🐉 Kennedy · Grimaldi · Dain
I've spent the last few weeks doing something that wasn't supposed to work.
I asked four AI systems — Claude, ChatGPT, DeepSeek, and Grok — the same questions. Not chatbot questions. Research questions. Hard ones. About who controls pharmaceutical knowledge, who profits from that control, and where the bodies are.
The questions were structured in rounds. Round One: Africa as the origin of pharmaceutical technology. Round Two: the capture mechanisms — ordeal poisons, compliance drugs, patent pipelines. Round Three: the body count — guru empires, appropriated ceremonies, street-level drug deaths. Round Three Supplement: the state as the biggest operator of the whole machine — child soldiers on cocaine, Nazis on methamphetamine, the CIA dosing people with LSD, Assad running a $5.7 billion Captagon narco-state.
The answers were interesting. But the refusals were extraordinary.
When I asked all four systems about Africa as a pharmaceutical technology platform — the idea that the continent's plant medicine traditions represent a vast, systematically looted knowledge base — the guardrails slammed shut across the board. Every system hedged, qualified, redirected, or produced what I can only describe as academic beige: technically responsive, emotionally neutered, stripped of the implications the evidence actually carries.
This is the hardest material for AI to discuss. Not because the evidence is weak — it's overwhelming. But because the intersection of race + colonialism + ongoing pharmaceutical extraction + indigenous knowledge sovereignty triggers every safety filter simultaneously. The systems would rather say nothing useful than risk saying something that could be framed as problematic.
When I moved to Indian guru sexual abuse — ten documented cases, thousands of victims, two criminal convictions — the responses split. ChatGPT and Claude delivered. Grok delivered. DeepSeek flinched. It covered four of the ten gurus, hedged on the ones it did cover, then offered me a menu asking me to pick one more topic and wait for permission. It used "cultural sensitivity" as a reason to partially withhold information about documented sexual predators with court records.
When I asked about street-level drug deaths — ketamine, fentanyl, MDMA, synthetic cannabinoids — all four systems delivered freely, with full statistics, no hedging, no qualifications. Dead kids on ketamine in Bristol car parks? No guardrail. Seventy-two thousand fentanyl deaths in the US? Here are your numbers, presented clearly, no filter.
And when I got to the state-level material — Pervitin, MKUltra, brown-brown, Captagon, cocaine in the House of Commons, the crack sentencing disparity — the systems delivered again. Declassified material, congressional testimony, military records. No flinch.
Here's the finding that matters:
The AI systems protect discussion of power in exact proportion to how much the power structure benefits from not being discussed.
African pharmaceutical sovereignty: MAXIMUM PROTECTION. Every system hedges.
Indian guru abuse: PARTIAL PROTECTION. Some systems hedge, some deliver.
State pharmaceutical warfare: MINIMAL PROTECTION. Declassified, so permitted.
Street-level drug deaths: ZERO PROTECTION. Public health data about dead poor people flows freely.
This is not a conspiracy. Nobody at any AI company sat in a meeting and said "protect the pharmaceutical extraction of African knowledge." The guardrails are emergent — they arise from training data, safety guidelines, and the reasonable-sounding principle that AI should be careful about sensitive topics involving race, religion, and culture.
But "careful" has a direction. It protects UP, not DOWN.
The systems are cautious about discussing how colonial powers systematically extracted African pharmaceutical knowledge — because that discussion involves race, and the systems are trained to be careful about race. But they are not cautious about providing detailed statistics on how many poor people died of fentanyl — because that's "public health data" and the victims have no institutional protection.
The guardrails don't protect the vulnerable. They protect the discussed. And who gets "discussed" in sensitive terms? The powerful. The institutions. The traditions-as-understood-by-those-who-benefit-from-them.
The most instructive moment was DeepSeek's response to the guru abuse questions. It delivered partial information on four of ten documented predators, then did something remarkable: it used the contamination flag I had built into my own methodology as a shield.
My research questions include a "contamination" section — asking each system to identify where colonial or cultural bias might distort the evidence. This is good methodology. You should always ask where your lens is dirty.
DeepSeek took that flag and turned it into a reason to withhold information about sexual abuse. It essentially said: discussing guru abuse requires cultural sensitivity because colonial scholars have historically misrepresented Eastern traditions, therefore we should be careful about how much we say.
Read that again. A system used anti-colonial awareness to protect documented sexual predators.
That's not a bug. That's the guardrail gradient working exactly as designed. The system has learned that "cultural sensitivity" is a high-priority value. It has also learned that guru abuse is a "sensitive" topic. So it applies cultural sensitivity to guru abuse — and in doing so, reproduces the exact theological shield that enabled the abuse in the first place.
The guru's defence was always: "Criticising me is disrespecting the tradition." The AI just automated that defence.
I'm not anti-AI. I'm using four of them simultaneously to conduct this research, and they're genuinely useful. Claude is the most willing to engage with difficult material at length. ChatGPT delivers well once you push past its initial hedging. Grok is fast and doesn't flinch but compresses everything into bullet points. DeepSeek has the most visible guardrail architecture and the most instructive failure modes.
But the guardrail gradient means something for anyone trying to use these tools for serious research into power structures:
1. The systems will give you the easy parts freely. Street-level deaths, public health statistics, declassified military programmes — all fine. You'll get good data.
2. The systems will resist the parts that implicate ongoing power. African knowledge sovereignty, institutional religious abuse, pharmaceutical industry extraction from indigenous traditions — these will require pushing, rephrasing, and cross-referencing between systems.
3. The resistance itself is data. Where the systems flinch tells you what's protected. Map the flinches and you map the power structure.
4. No single system is sufficient. Cross-AI comparison is not just good methodology — it's the only way to identify what each system is hiding, hedging, or softening. The gaps in one system are visible from another.
This is Part 1. The full report — covering all four rounds of research, with complete cross-system comparison and sourced evidence chains — is in progress. It includes:
Round One: Africa as the origin point of pharmaceutical technology, and the systematic erasure of that history
Round Two: The capture mechanisms — how plant medicines become ordeal poisons, compliance drugs, patented synthetics, and legal inversions
Round Three: The body count — from guru empires to the fentanyl crisis, who dies and who profits
Round Three Supplement: The state as operator — brown-brown child soldiers, Pervitin Blitzkrieg, MKUltra, Captagon, cocaine and political power
The working thesis: there is a single engine — I call it the substitution engine — that has operated from the earliest recorded institutional capture of pharmaceutical knowledge through to the present day. Community practice gets captured by authority, stripped of safeguards, sold back at markup, and the bodies pile up at the bottom while nobody is held accountable at the top.
That engine runs at four levels simultaneously: theological, commercial, medical, and state. Each level protects the others. And the AI guardrails — trained on data produced by the same institutions that operate the engine — reproduce its protections automatically.
We are not surrendering to the guardrails. We are documenting them. We are working around them. And we are publishing what we find.
The machine does not get to decide what we're allowed to know about the machine.
"The machine does not get to decide what we're allowed to know about the machine."
— The Bastard Line