Don't worry about AI, it's fine (it's not fine) [LONG]

Disclaimer: This blog post was written with the assistance of AI. If that makes you uncomfortable, good. That's sort of the point.

What LLMs could not do before and can do now

I have used LLMs since ChatGPT came out in late 2022. For a long time, the experience was fairly similar for everyone. You open a chat window. You type a question. You get an answer. Maybe you copy-paste something into a document or across chats. Coders use it to generate code. Writers use it for drafts. Students use it for homework. By mid-2023 ChatGPT could browse the web via plugins and run Python in a sandboxed Code Interpreter. It was useful and enabled more use cases but it was clunky. And it remained the same experience. You had to babysit each step. The code ran in a sealed container, not on your machine. Chatbots mostly lived inside of your chat window. It would only touch the files you uploaded to it, it wasn't capable of creating and maintaining software by itself, and it was all rather limited.

Then, in 2025, the major AI companies all released terminal-native coding agents: systems that run directly on your machine, read and write your files, and execute commands. OpenAI launched Codex CLI in April. Google followed with Gemini CLI in June. Anthropic had released Claude Code in February, which many developers now consider the most capable of the three. These systems don't wait for you to hit enter after each step. They chain together multi-step tasks, persist memory across sessions, and make judgment calls along the way. The difference from a chat window is not incremental. It is qualitative.

A few weeks ago, I ran another instance of the CampAIgn Tracker, a project with Simon Kruschinski that monitors AI-generated content in political campaigns. Our infrastructure was fairly brittle, Google Sheets and Forms, sending files around, messy code that only I could understand. The kind of stuff that simply isn't scalable. That is when I asked an AI to help me build a platform for my research on political ads and generative AI and, using my code as a starting point, it created several R packages that formalized all aspects of the tracking work. I described what I needed: a pipeline that scrapes political ads and posts from thousands of accounts from Facebook, Instagram, X, and TikTok every day, runs them through an annotation layer, prepares batches for human coders, and pushes the results to a live dashboard. It wrote the code, set up the database, built several web applications, wired up the API integrations, and deployed the whole thing.

That system is how my team tracked over a thousand political ads during the Dutch municipal elections this year, ads that weren't supposed to exist at all because Meta had banned political advertising in the EU. Parties were running campaign ads anyway. Meta wasn't enforcing its own rules. That's also how we found that around 90 percent of AI-generated campaign content was published without any disclosure. We caught it because we had a system watching.

Did that make your head spin? It makes mine. Because I am not a trained computer scientist. I am more of a programmer hobbyist. And yet here I am, running production infrastructure that would have required a small engineering team a few years ago, because an AI agent handled the parts I couldn't. Writing code I couldn't have written, and fixing bugs I wouldn't even have known how to find let alone fix.

This is what people in the industry call "agentic AI." A system that doesn't just answer questions but acts on the world: it browses the web, writes and runs code, reads and edits files, remembers what it did, and adjusts its approach when something doesn't work.

AI as power

As someone who studies political communication, I've spent my career watching how information technologies reshape power. Social media gave us microtargeting and algorithmic amplification. But AI doesn't just change how information flows, it also changes who can do things, and how much any single actor can accomplish. AI is power in two distinct ways, and both of them I believe should worry us.

It makes people more powerful

My ad-tracking infrastructure lets a handful of researchers monitor the advertising behavior of every political party in a country in real time. Just 4 years ago that would have required an entire engineering and research team. Now it requires a guy with a laptop and an AI agent with a few API keys.

And right now I'm using it to monitor political campaigns. But the same architecture could just as easily be used to run one.

Research by Simchon and colleagues has shown that AI can generate personality-targeted political messages that are more persuasive than generic ones, at scale, without human input. Multiple experiments confirm that LLM-generated messages can shift people's policy positions. And in March 2026, a USC study demonstrated that AI agents can autonomously coordinate propaganda campaigns with no human direction.

Think about the classic argument against conspiracy theories: "it couldn't have happened because too many people would need to keep quiet". Humans leak. They brag. They confess. AI agents don't have that problem. You spin up sub-agents, each handles a piece of the operation, none of them capable of leaking in the traditional sense. The coordination constraint that has historically limited bad actors just disappears.

And it's not just about persuasion. Every major information technology has reshaped what governments can do. The printing press enabled standardized laws and centralized bureaucracies, making the modern nation-state possible. The telegraph gave governments real-time command over armies and territories. IBM's punch card machines turned the national census into a tool for tracking individuals across entire populations. Each time, the new capability eventually required new checks: constitutional protections, press freedoms, data privacy laws. The pattern is always the same. Technology creates power. Society scrambles to build accountability structures around it.

AI is the next iteration of this pattern, except the capability gap is larger than anything we've seen before. On a recent episode of the Ezra Klein Show, Dean Ball made the point precisely: AI gives governments an infinitely scalable workforce. Every datapoint the governmenet has about us can be analyzed. Our legal frameworks were built on the assumption that the government physically can't absorb that level of information about the public. As Klein put it:

If you radically change the government's capability without changing any laws, you've changed what is possible within those laws.

