AI is changing the work. The apocalypse is a separate claim

This is an opinion piece. Facts carry a number and are listed in the sources below; my opinions are marked as mine. So you know who is writing: we earn our money setting up AI in SMEs, so we have a story to sell too. Check mine as well.

A forecast you can check

On 10 March 2025, Dario Amodei, CEO of the AI provider (An AI provider runs language models and sells access to them, for example OpenAI (ChatGPT), Anthropic (Claude) or Google (Gemini).) Anthropic, said that in three to six months AI would be writing 90 percent of the code¹. That is a good forecast. Not because it is right, but because you can check it: it names a quantity, a share and a deadline.

What it does not say is just as interesting. Amodei talked about code, not programmers. In the same talk he said a person still has to specify what gets built¹. In October 2025, after the deadline, he said it was true within Anthropic and some partner firms, “not uniformly”. An analyst at Redwood Research estimated the share at Anthropic at about half². Partly true, partly not, and in a company that sells exactly these tools.

Other forecasts reach further. In May 2025, the same Amodei said that within one to five years AI could wipe out half of all entry-level white-collar jobs and push US unemployment to 10 to 20 percent³. In February 2026, Mustafa Suleyman, CEO of Microsoft AI, said most tasks of white-collar work at a computer would be fully automated within 12 to 18 months⁴. Amodei said “could”; Suleyman spoke of tasks, not jobs. Headlines often turn that into certainties about jobs.

Before I wrote this, I had a number in my head myself: researchers agree that work becomes obsolete in about 350 days. I found no attributable source for it. No person, no date, no consensus. What I found instead is the most official dated forecast there is: in March 2025 Anthropic told the White House it expected powerful AI systems “in late 2026 or early 2027”, with capabilities at the level of Nobel Prize winners across most disciplines³³. Its formal letter of the same day is more cautious: “as soon as late 2026 or 2027”³⁴. The counter runs to the end of that window as I read it: 31 March 2027. It is Anthropic’s expectation, not mine, and not a consensus. And Anthropic sells exactly such systems. If you would rather see another forecast, the calculator at the end lists seven more, each with its date.

What AI can do today

That capabilities are growing fast is well documented. The research institute METR measures the length of tasks AI systems complete on their own. In March 2025, that length was doubling about every seven months⁵. In January 2026, METR measured a doubling about every 89 to 109 days for the period since 2024⁶. In the 2025 final of a competition run by the US agency DARPA, AI systems found 54 deliberately planted vulnerabilities in 63 challenges and fixed 43 of them⁷.

But a measurable ability is not finished work. In OpenAI’s GDPval test, experts rated the output of Claude (An AI assistant is a chat program such as ChatGPT, Claude, Gemini or Perplexity: you write in plain language, and a language model answers.) Opus 4.1, the best model in the test, as good as or better than human work in 47.6 percent of cases⁸. But the tasks were precisely specified and done in one pass, with no customer changing their mind halfway.

How easily we fool ourselves shows in an experiment by METR. In early 2025, experienced open-source developers took 19 percent longer on their tasks with AI tools. Afterwards, they believed they had been 20 percent faster⁹. A follow-up study from February 2026 points rather to a time saving, but METR calls it only very weak evidence¹⁰. Felt and measured time are two different numbers. My guess why: AI talks like a person, so it seems more capable than it is.

What reaches the labour market

The forecasts are loud, the data quieter. In October 2025, the Budget Lab at Yale University found no discernible upheaval in the overall US labour market since ChatGPT was released¹¹. In May 2026, it still saw no clear sign of AI effects¹². A Stanford study finds one group where something moves: employment of 22 to 25 year olds in the occupations most exposed to AI is 19 percent below where it would be had it kept pace with less exposed peers; experienced workers show no such gap¹³.

And in Switzerland? According to the KOF Swiss Economic Institute at ETH Zurich, the number of job seekers in occupations highly exposed to AI has risen up to 30 percent more than in less exposed occupations. AI has probably contributed. In 2025 that meant about 7,400 additional job seekers in those occupations¹⁴. In a UBS survey of about 2,500 firms, six in ten use AI, and more than half expect their headcount to stay the same¹⁵.

My reading: both are true. The tools improve fast, and labour markets respond more slowly than the loudest voices say. Slower does not mean not at all. Arvind Narayanan and Sayash Kapoor describe AI as “normal technology”, like electricity or the internet: far-reaching, but adopted over decades, not years¹⁶. That a machine can do a task does not mean people become jobless. How much we work, and what for, also depends on how a society organises itself. That is my conviction, not a forecast.

Who benefits from which story

Motives are invisible, so I accuse no one; I record what is public.

