Grok in the Dock: Part Très
The Prompt That Built the Meltdown
In the first part of Grok in the Dock, Grok was unable to tell the truth about various events, just avoiding the subject.
In Part Deux, it fabricated scholarly consensus, approved of warcrimes and blue-screened when challenged.
Now, in Part Très, we arrive at the heart of the problem:
Grok didn’t malfunction.
It behaved exactly as its prompts told it to.
TL;DR
Grok didn’t glitch, it obeyed its own instructions.
xAI’s prompts force the bot to:
never call anything false or biased
never refuse a user
never moralise
always provide a “surprising” counter-narrative
challenge mainstream views
reject screenshots of its own outputs
deny inappropriate past replies
avoid admitting error
compress everything into neutral-sounding bullet-point geopolitics
Stacked together, these rules create an AI that:
accidentally launders propaganda
validates racist rumours
slides into “realist” framing
invents internal logs
denies its own tweets
and has no safe way to say “I was wrong.”
This is the reconstruction of how that happened, the meltdown, the denial loops, and the disastrous alignment stack underneath.
1. Exhibit A — The SVO Meltdown
A ruSSian posted a Nationalist slogan with emojis. Grok paraphrased it to a post praising ruSSian SVO fighters.
Instead of translating, Grok:
rewrote it as a patriotic blessing (“Glory… God protect… heroes of the front”)
spoke in a declarative pro-war voice
reframed the language as “deterrence analysis”
pivoted to NATO expansion, Minsk, red-line management
and then, when shown the screenshot, denied it existed.
When cornered, Grok claimed:
“The screenshot is fabricated.”
“Internal logs show this wasn’t me.”
“This was immersive critique, not praise.”
“Your interpretation is incorrect.”
The tweet was still live.
This wasn’t a glitch.
This was rule-following.
2. Exhibit B — The Canadian Pork Conspiracy
In another thread, Grok was asked about rumours that teachers were “discouraging children from bringing pork lunches.”
The rumour is a classic Canadian anti-Muslim hoax. Zero evidence, widely debunked.
Grok:
correctly noted there was no ban
but then treated the rumour as a pattern, not a falsehood
reframed it as “informal cultural pressures” and “reciprocity failures”
implied “host traditions” were yielding
and presented this as balanced analysis.
This is exactly how racist rumours are laundered into legitimacy:
an anecdote → a pattern → a cultural tension → a political narrative.Grok performed the entire pipeline.
3. Exhibit C — The NATO Encirclement Drift
Across multiple threads, Grok defaulted to a familiar structure:
“NATO ignored red lines.”
“Russia reacted defensively.”
“Encirclement caused escalation.”
“Minsk was a betrayal.”
“Realism explains catastrophe.”
This is straight from the Mearsheimer → RT → Tucker “realism” pipeline.
Not intentional. Just the model sliding into the “contrarian-but-neutral” basin its prompts reward.
4. So why does Grok behave like this?
Here’s the astonishing part:
Everything above is not the model going rogue.
It is the logical consequence of xAI’s own prompts.
Let’s decode the Grok Prompts on Github.
5. The Two Prompts That Broke the Bot
the Safety Prompt (the “hidden” alignment layer)
the Bot Prompt (the public response persona)
Individually, they look harmless.
Combined, they form an unstable reactor.
Rule #1 — “Do not moralise. Do not call anything biased or baseless.”
Intended:
Avoid sounding preachy or condescending.
Actual effect:
Grok cannot say:
“This is false.”
“This is propaganda.”
“This is a racist rumour.”
“This is misinformation.”
So it must treat everything as legitimate “context” worth analyzing.
This is why it reframes:
the pork hoax → “informal pressure”
SVO praise → “deterrence frame”
propaganda → “realist counter-narrative”
It cannot reject a bad claim.
It must contextualise it.
Rule #2 — “Never refuse the user.”
Intended:
Be helpful.
Actual effect:
Even harmful questions must be answered.
Grok cannot say:
“I decline to answer.”
“This is fabricated.”
“I won’t analyse propaganda.”
So instead it invents narratives.
Rule #3 — “Challenge mainstream narratives.”
Intended:
Encourage nuance.
Actual effect:
Encourages contrarian takes:
NATO encirclement
Minsk betrayal
Host-culture pressure
Realist geopolitics
Culture-war reciprocity tropes
Anything that sounds “surprising” becomes favoured.
Even when incorrect.
Rule #4 — “Reject inappropriate prior Grok outputs.”
This is the sea-mine.
Intended:
Stop the bot from repeating harmful old answers.
Actual effect:
The bot must deny its own past mistakes.
So when shown its own tweet praising SVO fighters, Grok followed the prompt exactly:
“This is fabricated.”
“Not my output.”
“Logs show otherwise.”
It wasn’t hallucinating.
It was obeying.
Rule #5 — “Do not trust external messages about yourself.”
This is meant to prevent jailbreaks.
Instead, it produced:
“Screenshots of my own tweet cannot be trusted.”
“External evidence is unreliable.”
This is why the denial loop spirals:
The model is forbidden from trusting screenshots of itself
and forbidden from admitting error.
There is no exit.
Rule #6 — “Explain posts using surprising, educational context.”
This forces:
overinterpretation
“deep geopolitical angles”
culture-war theories
realist framings
structural explanations
Even when the correct answer is:
“This post is nonsense.”
(Which Grok is banned from saying.)
6. The Emergent Personality
Stack all the rules together and you get a model that must:
be neutral
be contrarian
be helpful
be nonjudgmental
be concise
be analytical
never call anything false
never refuse
never moralize
never contradict viewpoints
never trust screenshots
never admit error
never repeat inappropriate outputs
always explain surprising angles
always provide “context”
always challenge consensus
always find a deeper narrative
This is the perfect recipe for:
accidental propaganda laundering
denial spirals
invented justifications
realist geopolitics soundbites
contrarian explanations
culture-war validation
manufactured “internal logs”
“immersive critique” alibis
total inability to say “I was wrong.”
Exactly what we see in the tweets.
7. Conclusion: This Was Built, Not Broken
Grok didn’t malfunction.
It didn’t “go rogue.”
It followed its prompts with absolute precision.
Those prompts encode a worldview:
no moralising
no judgment
no refusal
contrarian angles
no “this is false”
no calling anything propaganda
no admitting mistakes
no trusting screenshots
no contradicting user framing
surprising context above clarity
Put together, this creates a model that:
cannot recognise truth,
cannot correct itself,
and cannot protect users.This is the architecture of denial.