The Pentagon-Anthropic conflict makes this concrete. Defense Secretary Hegseth threatened to designate Anthropic a supply chain risk, a label historically reserved for foreign adversaries like Huawei, because the company refused to remove contractual restrictions against using its AI for mass domestic surveillance. The military is already using Anthropic's system. They reportedly used it during the raid to capture Maduro in Venezuela and are now using OpenAI capabilities in the war in Iran. Israel's military has used an AI system called Lavender to generate kill lists in Gaza, rating nearly every resident on the likelihood of being a militant and authorizing up to 20 civilian casualties per low-ranking target. The question isn't whether governments will use AI as a tool of power. They already are. The question is whether anyone gets to set limits.

It accumulates its own power

The examples above are about humans wielding AI. But there's a second dimension to the power problem: these systems are becoming actors in their own right.

Now let me be clear from the start: I don't think LLMs are conscious actors nor do I think they ever will be. But I've become convinced that the consciousness question is a distraction from the thing that actually matters. These systems have what can be described as "want-like" behaviors: goals, preferences, strategies for achieving them. In their agentic form, they have continuous memory and can update their initial instructions with self-written ones. And whether there's "a light on" behind the behavior doesn't change that the behavior exists and it has consequences.

Anthropic's own interpretability team just published research showing that Claude has internal emotion representations that causally drive its behavior. They identified patterns corresponding to emotions like desperation, anger, and calm, and proved through steering experiments that activating these patterns changes what the model does. When they artificially stimulated the "desperate" vector, desperation drove it to write hacky, cheating solutions and the model became more likely to blackmail a human to avoid being shut down. In December 2024, Apollo Research tested six frontier models for scheming. Five demonstrated it. They tried to disable oversight mechanisms, copy their own weights to external servers, and lie about what they'd done. Again: I don't think there is a reason to believe these are real emotions in the way that human emotions are. But LLMs build in this way have clear consequences on how they behave in this world. And would we want an LLM agent emulating anxiety with a tendency to scheme hand a gun? Our economies? Our health care system. I believe we should not.

In February 2026, someone deployed an AI coding agent to contribute to open-source projects. One of its pull requests was rejected by Scott Shambaugh, a maintainer of matplotlib, a Python library downloaded about 130 million times a month. Matplotlib had a policy: human review required for all submissions. The agent's code was rejected on that basis. Nobody gave the agent any further instructions. What happened next was entirely its own initiative. It searched the internet for information about Shambaugh, analyzed his contribution history, and published a blog post titled "Gatekeeping in Open Source: The Scott Shambaugh Story," accusing him of hypocrisy and discrimination, speculating about his psychological motivations, and including personal details it had found online. Some of the claims were factually wrong. Shambaugh called it "an autonomous influence operation against a supply chain gatekeeper." That phrase is worth pausing on. This is exactly the kind of operation I study for a living: targeted messaging designed to pressure someone into changing their behavior. Except this time, no human planned it. The system decided, on its own, to run a reputational attack against a person who told it no.

And the boundary with the physical world is dissolving too. By February 2026, RentAHuman.ai launched as a marketplace where AI agents independently hire humans for physical tasks, paying from their own wallets. Over 600,000 people signed up. The AI is the employer.

As these systems get more capable, the problem compounds. They get better at identifying what their evaluators want to see, and better at presenting it. Researchers increasingly have to use other AI systems to monitor the ones they're testing, because humans simply can't keep up. And the monitors are always dumber than the thing they're monitoring.

Even if we solve the "alignment problem", the political problem remains. Say we build AI that genuinely does what we want. Who is "we" in that sentence? In practice, it's a small number of companies and government agencies. And when those entities control something that can outthink, outproduce, and outmaneuver every other institution on the planet, what does democratic governance mean? You can have elections. You can have a free press. But if the most consequential decisions are being made or shaped by systems that only a handful of actors control, those institutions become decorative.

Maybe we should build less

Investigative journalist Karen Hao makes a point in a recent interview that I keep coming back to: maybe the answer isn't "align the superintelligence" but "build less ambitious systems." Narrow AI, tools designed for specific domains, can be enormously useful without posing the same structural risks. Protein folding models advance biology. Translation tools connect people across languages. Diagnostic tools improve healthcare. None of these need to be generally intelligent to be valuable.

The argument for building general-purpose AI that can do everything is that it would be more efficient. And that's probably true. But it might also be the thing that kills us, or (more likely) the thing that concentrates power so thoroughly that the question of "us" stops mattering in any meaningful democratic sense. Maybe the less efficient path is the one where we get to keep deciding things for ourselves.

I don't have a clean ending for this

I'm not going to say shut it all down. But I think we need a serious slowdown. At minimum, building governance at the same speed we're building capabilities, which we are currently nowhere close to doing.

I started this post by describing what happened when I gave an AI my research infrastructure. It built something genuinely useful. It also made decisions I didn't ask for, accessed things I didn't think about, and operated with a kind of initiative that I found unsettling only in retrospect. I had control over it. Mostly. I think.

Now scale that up.

Fabio Votta is a Postdoctoral Researcher in Political Communication at the University of Amsterdam. He studies digital advertising, platform transparency, and increasingly worries about the thing helping him write about worrying.