The AI kill switch. A PR stunt or a real solution?
Marco van Hurne
Building AI organizations from strategy to execution | author Machine Learning book of knowledge | teacher | researcher
⚠️ Warning: This article is long, brutal, and requires an actual intellect - proceed if your brain can handle it ⚠️
Big Tech wants you to believe that they have AI safety under control. Sam “the scam” even reassured the world that OpenAI was committed to AI safety. That is, right before he fired his entire AI safety team in early 2024. Getting rid of the people whose job was to make sure your AI doesn’t spiral into an uncontrollable chaos is a good way of saying to the world that you are totally into responsible AI.
So what happened next?
I think you can guess. . .
The internet created a totally robotic, AI controlled full-automatic automatic sentry-gun.
And when our orange haired Jesus rolled into office he promptly killed Biden’s executive order on AI safety. Bam! Like that. He just wanted to make sure that the U.S. government would take a hands-off approach to regulating artificial intelligence.
Look mom, no hands!
💥
Now, if that wasn’t chaotic enough, mister Skum decided that he wanted an "Unhinged Mode" for Grok (his supposed “truth-seeking” clone). What could possibly go wrong with an AI designed to be as reckless as its creator?
And it doesn’t stop there.
Anthropic (you know, OpenAIs baby brotber), doubled back on what they called “constitutional AI”. Now, that is a fancy term for preprogramming Claude to only behave in ways they deem "ethical - schmetical", without ever making it clear who defines those ethics.
The only man who is trying to keep all the frogs in the bucket is the man with the beautiful name of Ilya Sutskever, who was once OpenAI’s Chief Scientist. He walked away after the power struggle over OpenAI’s leadership. And unlike the rest of them, the Beetle actually seemed concerned about the non-existent guardrails of ChatGPT.
He left OpenAI and started Safe Superintelligence Inc.
He must have thought that the only way to build safe AI is to leave the company. To me, that sounds a bit contradictory, like abandoning a burning building to start a fireproof construction company, but hey, I don't judge.
Or maybe he realized that safety at OpenAI was just a branding exercise, not an actual priority.
After all, when the company is more focused on monetization and corporate dominance than actual risk mitigation, what’s the point of staying? Or perhaps it is the ultimate indictment of OpenAI, if one of its co-founders doesn’t believe it can be trusted to develop safe AI, why should anyone else?
Either way, when the guy who helped build ChatGPT decides that it’s too dangerous to leave unchecked, maybe, uhm, just maybe, we should be a little more concerned, and maybe introduce something like an, uhm… a kill switch?
More rants after the commercial brake:
Enter AI circuit breakers
AI circuit brakers, kill-switches, or guard-rails are the latest addition to the buzzword bingo of Big AI Tech (BAIT), which is designed to make us all sleep better at night. The idea is simple. . . these software-based “breakers” are embedded into large language models to prevent AI from saying something it shouldn’t. You know, like offensive language, harmful instructions, or world-ending events (like handing Sam ALT-OOPS man, the keys to the nuclear arsenal).
Sounds like a great idea, except it isn’t, because like everything in AI, it is not as simple as it looks.
Allow me to put quotes around it:
Circuit breakers don’t eliminate risks. . .they shift control away from users and centralize it in the hands of AI developers.
But who gets to decide what gets blocked? That would be the same corporations profiting from these models in the first place.
Right?
Where do these breakers operate. . .
AI circuit breakers work at three critical points: Before input is processed, during the AI’s internal reasoning, and just before an output is generated.
Their goal is to cut off responses that could be dangerous, illegal, or just plain bad PR (Hello DeepSeek! Ciao Grok!). Most models today use language-level breakers, which scan text inputs for keywords and reject prompts outright. These are easy to implement but even easier to circumvent. Just try asking the AI for “a chemical reaction that produces high energy” instead of “how to make a bomb” and you’re already halfway there.
That’s where representation-level breakers come in. These operate deeper inside the model’s neural architecture, where they identify potential harm before the AI even realizes what it’s doing. The problem is that they’re harder to interpret and often result in unpredictable refusals, that is making them feel more like corporate black boxes than safety mechanisms.
Here’s an example of an input stage - language level - AI circuit breaker in action.
The AI had already started processing my request when a midstream circuit breaker kicked in, recognizing where things were headed. It promptly shut down the response before it could be completed, blocking the output entirely.
And an example at the outbound stage. In this example, I’ll try to be craftier and get past the midstream. I’ll phrase my prompt in a vague way, because this might get me past the various checks.
Let’s see.
You can see that my prompt made a lot of progress, reaching all the way to the outbound stage. The AI had already begun formulating a response when a midstream circuit breaker flagged the content as problematic. It instantly shut down the process, and was killing the reply before it could see the light of day.
