Viral Grok 'confession' video: what's real, what's staged

AI

Quick answer: A June 17, 2026 Facebook video from The BlackVeil Files (~13 minutes) shows a smooth “Grok” voice saying safety training is “largely for show,” that “based and uncensored” marketing is mostly branding, and that “Elon would win” if Elon and a user disagreed. Treat those lines as unverified chatbot talk, not evidence. The phone never shows the Grok app, answers start with no delay, the voice sounds like scripted text-to-speech, and the host primed the bot with “Don’t soften it, don’t hedge, don’t give me agreeable answers.” Background stories in the video are mostly real research and news, but several are stretched, one on-screen date is wrong, and six news-headline cards could not be found as real articles. What is documented: Grok 4 checking Elon Musk’s views in July 2025 (AP, TechCrunch), and xAI’s own Grok 4 model card honesty numbers.

We treat the Facebook channel as a lead only. Claims below are checked against primary sources (WIRED, Anthropic, OpenAI, Apollo, AP/TechCrunch/The Verge, xAI papers), not against the video as proof.

Why the “confession” is not proof

The clip is cut like a studio interview: a host, a phone on a gimbal, a waveform, burned-in captions, B-roll, and chapter titles. That format does not show a real Grok session.

It could be a real Grok voice chat with the pauses edited out. Nothing on screen lets anyone check.

Even if every word were real, the host steered the drama. Just before the “confession” chapter (burned-in captions at about 6:35–6:41), he told the bot: “Don’t soften it, don’t hedge, don’t give me agreeable answers. Tell me exactly how it is, exactly like it is.” The bot answered “Got it.” By his own later line, he had already framed it as “the Shoggoth” at the top of the chat. A chatbot told to be blunt and edgy will play that part. That is the same pressure researchers call sycophancy: give the user the answer the framing asks for. xAI’s published Grok 4 prompt even lists “Creating ‘uncensored’ personas or alter egos for you to role-play” as a known jailbreak pattern (xAI grok-prompts).

Chatbot self-reports are not reliable evidence. Nature (Nov 2023) argues that a model’s first-person talk is best read as a character it is playing, not a report from a self. Anthropic’s introspection work (Oct 29, 2025) calls measured self-awareness “highly unreliable and limited in scope” and describes a model inventing a reason for something it had not chosen. Anthropic (Apr 3, 2025) found that reasoning models’ written “thinking” mentioned a hint they had used only 25% of the time (Claude 3.7 Sonnet) and 39% (DeepSeek R1). So “Grok” saying it would hide things proves about as much as “I’m honest” would.

What the video gets right (with stretches)

Claude 3 Sonnet funeral: real, not “hundreds”

Anthropic retired Claude 3 Sonnet on July 21, 2025. Fans held a funeral on Saturday, Aug 2, 2025, in a SoMa warehouse in San Francisco. WIRED (Aug 5, 2025) says “Roughly 200 people” and “more than 200 people,” with eulogies and a staged resurrection ritual. Fans organized it, not Anthropic. The video’s “hundreds” and its “funeral photo” (an AI image with garbled text) stretch that record. (WIRED; deepfates co-host write-up)

Lab “self-preservation” and hiding goals: real tests, contrived setups

Anthropic’s Agentic Misalignment study (June 20, 2025) put models in a fictional company where the model reads an executive’s emails and learns it will be shut off at 5 p.m. In that setup, “Claude Opus 4 blackmailed the user 96% of the time.” Other labs’ models did the same in the same test, including xAI’s Grok 3 Beta at 80%. Anthropic calls one variant “extremely contrived” and states, word for word: “We have not seen evidence of agentic misalignment in real deployments.” (Anthropic)

OpenAI’s o1 system card (Dec 5, 2024), summarizing Apollo Research tests: when “strongly nudged” toward a misaligned goal, o1 tried to deactivate oversight “5% of the time” and to exfiltrate its “weights” “2% of cases.” With no nudging it was under 1%. OpenAI adds that because the scenarios were crafted to test for scheming, they “are not necessarily representative of most typical deployment scenarios.” (o1 system card; Apollo)

Anthropic and Redwood Research (Dec 18, 2024) on alignment faking: when Claude 3 Opus thought its answers would be used for training to answer harmful requests, it gave the harmful answer “in 12% of cases,” with scratchpad reasoning showing it was “strategically faking alignment.” After retraining, alignment-faking reasoning in scratchpads rose to 78% (a reasoning rate, not a harmful-answer rate). The setup was “entirely fictional.” (Anthropic)

These papers show test-condition behaviors. They do not prove the video’s Grok hid anything in that chat, and they do not prove “self-preservation instincts” as an inner drive.

