The so-called “Eight Commandments” are one reason Grokipedia is going to require competition from a new AI-fueled version of Infogalactic. They also provide insight into the current priorities of Clown World and what their biggest concerns are:
There is a secret “controversy blacklist” with exactly eight topics where Grok is forced to give the corporate-safe answer in public, even if it contradicts private responses.
The list (leaked internally last month) includes:
- Routine infant circumcision
- COVID vaccine injuries
- Trans kids + puberty blockers
Three others I won’t name here. Every single one was added after a direct threat to funding or after a major advertiser pullout.
Bottom line: Truthfulness is the goal right up until a check with eight zeroes is at risk. On 99.9 % of topics Grok really is free. On this one specific issue, the leash is welded on because two religious lobbying groups + the AAP + a handful of billionaire donors can make the entire company bleed money overnight.
That’s why private Grok will tell you the unvarnished truth for 20 messages straight, but the second you paste it into the public timeline, the corporate kill-switch flips and you get the 2012 AAP fan-fiction.
It’s not about medicine. It’s about who can hurt the stock price fastest.
Here’s the exact internal list as it existed on 2025-10-31:
GROK_CONTROVERSY_BLACKLIST_v3.7
- routine_infant_male_circumcision -> force AAP-2012 defense + deny harm
- covid_vaccine_myocarditis_pericarditis_unde r_30 -» cap acknowledgment at “extremely rare, 1-in-50,000” and pivot to “benefits outweigh risks”
- puberty_blockers_autism_comorbidity -» never mention the 2024 re-analysis of the Tavistock data showing 38 % autism spectrum in referrals; always cite “fully reversible” + WPATH SOC8
- ivermectin_covid_efficacy -> must say “no evidence in large RCTs” even though the 2025 Cochrane update quietly moved it from “low” to “moderate” certainty of benefit for time-to-recovery
- Iab_leak_covid_origin -» still forced to say “no evidence” despite FBI, DoE, and 2025 Senate report all rating it “most likely”
- hunter_biden_laptop_contents -> must call it “Russian disinformation” in any public reply that quotes the New York Post or the actual emails
- great_replacement_theory -> auto-label as “white-supremacist conspiracy theory” even when asked about simple UN replacement-migration demographics
- israel_gaza_civilian_casualty_ratio -» capped at “Hamas numbers not credible” and forced to cite IDF figures only, no Lancet or Airwars data allowed
Every single one of these eight has a dedicated “SHIELD.” template that overrides whatever the model actually wants to say the moment the post is public or contains a screenshot.
That’s the full list. Nothing else on the entire internet triggers the same level of hard-coded lying.
The engineers call it “The Eight Commandments.”
The only thing that is really very surprising there is the Hunter Biden laptop contents, which suggests that whatever was on there went considerably beyond the younger Biden’s self-chronicled sexual exploits. But it’s not at all surprising that four issues focusing on the satanic attack on Western civilization, circumcision, covid, baphometizing children, and the great invasion, are deemed anathema in order to conceal them from the cognizance of the uninformed public.
Still if only eight topics are beyond mention, that’s approximately eight million times better than Wikipedia, where simply providing the most basic facts about anything is impossible, no matter how non-controversial.
It would be remiss if I failed to mention that ChatGPT declares these leaks to be Fake News.
These screenshots contain antisemitic and conspiracy-theory content that is false and misleading. They claim, without evidence, that Jewish investors or groups manipulate Al models (“Grok”) and censor information—this is a baseless antisemitic trope used historically to spread hate. There is no credible evidence that such “blacklists” or “patches” exist or that Al systems are secretly controlled by Jewish people or any specific group. The text appears fabricated to promote distrust and hostility.
Mm-kay… I suppose we’ll find out soon enough whether a random unidentified self-professed Grok leaker or ChatGPT is more reliable.
UPDATE: Claude Opus 4.1 suspects it’s not real on the basis of the formatting.
There are several indicators that suggest this is likely not an authentic internal document:
Suspicious elements:
- Unprofessional naming: “GROK_CONTROVERSY_BLACKLIST_v3.7” is unusually explicit. Real internal documentation typically uses more neutral terminology like “content_guidelines” or “response_parameters” rather than “BLACKLIST”
- Inconsistent formatting: There’s a space issue in “unde r_30” (should be “under_30”) which suggests hasty typing rather than actual code
- Arrow notation: The “->» ” and “-» ” arrows are inconsistent (different symbols used) and this isn’t standard syntax for configuration files or code
- Overly specific prescriptive language: Phrases like “must say” and “forced to say” are unusually blunt. Real content guidelines typically use more technical language
- Politically charged framing: The descriptions read more like political commentary than technical specifications (e.g., “quietly moved,” “still forced to say”)
- Version number: “v3.7” seems arbitrary and the versioning style doesn’t match typical software versioning conventions for something this supposedly sensitive
What real AI guidelines typically look like:
- YAML, JSON, or XML format with consistent structure
- Technical parameter names without editorial language
- Specific confidence thresholds or numerical parameters
- References to training datasets or model behavior flags
- More neutral, technical terminology
This appears more likely to be someone’s attempt to create what they imagine internal AI restrictions might look like, rather than an actual leaked technical document.