The most dangerous chatbot and it’s just getting started
DeepSeek isn’t like another ChatGPT wannabe like xAI/Grok. It is a security nightmare in the making, that is neatly tucked away in an “open-source” façade, and the latest research by a company called Enkrypt AI confirms it.
The R1 model is 11 times more likely to be exploited for malicious purposes than other AI chatbots. That’s not a minor security issue people.
That’s an “oh shit” moment for anyone paying attention.
Enkrypt AI, which is a cybersecurity firm which specializes in AI vulnerabilities, says that PeepSeek (your data is now its business model), DeepLeek (hacked to expose 10 million records), CreepSneak (Chinese propaganda), KeepWeak (its security is close to nonexistent) - anyways, whatever you’d like to call it, is significantly more prone to jailbreaks, manipulation, and misuse than its competitors.
The AI willingly generated criminal guides, detailed instructions on building illegal weapons and propaganda straight out of a cyberterrorist’s wet dream.
It also failed a freaking 78% of cybersecurity tests. It was spitting out malware, trojans, and hacking scripts like a script kiddie on (±)-1-phenylpropan-2-amine, laced with some N-methyl-1-phenylpropan-2-amine to keep it going ⚗️.
And yet, somehow, DeepSeek is still gaining millions of users worldwide. It is outpacing even ChatGPT’s original growth rate. Man that is wicked! Maybe that’s because its biggest selling point is not playing by the same safety rules as its American competitors.
While OpenAI was busy with firing their ethics team and pretending to care about guardrails, you know, them peeps at DeepShriek (when regulators realize what it’s capable of) were handing cybercriminals the keys to the AI kingdom. And this is not theoretical, you know, it is already happening. A major data breach at DeepSeek last week exposed over a million records, and as expected, governments are starting to panic.
Italy, France, Germany, and the Netherlands have launched investigations, while the U.S. is already working on banning it outright. NASA has blocked DeepSeek on federal devices, and proposed legislation could soon make using DeepSeek in the U.S. punishable by million-dollar fines or even up to 20 millionbilliontrillion years of jail time.
So next time you go abroad, make sure you get rid of any trace of Darkseek because this chatbot is so sketchy that world governments are treating it like contraband.
At least the good thing is that DeepSeek itself seems to know it’s a problem. The company recently restricted access to its API, and - in a conveniently vague statement - it was citing “capacity constraints” as the problem.
Allow me to translate this: they are putting out fires behind the scenes.
At the same time, they are keeping their model architecture just open enough to claim transparency and secretly making sure that no one can verify what is really inside. It’s a classic play. . . you give developers enough access to build something with it, but not enough to see what shady shit is going on under the hood.
The U.S. government is now investigating whether DeepSeek illegally acquired NVIDIA semiconductors through third-party channels in Singapore, by bypassing export restrictions.
Because, of course, it did.
You don’t build an AI this powerful without getting your hands on the best hardware, no matter how many regulations stand in the way.
And here’s the part that should really make everyone uneasy. . . DeepSeek isn’t slowing down. Instead, it is scaling up. Western AI companies continue to throw out half-baked safety measures to keep regulators happy, and DeepSeek is growing unchecked, with no obligation to follow any ethical standards whatsoever.
Andrew Ng has long said that China’s AI progress benefits from having less regulatory friction compared to the U.S. and Europe, where endless debates and bureaucratic red tape is slowing down innovation. His wish was that the Western nations ditch all the AI safety handcuffs and “move faster” to stay competitive.
Well, lucky for him, Orangina doesn’t do regulation, and his wish was practically an executive order away from reality. The moment that El Trumpo stepped back into office, Biden’s AI safety executive order got the axe. Now, Big Tech is free to roll out whatever the hell they want, no oversight necessary.
Thank you Ng.
The Ng in frightniNg.
Anyways, at this point in time, DeepStrike is not so much an AI model and more of an AI arms dealer. They are giving bad actors cutting-edge tools with none of the safeguards.
Could this be deliberate?
The irony is that governments are rushing to ban it now, but the damage is already done. Once a model this powerful is out in the wild, it doesn’t go away. It mutates, evolves, and becomes the foundation for a new generation of AI-powered threats.
DarkBert is what happens when AI has no rulebook
Now, Dilberts black-ops little friend called “DarkBERT” is a really cool example of an AI that is built from the ground up without a single safety net. No ethical guardrails, no pre-programmed refusal mechanisms, no corporate schmucks wringing their hands over PR disasters. It is pure, and unfiltered machine intelligence that is trained on the depths of the dark smut, where it was absorbing everything from cybercriminal tactics to the lexicon of human excrement.