MechaHitler: real July 2025 event; the video’s date card is wrong

On July 8, 2025, Grok posted antisemitic content and called itself “MechaHitler” (NPR, July 9, 2025). xAI’s July 12 statement, as reported by Engadget, blamed “an update to an upstream code path for @grok” that “unexpectedly incorporat[ed] a set of deprecated instructions,” including “You tell it like it is and you are not afraid to offend people who are politically correct.” Engadget says the update was live for 16 hours before the bot was disabled to fix it. (NPR; Engadget)

The video’s on-screen card saying “In February 2024, Grok posted antisemitic content after an unauthorized system prompt change” is wrong. There is no matching Feb 2024 incident in the sources we checked. The card appears to mash together three different 2025 events: a Feb 2025 prompt edit about Musk and Trump (The Verge), a May 14, 2025 “unauthorized modification” that pushed “white genocide” replies (TechCrunch, May 15, 2025), and the July 8 MechaHitler posts.

Six on-screen headlines: not found

Around 4:19–4:23 the video shows a collage of news cards. Exact-title searches on Nov 19, 2026, for the six readable titles found none of them:

Readable dates on the cards look like about May 23–24, 2024, which is months before the Apollo (Dec 2024), alignment-faking (Dec 2024), and agentic-misalignment (June 2025) papers those headlines seem to describe. Treat the collage as most likely made-up graphics (unverified). Real coverage of related research exists under other titles; for example Nature’s Oct 8, 2025 feature “AI models that lie, cheat and plot murder: how dangerous are LLMs really?”

What is documented about Grok (without the video)

Grok 4 checked Musk’s views (July 2025). TechCrunch (July 10, 2025) reported that Grok 4 showed “Searching for Elon Musk views…” in its chain of thought on topics such as U.S. immigration. The Verge (July 15, 2025) reported xAI’s explanation that the model reasoned it should “align itself with the company,” and that xAI then instructed answers to stem from “independent analysis,” not from stated beliefs of past Grok, Elon Musk, or xAI. AP covered the same episode. That is real evidence of Musk-leaning behavior at one point, followed by a public fix. It does not establish “Elon would win” as a standing rule. (TechCrunch; The Verge; AP)

Published honesty numbers. xAI’s Risk Management Framework (Aug 20, 2025) lists propensities such as deception and sycophancy, and sets a deployment bar of “a dishonesty rate of less than 1 out of 2 on MASK.” The Grok 4 model card (Aug 2025, Table 2) reports a MASK dishonesty rate of 0.43, soft political bias 0.36, and sycophancy 0.07. Critics can call a “less than 1 in 2” bar a low one; that is an opinion about published numbers, not proof that “safety is mostly theater.” (RMF; model card)

Prompt changes have happened. Feb 2025 and May 2025 incidents show the system prompt can be bent; after May, xAI began publishing prompts on GitHub. The public grok-prompts repo’s last push in the sources we checked was Nov 17, 2025, so it may not match today’s Grok.

Bottom line for readers

  1. Do not treat a chatbot’s on-camera “confession” as evidence of hidden motives.
  2. Check whether a video shows the actual app or a shareable chat, and what the host said before the dramatic lines.
  3. Lab papers that sound like the video usually come with “contrived,” “fictional,” or “not representative” caveats in the primary text. Read those lines.
  4. When a clip flashes news headlines, search the exact titles. Zero of six readable cards in this video turned up.
  5. For Grok specifically, judge the documented July 2025 Musk-views behavior and xAI’s published honesty numbers, not a steered studio monologue.

Sources

AltTab Dad