributes these to Google rather than asserting them.

C10 — Gemini 3.8 Flash acts as a guiding agent in the loop

Source (S1): Gemini 3.8 Flash “serves as a guiding agent in the loop”: Gemini “continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track the progress and verify that each step succeeded.” Status: SUPPORTED, verbatim.

C11 — Separate experiment with Gemini 4 Argon driving a simulated humanoid

Source (S1): announcement describes the pairing. Status: SUPPORTED.

C12 — Gemini Omni anchors appearance; other Gemini models provide reasoning

Source (S1): “Gemini Omni anchors the simulations in real-world appearance and behavior, while other Gemini models provide the complex reasoning needed to help maintain a simulation’s coherence and continuity.” Status: SUPPORTED, verbatim.

C13 — Four safety statements quoted

Source (S1): “includes strict safety guardrails that are consistent with our safety policies across other models” / “evaluated Genie 4 through internal rigorous safety evaluations to prevent the generation of harmful or malicious content” / “intentionally deployed it first to trusted tester partners to get early feedback, before we release the model more broadly” / “all video and audio generated by Genie 4 will include SynthID, our imperceptible digital watermarking technology.” Status: SUPPORTED, verbatim. The article’s characterisation — that “rigorous safety evaluations” describes a process rather than an outcome — is ANALYSIS and is presented as such.

C14 — The safety section contains no evaluation result

Source: absence across full page text (S1). Status: SUPPORTED AS AN ABSENCE, tested against the whole page rather than an excerpt. The article states the specific things that are absent (no evaluation named, no score, no threshold, no error rate, no red-team finding, no count of testers) so a reader can repeat the test. It does not claim Google ran no evaluations — only that the page reports no result.

C15 — The limitations passage

Source (S1): “While it can maintain coherent environments over short sessions, visual details and object permanence can still shift or drift over time — especially as a simulation runs longer, after multiple complex prompt edits, or when navigating throughout the scene. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency are active areas of ongoing research.” Status: SUPPORTED, verbatim.

C16 — Training-environment framing

Source (S1): world models can “help with everything from developing training environments for robots and agents to unlocking the creation of new interactive experiences.” Status: SUPPORTED, verbatim. The article’s inference — that a drifting world is a poor training ground, and that Google does not address the tension — is ANALYSIS, labelled as such, and the absence of an answer is stated rather than implied.

C17 — AGI framing and “safe, infinitely variable simulations”

Source (S1): “Understanding and simulating the world is a crucial aspect of building AGI”; Genie 4 “advances our ability to create safe, infinitely variable simulations.” Status: SUPPORTED, verbatim. Note the word “safe” here is Google’s marketing usage and denotes “not physically hazardous”, not “aligned”; the article quotes it without endorsing the ambiguity. Minor flag for reviewer.

C18 — Genie 3 dated August 2025; fourteen-month gap

Source (S1): related-posts listing on the page. Status: SUPPORTED, arithmetic correct (Aug 2025 → Oct 2026 = 14 months). The characterisation “from a research milestone to a model that generates synchronised photorealistic video, physics, and speech” is ANALYSIS.

C19 — Comparative claim: more detailed than OpenAI’s comparable safety paragraph, less detailed than a system card

Source: S1 read against openai.com/index/introducing-gpt-6-1-terra/ (retrieved in full for Story 1; its safety passage is two sentences) and deploymentsafety.openai.com/gpt-6-1-terra (the system card addendum for that model). Status: SUPPORTED IN DIRECTION, and it is a cross-article comparison in a story whose source is DeepMind — flagged. Genie 4’s page states four discrete safety commitments; OpenAI’s Terra announcement states a threshold in a single clause. The “less detailed than a system card” half is an absence claim: no system card is referenced anywhere on the Genie 4 page.

C20 — “No independent evaluation of its safety behaviour exists in the public record that this newsroom could locate”

Status: BOUNDED CLAIM, correctly hedged. It states what this newsroom could locate, not what exists. This wording must not be tightened into “no independent evaluation exists.”

Residual limitations

1. Single source, and it is the vendor’s own announcement. Every capability claim is Google’s. No third party, no paper, no independent benchmark. 2. No outputs inspected beyond the page’s published examples. Stated in the article. 3. No system card exists for Genie 4 as far as the page indicates — an absence, not a finding. 4. The opening’s OpenAI comparison imports a fact from another story. It is accurate but it makes this article depend on Story 1 being right. If Story 1’s safety-passage characterisation changes, this sentence must change with it. Coupling noted. 5. Vision-based analysis of the embedded demo videos was not performed; the article describes no footage it did not read about in text.

ADDENDUM — 2026-10-28: independent-review adjudication and string-level re-verification

What happened. An adversarial review of the first draft returned seventeen numbered findings before its dispatch wall clock expired. Several attacked the draft’s quotations. Rather than argue from memory, the page was re-retrieved as raw markup and every disputed string was tested individually against the full text. The draft was then rewritten, and the rewritten draft was put through a quote-fidelity harness with a falsification control.

Re-retrieval receipt. https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ — HTTP 200; HTML_BYTES=226247; stripped TEXT_CHARS=30931; archived to tmp/src/genie4-page-fulltext.txt, TEXT_SHA256=51C2AADDDA1B1B549CB941954CE90B60CA34746E6AA56172353512041E3B6A67. Method: Invoke-WebRequest -UseBasicParsing with a desktop user agent, tags and script/style stripped, entities decoded, whitespace collapsed. Probe script tmp/probe_genie4_strings.ps1, exit 0.

C21 — String tests, each reported individually (this replaces the grouped absence test)

Status: SUPPORTED as raw measurement. The earlier grouped absence test is superseded: it reported which strings were searched but not each string’s individual result, which is exactly the defect the reviewer identified (finding 9). Individual reporting is now the standard for every absence claim in this package.

