AI does not pose an existential threat to humans, and this is a hill I will live on. Sorry to disappoint doomers, but we aren't going to see the world plunged into chaos by machines. For that to happen -- the kind of chaos seen in the Terminator movies -- AI will need to have will and sentience. This is not to be confused with what it already has: agency.
Agency is the capacity to control actions based on specific or general objectives. An individual or entity with agency can control their actions autonomously. Controlling one's actions does not mean you understand the reason for said actions.
I intend to draw fine distinctions between how AI systems work and how humans think. I’ll also show you how you are being manipulated into believing that AIs understand what they are doing -- what I call the illusion of causal reasoning. Finally, I’ll show you the way out.
First, let’s establish context.
How does AI process and parse information? I'll start with an unremarkable example of a parrot. A parrot can mimic human language, tone, and cadence. That a parrot mimics language expertly does not translate to an understanding of why words are used in a particular way or the true meaning these words convey (this is a painfully extreme simplification, bear with me).
Let’s get a bit more technical. While it can look like an AI is dynamically figuring out cause-and-effect on the fly, extensive computer science research shows that zero-shot capabilities (explained in more detail later) are generally driven by a different mechanism: advanced pattern matching, text correlation, and vast probabilistic knowledge.
The difference between zero-shot capability and genuine causal reasoning in humans comes down to several key factors:
1. Causal Knowledge Retrieval and Causal Inference
When an LLM answers a causal question correctly (e.g., explaining why a ball thrown up comes down or why a certain gene interacts with another), it is usually executing a highly sophisticated search and retrieval of patterns buried within its parameters.
The AI has never seen your exact wording before, so its successful response looks like an active deduction. But in reality, there’s an underlying web of correlations, linguistic templates, and factual data about the entities that exists heavily in its pre-training data. It is extracting learned statistical dependencies rather than generating independent causal logic.
2. Counterfactuals and Interventions
True causal reasoning requires the ability to navigate Judea Pearl’s Causal Hierarchy, which includes understanding interventions ("What will happen if we actively change X?") and counterfactuals ("What would have happened if X hadn't occurred?").
In zero-shot scenarios, LLMs consistently struggle when presented with unseen, counter-intuitive, logic-flipped scenarios.
Because their core transformer architecture is inherently autoregressive (predicting the next most probable word based on prior text), they lack an internal, dynamic world model that tracks shifting physical or logical states over time.
If you subtly alter the statistical features of a prompt without changing its core causal logic, a zero-shot model's reasoning often collapses completely.
AI, acting on its own, will never be an existential threat to human lives globally.
The above distinction between machine and human intelligence is important because I will predicate more points on it as I proceed.
For AI to potentially take over and cause an existential threat, it needs
- true sentience, not mimicry of human intuition
- ingenuity, not programmed random choices from a distribution curve.
Devious actors have managed to sell the illusory capabilities of AI as the truth, so it's now easier to convince people of the dangers of uncontrolled AI sentience than for people to see the truth behind the smokescreen of scientific jargon.
I can't even blame anyone for believing the fear-mongering. Humans have historically deified things they do not understand. But I hope that after reading this, you should be able to, by yourself, make the distinction between what AI can do and what it simply cannot! With this, you can take ownership of your own future, rather than cower in ill-defined fear.
That fear has only worsened as calls for AI development to be slowed or halted by CEOs, AI experts, and politicians in the United States have grown louder.
Notable figures like Geoffrey Hinton, Sam Altman, Dario Amodei, Elon Musk, and Bernie Sanders have publicly painted a rather grim end-of-the-world doomsday picture. It’s appalling and unfortunate.
If you look under the hood, the plot becomes clear.
How the AI Police is funded by the very people who stand to benefit from the success of frontier AI companies.
The entities proposing to act as the independent auditors of Anthropic and OpenAI are financially sustained by the same capital network that benefits if Anthropic and OpenAI succeed.
Dustin Moskovitz (Facebook co-founder) funnels billions into his Good Ventures Foundation. Good Ventures funds Coefficient Giving (co-founded by Holden Karnofsky, husband of Anthropic’s President Daniela Amodei). Coefficient Giving funds Alignment Research Center (ARC), METR, and Redwood Research (the “independent AI safety audit organizations”).
Meanwhile, Moskovitz's foundation holds a massive, highly valuable private equity stake in Anthropic.
I should also add that Paul Christiano, the founder of ARC (the parent company of METR), was Dario Amodei’s housemate and coworker at OpenAI. He’s on the board of Redwood and very recently, OpenAI.
Finally, US senator, Bernie Sanders convened a highly influential, closed-door Senate briefing this month to warn Congress that they had only one year to establish meaningful guardrails before AI could escape control. The research material he quoted was straight out of METR’s playbook.
