AI Grows Up: From Hallucinations to Agents

Roman Z., VC Market Analyst at Raison

Four years ago, ChatGPT launched to public access. Back then, everyone had two main questions about AI: why does it get things so wrong — and will it replace us? In 2026, we know how to answer both.

Introduction

Over the past few years, humanity has been living through a technological revolution: artificial intelligence has entered our lives. And like many technologies before it, AI was met with mixed reactions.

In 2022, ChatGPT became publicly available. Even the first version could answer questions, write code and text, and assist with research. All of this seems unremarkable now — neural networks handle these tasks at a high level. But in 2022, its competence gave plenty of reasons for skepticism.

Illustration: ChatGPT (OpenAI)

The media of that era illustrates the ambivalence of the moment well. The New York Times and The Guardian were writing about roughly the same thing: a machine that can hold a human-like conversation is impressive. But that same machine can deliver confident, utter nonsense.

ChatGPT handled simple tasks well, but struggled badly with complex ones. This was most visible in programming: the neural network generated so many incorrect solutions that Stack Overflow, the world's leading IT forum, banned posting AI-written answers. Even so, the potential was obvious to anyone paying attention.

The NYT wrote at the time that the possible societal consequences of ChatGPT were too vast to fit in a single column. Some said it was the beginning of the end for all knowledge-work professions. Others said the whole thing would be forgotten within a year.

Four years later. Both camps were wrong, and both were right. Let's unpack that.

Chapter 1. AI Delivers

The central obstacle for neural networks has always been output quality. Early versions suffered from hallucinations — a phenomenon where the model generates confidently worded but factually incorrect text.

The technology didn't stand still. New training methods, reasoning chains, cleaner datasets, feedback loops with users and other neural networks — all of this drove enormous progress. A public hallucination leaderboard shows the gap between model generations in stark terms. GPT-o3-pro — the flagship of its time — produced around 23% hallucinations. ChatGPT 5.4-nano, a smaller model from the current generation, sits at 3.4%. Where one in four answers used to contain factual errors, it's now one in twenty-nine.

Simultaneously, another problem was being solved: interaction with the outside world. In 2022, there was only the chat window. If you needed to query a database, you had to know SQL yourself, set up your own database connection, understand the schema, and explain the task to the neural network precisely enough. The time saved by having AI write the code could easily be negligible.

Illustration: Claude (Anthropic)

Now we've entered the era of agentic AI. Protocols like MCP (Model Context Protocol — a standard that lets AI agents connect to external services, databases, and tools without human intervention) have emerged, and companies have restructured their processes and software around integration capabilities. The result: a business analyst doesn't even need to know what's running under the hood. They simply formulate a request — and the agent pulls the data, runs a preliminary analysis, builds charts, and assembles a presentation.

This is most dramatically visible in software engineering. AI can now independently analyze, edit, and write code across entire projects, automatically creating pull requests on GitHub. This has generated an unexpected problem: there are now so many of these requests that open-source developers physically can't review them all. Jeff Geerling, a well-known open-source developer and the author of one of the popular YouTube channels on infrastructure and DevOps, writes about this in detail. A significant portion of the AI-generated pull request flood is slop: programming has become accessible to people who don't know how to properly define a task.

That said, the economic consequences of hallucinations remain real. Fourdots tallies the losses from AI fact-checking alone — and the numbers aren't trivial. Debugging errors in AI-generated code still costs hours that weren't supposed to exist.

AI still needs a human to define the problem. But once a quality prompt is written, you can simply watch as the agent does in minutes what used to take an engineer hours. Programmer-blogger Teo.gg breaks down exactly how software engineering has changed — we recommend starting at minute 20.

Programming has become accessible to people who don't know how to properly define a task.

One story stands apart: security. Anthropic announced Claude Mythos, a model specializing in code and cybersecurity — but not its public release. They announced that it would be delayed. Through Project Glasswing, it became clear that the model is so effective at finding and exploiting vulnerabilities that Anthropic decided to first introduce it to a small group of trusted companies: they need time to secure their infrastructure before the same techniques are picked up by hackers.

And Stack Overflow, by the way, is now on the verge of shutting down. It was displaced by neural networks.

Chapter 2. We Won't Be Replaced

So AI is becoming more accurate and better at interacting with the outside world. When does it start taking jobs?

MIT recently conducted a large-scale study on what share of specialists AI is capable of replacing. The task turned out to be complex: jobs vary enormously. If you count only the positions AI can replace entirely, the picture looks surprisingly modest — around 2% of the US labor market.

The researchers went further. Using the official Bureau of Labor Statistics registry, they broke down job listings by skill, accounting for the complexity and demand of each. They then compiled a list of skills AI can already perform and overlaid the two datasets.

The result was named the Iceberg Index. Fully AI-replaceable jobs are just the tip of the iceberg. Counting partial replacement, up to 11% of the labor market falls under AI influence. The paper's preprint is available on ResearchGate, and Economics Explained breaks it down in accessible terms.

It's worth understanding what this kind of research is actually for. It's not an attempt to scare anyone about unemployment. Everyone in the market needs to understand where and how AI is moving. Future graduates, so they can choose their courses with intention. Executives — not to race toward layoffs, but to implement AI thoughtfully where it genuinely strengthens a team. AI companies themselves — to understand what kinds of tasks to prepare their models for.

Everyone in the market needs to understand where and how AI is moving

The study's conclusion matches something that's been in the air for a while now: AI won't replace us; it will more likely augment us. The list of AI skills will keep expanding, but job losses won't follow in the same proportion. A much larger share of work is simply well complemented by neural networks: they free employees from routine and let them accomplish more in the same amount of time.

The Technology Advances, Just Not Toward Us

At the beginning of this piece, I outlined two theses that dominated in 2022: "AI can't handle the tasks, though the potential is there" and "AI will soon replace white-collar workers, unemployment awaits." Both sound pessimistic. But looking at them together, in the context of 2026 — they actually give reason for optimism.

AI won't replace us; it will more likely augment us

Yes, AI has changed the world, in far more ways than can fit in a single article. Yes, it affects the labor market. But not so much by replacing people as by supplementing them. This still shows up in hiring: experienced employees now handle larger workloads, and the demand for new hires decreases. But AI is far from the only reason for that — and younger people, as a rule, tend to adapt to new technologies more easily.

Yes, AI is still imperfect. The most critical decisions can't be entrusted to it. But its capabilities will grow aggressively — in both the encouraging and the unsettling directions.

Image: Gemini (Google)

The ambivalence in the discourse around AI hasn't gone away. It's just grown up. We no longer joke about neural networks talking nonsense. Now they exert pressure on labor markets and stock prices. One person with a neural network can now meaningfully solve more important tasks in a week than ever before — and this is only the beginning.

The illustrations were generated by three different AI models using the same prompt, to see how each one interprets the same brief.

This article was written with the help of Gemini and Claude.