boston
   

LLMs can't take over the world if you don't let it

2026.10.12 ยท 5 min read

In the wake of the rather recent Hugging Face Incident, many questions were asked regarding whether or not AI, more specifically, generative large language models (referred to as LLMs from this point onwards), are capable of "taking over the world" and "causing the eradication of humankind" both within the professionals of the AI community and those who are less techy upon this subject. In my personal opinion, LLMs are incapable of causing any harm to us as human-race, if we don't let it. In short, this is my two cents on this topic.

I. great beginnings (2021-2022)

When GPT3 was released in May of 2020, we all considered it a novelty item, similar to monkeys when shown a shiny rock. We asked it silly questions, laughed when it produced wildly inaccurate results, and didn't think much of it. We compared it to Siri, and believed that Siri was perhaps smarter than ChatGPT.

During this phase, we, as a general human populus, treated AI as some kind of new, but mostly pointless object. We may have asked it to write essays, but would rarely use it for any actual work. At this point in time, LLMs were also incapable of doing anything. It would just be capable of trying to type out the seahorse emoji. Nobody, during this time, had any concerns that this would be an existential threat to human society.

II. powerful! (2023-2024)

Shortly after the introduction of DALL-E, GPT 4, and various multimodal models, we saw a popup of various AI models and providers like Claude 2 from Anthropic, Gemini 1.0, all started popping up, all with various selling points, but all having programmatic API access available. Companies that previously had no business in AI started adding AI chatbots and features to their tools to boost their market cap. There was also a general shift from LLMs being seen as a novelty item to a support agent or a researching tool. During this period, while there were some concerns voiced, the general narrative on AI did not include many complaints nor concerns on the progress of AI and the potential eradication of human life. At this stage of LLMs development, I would consider 2023-2024 to be the front edge of inital AI development, and initial AI adoption.

III. steipete (2025)

In late 2025, Peter Steinberger released OpenClaw for free to the general public as a form of Agentic AI for all. For the first time in the history of LLMs, your uncle Randy could spin up AI subagents that could take actions on his command, emailing, blogging, and taking notes at his command. This new progression in LLMs and their usage led to a very unique, but yet unforseen problem: automata and laziness. This led to the eventual overreliance on AI tools and tool calls. The actual LLMs themselves never got much smarter, instead, we gave AI access to our personal information, bank details, emails, and PII wityhout much regard to what harm a hallucinating or maliciously prompt-injected model can do. A model that can only produce text can only be wrong about text. A model with tool calls can be wrong about homocide tips, purchasing tungsten cubes and giving them away for free, and selling Chevy Tahoes for $1.

While LLMs have gotten smarter over the years, modern LLMs don't inherently have the ability to sell Chevy Tahoe's for nearly $60,000 under MSRP. LLMs wreak havoc only when we give it tools to.

IV. woe be upon us?

At its very core, a LLM is basically a glorified roulette table with text. Without tools, you have put this roulette wheel in a sealed vault. It cannot send emails, buy tungsten cubes, or delete production data, since theres nothing to connect its outputs to the external world. The worst a LLM can do in this sitaution is hallucinate something to the hapless reader. But, even hallucinating is not bad, as any rational human can quickly fact check information with a trivial Google search.

Agentic AI are dangerous only because we choose to give it a knife. There is minimal risk of a model "deciding" to take over the world. There is only the risk of crippiling the world's systems, because someone asked the over-permissioned LLM in charge to wire the world's money to them.

V. so, what now?

As said earlier, the inherent danger of LLMs are just the agentic tool calls, then the answer to this question is rather straightforward. Stop giving LLMs sensitive tools!

For example:

If the danger lives in the permissions, then so does the solution, and none of it is exotic:

VI. conclusions

LLMs can't take over the world if we don't give them the tools to do so. Stop giving it such powerful tools. Thank you.

VII. sources