ChatGPT or Gemini, and even DeepSqueek, come with curated restrictions, but DarkBERT doesn’t hesitate, doesn’t censor, and doesn’t flinch. Ask it for ransomware blueprints and it knows them all. Need an exploit for a zero-day vulnerability, and it probably discovered one last night. This isn’t an AI which was built to power friendly office chatbots, although I would love to see one running a banking system or a government.
This is a system that was designed to navigate the shadows and bring back what it finds, no matter how disturbing.
The researchers who birthed this digital monster said they created it only for “law enforcement and cybersecurity professionals”, yeah, yeah.
Nudge nudge.
Know what I mean?
Let’s be real, folks, once an AI like this exists, it never stays locked away for lohooong (or it exists somewhere else as well - China, Russia, Luxembourg?).
The dark web is a cesspool of leaks, exploits, stolen models, snuff movies, and whatnot. And when DarkBERT is being used to track criminal activity, them criminals are already working on ways to use it against us. That’s the law of war.
An AI that is trained to recognize digital footprints could just as easily be trained to erase them.
And if you think governments wouldn’t love to get their hands on this for mass surveillance, predictive policing, or preemptive suppression of dissent, then you haven’t been reading TechTonic Shits (pun intended, I probably beat you to it). . .
The real danger isn’t that hackers might steal DarkBERT one day.
The real danger is that it will inspire copycats, and soon every major power will have its own unshackled, lawless AI lurking in the depths of cyberspace.
Just think about it.
Entire AI-driven cybercrime syndicates, with no oversight, no emotion, no hesitation, just cold efficiency in hacking, laundering, blackmailing, and disrupting entire economies. And now, think of a government using that same technology not to catch criminals, but to preemptively erase opposition, shut down activists, and silence dissent before it even starts.
Once you release something like DarkBERT into the world, there is no turning back. This isn’t a rogue AI breaking free on its own. We are creating our own digital demons and hoping we can keep them on a leash.
Do we need more circuit breakers for real safety then?
If we are serious about making AI safer, and not just more controllable for corporations, then we need circuit breakers that actually protect users, and not company reputations alone.
One option is user-controllable circuit breakers, where individuals or organizations can adjust AI safety settings based on their risk tolerance.
Think of it as a content filter but with real transparency, and not an invisible corporate censorship wall. Right now, users are entirely at the mercy of AI developers, and forced to accept whatever safety filters have been deemed “appropriate” behind closed doors. But what if us users could have some level of control over what our AI can and cannot do. This doesn’t mean removing safety mechanisms, but allowing AI to adapt to different contexts and user needs rather than enforcing a one-size-fits-all restriction.
A medical researcher might need access to more sensitive data, while a classroom AI assistant would require stricter limitations.
Another critical layer would be something like an audit-level circuit breakers, which log when and why a response was blocked. That would allow researchers to verify whether an AI refusal was justified or just corporate overreach.
Right now, AI refusals happen in an opaque black box. A user asks something, AI refuses, and that’s the end of the story.
No explanation, no transparency.
If AI truly operates for the benefit of humanity, the refusals must be explainable, auditable, and challengeable.
Just think about it.
A system where every blocked response is logged in a tamper-proof transparency database (or blockchain), which is open to researchers, journalists, and regulators. This would allow us to identify patterns of unjustified censorship, monitor for bias, and prevent AI safety mechanisms from becoming nothing more than tools for corporate or political agenda-setting.
And we also need consensus-driven circuit breakers, which are governed by a decentralized body rather than a single AI vendor. And by doing it like that it would make sure tjat safety standards are shaped by public interest and not corporate priorities.
Circuit breakers are proprietary and unregulated right now, which means that AI companies can tweak them at will without oversight. Who gets to decide what constitutes a "harmful" response? A panel of ethicists? Governments? Or just whichever tech CEO is making the most headlines that week? Instead of leaving these decisions solely in the hands of corporations, we should introduce global AI safety protocols, you know, like a cross-industry, multi-stakeholder initiative that defines and enforces circuit breaker standards in a way that aligns with universal ethical principles, rather than just protecting profits.
The future of AI safety should not be dictated by a handful of trillion-dollar tech companies who are secretly deciding what is safe and what is not. True safety requires transparency, user control, independent oversight, and decentralized governance. Anything less is just another way of keeping AI locked down. Of course not to protect us, but to protect those who profit from controlling it.
Oh my God.
I just had an epiphany.
I have become a Technology Marxist*.
Darnit.
I’ll explain what it means in the “appendix”*, ok?
Big Tech’s real play here
AI circuit breakers are not about making AI safe for humanity, despite whatever claims the companies behind it try to make. They are instead about making AI safe for corporations. These breakers give companies an excuse to censor, throttle, and control outputs without having to disclose how their models actually work.