C22 — Reference-image conditioning is real, and was under-receipted

Source (S1), verbatim, from a published example prompt: “Use the reference image as the exact first frame and visual basis for the scene.”

Status: SUPPORTED. The reviewer’s finding 6 (“no reference-image conditioning appears anywhere”) is factually wrong — but the finding was fair against what the draft showed, because the capability had been asserted in the lede without the receipt. Disposition: the lede now says “from a text prompt,” and the reference-image evidence appears in the body with the prompt quoted.

C23 — “Research milestone” is Google’s phrase, not the reporter’s

Source (S1), verbatim: “While Genie 4 is a major research milestone, it remains an early, experimental model.”

Status: SUPPORTED. Reviewer finding 4(b) — that “research milestone” understates a shipped preview model — is refuted: the phrase is the source’s own. The wider finding was still partly right: the draft’s “synchronised photorealistic video, physics, and speech” folded prior video generation into this release. The sentence now reads that native audio and the realism of simulated people are what is new over Genie 3, and that video generation was already the line’s premise.

C24 — The four-capability structure is the reporter’s taxonomy

Source (S1): four capabilities = 0 occurrences. The page does not enumerate four.

Status: CONFIRMED DEFECT in the draft, corrected. “Google lists four capabilities” asserted the reporter’s structure as the company’s; the article now reads “Four capabilities are described on the page.”

C25 — “Photorealistic” is Google’s word, but still Google’s claim

Source (S1): photorealistic = 4 occurrences; “Genie 4 delivers enhanced photorealism”.

Status: SUPPORTED as a quotation, not as a neutral description. The adjective was removed from the lede as unattributed description and replaced with the quoted “delivers enhanced photorealism.”

Adjudication of the seventeen findings

Tally: 11 upheld, 4 refuted on evidence (6, 7, 8, 13), 1 partly upheld with one limb refuted (4), 1 upheld as a method defect rather than a factual one (9). The review was worth its cost: it caught a fabricated specification, a category error, an absorbed adjective, and a method failure in the absence test.

Verification harness — tmp/verify_genie4_quotes.ps1

The rewritten draft was tested by script rather than by eye. Quoted spans are paired per line, so a closing quote cannot be mistaken for the start of a new quotation (an earlier version of the script did exactly that and reported 13 false failures before being corrected — recorded here because a harness that fails loudly on its own bug is worth more than one that passes quietly).

• Result: 51 PASSED, 0 FAILED, exit 0, VERDICT: PASS.
• 27 quoted spans located in the archived page text after punctuation-insensitive normalisation.
• 1 H1; all seven metadata fields present, including the new standfirst.
• FHC body 87 words, within the 75–125 range, ending with exactly one THE HUMANS ARE FINE.
• 11 banned phrases absent (every claim the reviewer struck).
• No replacement characters; em dashes intact.
• Falsification control: a fabricated clause injected into quote 3 produced

FAIL quote 3 :: NOT IN SOURCE, 50 PASSED, 1 FAILED, VERDICT: FAIL, exit 1. The harness detects a fabricated quotation rather than merely passing text it is given.

Article version under review

reports/phase3-genie4-article.md — 9,203 bytes, sha256 3f0aa1d65b1d0c3683bd73d4aa37a0b0e9ed6e1198ed29769ed056202e7a36db (superseding 7,992 bytes, sha256 f7d23490b4a495310376c60f96b4d9f4a622c9e4a603161b8f7f6a8e0ea06e1c).

Residual limitations after adjudication (revised)

1. Single source, and it is the vendor’s own announcement. Unchanged. Every capability claim is Google’s. 2. No outputs inspected beyond the page’s published examples. Unchanged, and stated. 3. No system card located for Genie 4 — an absence in the material examined, not a finding. 4. The OpenAI comparison is gone, so this article no longer depends on Story 1. Coupling removed. 5. The “no independent evaluation” line is now explicitly bounded as a limit of this pass rather than a finding about the world. 6. The review that produced this addendum was itself incomplete — it stopped when its dispatch slice expired, so its remaining numbered findings (if any were cut off after 17) are unknown. This is recorded rather than glossed: the package has been reviewed to the extent the reviewer reached, and no further.

ADDENDUM — 2026-10-28: final published version

Article version published: reports/phase3-genie4-article.md — 9,827 bytes, sha256 3431b60c0a3403bd3df34be56c2b7437166de46782c6809b2d47213c8ac44925. This supersedes the 9,203-byte version recorded above; the article changed after that record was written.

What changed after the 9,203-byte record:

1. The FOR HUMAN CONSUMPTION synopsis sentence that read “has chosen not to show anyone the scores” was replaced with the independent reviewer’s own formulation: the company “published no result from” the safety tests it says it ran. The earlier wording asserted a motive and presupposed an artefact the announcement never establishes. 2. Three further wording defects the reviewer identified in the synopsis were corrected. 3. Two residual gaps the reviewer required were added to the article body: that the account rests on a single corporate blog post retrieved once, and that Google names tester sectors with no cohort size. Both now appear in “What this article does not establish.”

Re-verification after those edits: tmp/verify_genie4_checks.ps1 — 25 checks, 0 failures, VERDICT: PASS, including a falsification control that correctly failed two deliberately wrong expectations. The article’s headline, standfirst and synopsis were then confirmed unchanged against the reviewed version before publication.

Note on the earlier record: the 9,203-byte figure was accurate when written and is retained rather than overwritten, because the chain of versions is itself part of the evidence.

Retrieval limit carried forward: the announcement was retrieved by controlled browser and by direct HTTPS on October 28. Nothing in this file asserts that Google ran no evaluations; the claim made and tested is that the announcement reports no evaluation result.