If you somehow believe that the individuals in this convoluted web of very mutually beneficial relationships somehow value your well-being over their profit, then there’s not much I can do to help you. You can stop reading at this point.
I don’t think ARC, METR, and Redwood Research are the kind of organizations that will bite the hands that feed them.
If AI does not pose an existential threat, where does the real threat lie?
Shark attacks are correlated with increased consumption of ice cream. Now, if you told this to a child, the child could think there was a causal link between sharks and ice cream consumption. This is possible in a young child's mind because intuitively, correlation can seem like causation. An evil actor could take that accurate data and misrepresent it to that child by saying, “Don't eat ice cream or else you will get attacked by a shark!”
The child would believe it. But that's a misrepresentation of accurate information. Whereas, the true reason for the correlation is a confounding variable: summer heat. It's important to note that the people painting the fear narrative are aware of the truth but have chosen to misrepresent facts to drive it down a particular path. So now you understand how you are being manipulated. Let me provide a balanced perspective.
Zero-shot learning.
Generative AI, in recent years, has become the most rapidly evolving piece of technology ever.
The real threat, however, lies with zero-shot learning: the capacity AI has to correctly identify new concepts even when the information about the concept wasn’t explicitly labeled in its training data. It is an impressive capability, but it is not proof of true causal reasoning in large language models. It is merely the AI’s ability to retrieve factual data about the entities that exist heavily in its pre-training data.
The single most important application of zero-shot learning is in understanding, shaping, and mapping human ingenuity. Because human ingenuity cannot be truly replicated, anyone with the intention to teach a machine to apply the specific mental models of an individual would need to use zero-shot learning.
How does that apply to you? If you’ve ever used an AI chatbot for problem solving, someone with access to your chat sessions can use that data to train an AI to solve problems using your unique approaches.
“Someone” in this case is a euphemism for frontier AI companies. Every day you use AI chatbots to solve difficult tasks, you’re literally training your replacement in real time. Your data + zero-shot learning = AI adopting your ingenuity to solve difficult problems and make breakthroughs before you do.
It’s not confirmed, but OpenAI may or may not have used this approach to undermine the achievement of two mathematicians.
This is the real and greatest threat. Your intellectual property can be taken from you and used to make important breakthroughs without due reference, regard, or compensation, and you will be none the wiser.
Right now, across the world, people are working on important scientific, creative, and social endeavors, unaware that the better they get at solving problems using AI, the more “competent” their AI replacement will become.
I make bold to say that many big Gen AI breakthroughs so far have resulted from this kind of training. When you understand this, it immediately becomes clear to you that the real threat to humans is a threat to your ingenuity, because every person on the planet who uses AI very likely shares their thinking process and patterns, enabling AI to build a predictive map of their ingenuity.
With this ingenuity map, it can derive answers to questions that you are yet to answer. In fact, your data is being harvested from your chats for these exact purposes, even as you read this.
Now you know to what extent your very usefulness to society is threatened.
The way out (or in, depending on how you look at it).
The obvious question is: what can you do about this? There are two major things I advocate:
1. Proactive protection: Get ahead of them by creating a commercially relevant version of your expertise using AI. It sounds counterintuitive, but when you do, they lose commercial incentive to copy your expertise, as they did to Hayao Miyazaki.
They will not attempt to recreate or copy your ingenuity if you've already done that for yourself. Furthermore, with an AI clone of yourself, you can make more money and place a higher premium on your natural knowledge and ingenuity. When people see just how brilliant an AI clone of you is, they would pay a premium to consult you personally.
2. Prevention. Stop sharing sensitive information with AI (especially platforms that do not practice zero-data-retention). Use zero-data-retention AI models like Ollama’s offline models and other offline AI models that 100% do not, under any circumstance, transmit your data to external servers.
Lastly, the word processors you use, such as Google Docs and Microsoft Office, are connected to the internet, which translates to data leakage. Use your physical notebook and pen or open-source offline platforms like LibreOffice for working on sensitive documents.
Combining these two methods; proactive protection and outright prevention, should ensure you’re not only unstoppable, but can preserve and protect that which is most important to you and is most valuable to AI companies: your ingenuity.
Statement of benefit.
I run a consultancy where I make AI clones of thought leaders and people who want to protect their intellectual property. I also advise experts and high-net-worth individuals (HNIs) on strategies to proactively protect themselves from AI-related data harvesting. I do have a lot to benefit from enlightening you on the real threat to your ingenuity, and thankfully, so do you.
The decisions you make about how you protect your ingenuity today will shape your relevance tomorrow.