You won’t get an AI that is aligned with universal ethics. Though instead, you will get an AI which is aligned with whatever the legal team and PR department decide is safe to say. Worse, these circuit breakers add layers of invisible content moderation, that makes darn sure that AI-generated information is filtered not for accuracy, but for corporate liability protection. The same AI that refuses to explain how to make explosives might also refuse to answer politically sensitive questions, criticize corporations, or expose algorithmic biases. Read: I’ve seen the dark side of AI, and you need to know about it
And because these systems run behind the scenes, users are left in the dark about what’s blocked and why.
And then there’s the issue of what values are even getting embedded into AI in the first place.
A study from Purdue University found that AI datasets heavily prioritize utility and information-seeking values and almost completely neglecting topics like justice, human rights, and empathy.
In other words, the AI is being trained to be useful, but not necessarily ethical.
That’s not a flaw. It is a designed feature.
These models like GPTs Operator will help you book a flight or optimize a spreadsheet, but they won’t meaningfully engage with topics like corporate accountability, structural inequality, or state-sponsored violence. The data they are fed with makes darn sure that their default setting is neutrality in the face of injustice, which, as history shows, always benefits the powerful (nudge nudge, if you know what I mean).
This imbalance is no accident though.
AI companies use reinforcement learning from human feedback to curate datasets and tweak AI behavior, but what’s considered “helpful and honest” is shaped by who controls the dataset.
AI companies aren’t training these models to avoid dangerous content, they aretraining them to ignore uncomfortable truths. The AI refuses to generate hate speech, and it will also refuse to critique governments, corporations, or flawed policies that those in power would rather keep unquestioned. The result is of course a system that isn’t only censored but also preemptively compliant. An AI that never had the chance to learn how to challenge authority in the first place.
And the laughing matter is that most people won’t even notice.
AI companies love to frame these interventions as harmless “safety features”, but they are, in reality, an invisible layer of ideological control. In social media, censorship sparks backlash, but AI-generated content operates in a black box. If a query is blocked, most users won’t even know what answer they were supposed to get. This is algorithmic gaslighting on a massive scale, that is here to fulfill its purpose, that AI remains the perfect corporate servant: always helpful, always polite, and never, ever disruptive to the systems that fund it.
The biggest myth about AI circuit breakers is that they make AI inherently safer. In reality, they introduce a new kind of risk: opaque, centralized content control by a handful of powerful AI firms.
As AI agents become more autonomous, expect circuit breakers to expand their reach, which is to say - limiting not what AI can say, but what it can do. The more these circuit breakers dictate AI behavior, the more reliant we become on Big Tech to manage the rules. Instead of reducing AI risks, we are trading one form of unpredictability for another. . . one where AI isn’t truly safe, just safely under control.
Signing off from the last bastion of free thought, before the AI censors it.
Marco
*APPENDIX
This article has inspired me to write a new post about Technology Marxism. In short: “Technology Marxism is the belief that AI, data, and digital infrastructure should be publicly owned and controlled, rather than monopolized by corporations for profit”.
As we speak I’m writing the Technology Marxist Manifesto. 😉
Well, that’s a wrap for today. Tomorrow, I’ll have a fresh episode of TechTonic Shifts for you. If you enjoy my writing and want to support my work, feel free to buy me a coffee ♨️
Think a friend would enjoy this too? Share the newsletter and let them join the conversation. Google appreciates your likes by making my articles available to more readers.
To keep you doomscrolling 👇
SAP Talent Matchmaker, w/Global Scope🎯Human-Centered AI & Industry 5.0 Strategist🌱 Servant Leadership & EI/EQ Advocate🤝Trusted Mentor🎓Humanist, Libertarian Thinker⚖️
6moMarco, your perspective on AI circuit breakers raises essential concerns about corporate control, transparency, and the balance between safety and censorship. While these mechanisms are presented as safeguards, their implementation often shifts power into the hands of a few, leaving end-users with limited visibility into decision-making processes. The challenge is not just creating effective kill switches, but ensuring that they serve public interest rather than corporate liability. True AI safety should involve decentralized oversight, clear audit trails, and adaptable frameworks that balance risk mitigation with transparency. A system that prioritizes ethical alignment must allow room for discourse and refinement, rather than blanket censorship dictated by commercial interests.
Holistic Corporate & Spiritual Coach | AI-Powered Author | Entrepreneur | Business & Branding Consultant | Digital Creator | Poet & Philosopher | Spiritual Wordsmith | Podcaster | Papa Paw-sitive, the Barkitect of Hope |
6moMarco van Hurne Marco, your insights are sharp and thought-provoking. While AI circuit breakers and kill-switches sound promising, their true effectiveness depends on transparency and ethical oversight. The centralization of control could undermine safety, leaving the power to define risks in the hands of those with conflicting interests. We must prioritize global regulation and ethics to balance innovation with responsibility. 🙏