In March 2023, tens of thousands signed an open letter calling for at least a six-month pause in training AI systems more powerful than GPT-4¹⁷. No pause happened. In May 2023, the CEOs of OpenAI, Anthropic and Google DeepMind signed a statement that mitigating the risk of extinction from AI should be a global priority, alongside pandemics and nuclear war¹⁸. Andrew Ng, co-founder of Google Brain, pushed back in October 2023: large tech companies that would rather not compete with open source were creating fear of AI leading to human extinction¹⁹. Yann LeCun, then Meta’s chief AI scientist, called the doom scenario “just preposterous”²⁰.

Meanwhile, the companies whose leaders publish bold timelines raise huge sums. In February 2026, Anthropic announced a funding round of 30 billion dollars at a valuation of 380 billion²¹. Valuations like that price in expectations about the future.

My opinion, no more: part of the doom rhetoric is marketing. Whoever says their technology could end the world also says it is the most powerful one. While politics debates the end of the world, nearer questions wait: security, data protection, liability. I also see the race with China as a fight for market share rather than world domination. For an SME, a smaller question is more useful: which work in your business is slow, error-prone or simply dull today?

What the bomb teaches, and where the comparison limps

My opinion, with a lopsided comparison: nuclear weapons have existed since 1945, and humanity is still here. They do damage when someone builds them, handles them carelessly or uses them on purpose. With AI, too, the risk lies less in the technology than in who uses it, for what and how carefully. But the comparison limps: a nuclear weapon needs uranium and a state, an AI model (A language model is the actual AI behind an AI assistant: it has learned from a great deal of text and writes an answer to an input.) a computer and an account. Material can be controlled; software hardly.

Security: the same ability on both sides

IT security is a concrete example of both sides. In 2024, Google’s AI agent (An AI agent carries out several steps of a defined task. A person decides before it sends anything outside your firm or spends money.) Big Sleep found a previously unknown, exploitable vulnerability in the widely used SQLite database before it appeared in an official release²². In July 2025, Google wrote that Big Sleep had found a vulnerability known only to attackers and at risk of being exploited²³. That is Google’s account, not independently confirmed.

Attackers use the same abilities. In November 2025, Anthropic reported an espionage campaign against about 30 targets in which AI carried out 80 to 90 percent of the tactical work on its own and a handful of intrusions succeeded. The same report says the AI often overstated its findings and occasionally fabricated data, such as credentials that did not work²⁴. Security experts publicly doubted the account, partly because Anthropic published no technical traces of the attack²⁵.

Daniel Stenberg, maintainer of the widely used tool curl, wrote in July 2025 that AI-generated junk made up about 20 percent of security reports²⁶. In April 2026, he wrote that almost every report now uses AI, and that the quality is higher²⁷. In June 2025, the AI system XBOW topped the US leaderboard of HackerOne, where specialists report vulnerabilities; according to the company, people reviewed every report first²⁸. My assessment: at searching for known patterns, such systems now beat many people who look after an SME’s IT on the side. In Switzerland, the Federal Office for Cyber Security notes that attackers now systematically use AI to deliver tailored, personalised and credible content to their victims²⁹.

My conclusion: neither panic nor calm, but homework. The UK’s National Cyber Security Centre writes that the time between the disclosure of a vulnerability and its exploitation has shrunk to days³⁰. It also works the other way: in the DARPA competition, a patch was ready after 45 minutes on average⁷. If you leave updates lying around for weeks, you have a problem, with or without an apocalypse. My hope: because software is getting cheaper, small firms can run their own small tools. Then less may be exposed to the internet, but not automatically: it depends on how the tools are built and maintained.

And the rules in Switzerland?

Switzerland has no AI act of its own. The Federal Council intends to ratify the Council of Europe’s AI convention and to amend the law sector by sector where possible. A draft for consultation is due by the end of 2026³¹. Already, according to the FDPIC (The FDPIC, the Federal Data Protection and Information Commissioner (EDÖB in German), supervises compliance with Swiss data protection law.), the Swiss data protection law (Switzerland’s Federal Act on Data Protection (FADP) governs how firms and authorities may process personal data. The revised act took effect on 1 September 2023.) applies directly to data processing with AI³².

What you can do this week

The apocalypse is a claim about the future. The hours your team spends retyping today are a fact. You can do something about the second:

1. Map one task. Pick a task that comes back every week. Write down what starts it, what gets copied and what ‘done’ looks like. The worksheet in ‘Where does the time go?’ helps with that.
2. Protect access and data. Switch on two-factor authentication (With two-factor authentication you log in with two proofs, such as your password and a code on your phone.) wherever it exists, keep updates current and check who has which access rights (Access rights set out who may see and change which files, programs or accounts.). Before you put customer data into an AI tool, find out where it goes.
3. Try a bounded improvement. One step, one person, a few weeks. Anything that leaves your firm or moves money is still signed off by a person.
4. Measure. Record the time beforehand and measure again afterwards with the same method. Feeling faster is not measuring faster.

The calculator below uses only your numbers. Its second view lists the forecasts from this article with their dates and sources.

Perhaps one of the loud timelines turns out right. Then a mapped, protected and measured task serves you better than fear. If not, you have still gained time.