According to social media, you should be preparing two documents: your résumé and your will.
Artificial intelligence is going to take your job. Then it will wipe out humanity. There is some debate about the order, but apparently you should have both in PDF format.
Meanwhile, someone is offering you a course on surviving all this. It comes with lifetime access.
I admire anyone who can offer that kind of guarantee in the middle of an apocalypse.
Before you buy the course or say goodbye to your loved ones, let's understand what we are building. No equations. No words that sound like prescription drugs. And, above all, no assuming that a machine that writes emails is already arranging our funeral.
First, Meet the Alleged Killer
Let's talk about the AI you find in chat apps.
Imagine reading: “The coffee was very hot, so I waited for it to…”
You would probably finish the sentence with “cool down.” Not “run for office.”
A language model learns during training to anticipate how a text might continue. It practices on enormous numbers of examples, and its internal numbers are adjusted to improve its performance. When you write to it, it uses what it has learned and the conversation's context to build a response, piece by piece. It does not need to have every answer stored in advance. (Google for Developers)
Those pieces are called tokens. You now know a technical term. Please finish the article before adding “AI consultant” to your LinkedIn profile.
Learning to produce responses can develop capabilities far more complex than completing a sentence. With additional training, models can get better at solving problems, checking results, and changing strategies. Not all of that comes from programs added on the outside. (arXiv)
But answering fluently does not guarantee answering correctly. They can invent information or make mistakes while sounding completely confident. (Google for Developers)
So far, an extraordinary technology with a flaw that will feel familiar to anyone who attends work meetings.
Knowing How to Do Something Does Not Mean Having the Keys
Suppose you ask it for a pizza recipe.
It writes one. There is still no flour in your kitchen.
Now imagine connecting the model to a grocery app and allowing it to make purchases. It could look up products, choose ingredients, and place an order.
When a system can choose steps and use tools to complete a task, we call it an agent. It can work without you specifying every click: it decides which tool to use, observes the result, and continues. (Anthropic)
The difference is not that the AI has developed an appetite.
The difference is that a program can now take actions based on its decisions.
A recipe is text. A purchase involves access to a store, an address, money, and authorization. Those are different things, even if both begin in a chat window.
So when you hear that an AI “can do anything,” it is worth asking:
With which tools, under what conditions, and within what limits?
“It's very intelligent” is no substitute for any of those answers.
“Despacito” Is Not a Launch Code
The story of “I asked Claude who sings Despacito, and thirty seconds later the Earth disappeared” skips a few steps.
Quite a few, actually.
You would need to explain how it went from answering a question to accessing dangerous systems, taking actions, and overcoming safeguards or exploiting security flaws.
There are several missing connections between the chorus and Armageddon.
That does not mean we should stop worrying. The risks include malicious uses, errors, and potential losses of control. Future systems bring uncertainties, and current safeguards have limitations. None of that requires a machine that hates us. (International AI Safety Report)
Let's return to the grocery store with an imaginary example.
The agent buys the ingredients. The confirmation is slow to arrive. It assumes the purchase failed and tries again. Then it tries once more.
You end up with enough food for a wedding you were not planning to have.
There was no evil consciousness. There was poor error handling.
Now replace “ingredients” with “supplier payments,” and you will understand why we need limits, checks, and ways to stop or correct an operation. Those measures are part of responsible risk management, not optional decorations for a presentation. (NIST AI Resource Center)
You do not need to believe in the Terminator to take safety seriously.
You also do not need to picture the Terminator every time a machine learns something new.
The Deep Fryer Is Not the Restaurant
Let's move on to the question that probably worries you more than missiles: work.
Imagine Laura, who prepares quotes for a company.
She used to spend two hours gathering information and writing each one. With a new tool, she gets a draft in fifteen minutes. Those times are made up, but the dilemma is clear: she can now produce much more.
Does that mean Laura is no longer needed?
We have not looked at the rest of her job yet.
One customer asks for something that will not meet their needs. Another needs an impossible delivery date. A third wants the cheapest product, with the best quality, delivered yesterday.
That last one is not an artificial intelligence problem. It is a commercial imagination problem.
Laura also has to clarify needs, check prices, negotiate, and stop the company from promising things it cannot deliver.
Automating the document does not prove you have automated the entire job. The International Labour Organization's 2025 assessment considers job transformation the most likely effect of generative AI, because many occupations combine tasks that still require human involvement. That does not guarantee that every position will survive. (International Labour Organization)
Some companies might hire fewer people or shrink their teams. Others might serve more customers or launch services they could not previously offer.
The reassuring conclusion is not “nobody will lose their job.”
It is this: a machine doing part of your job does not mean all your experience has stopped being useful.
Buying a deep fryer does not automatically make you a restaurant owner either.
Even if the fries look spectacular in the demo.
The Apocalypse Is Still Waiting for the Delivery Guy
Let's stick with our imaginary restaurant.
The owner buys an oven that makes ten times as many pizzas. He is delighted with the productivity.
But he keeps the same delivery guy and his bicycle.
An hour later, he has a mountain of pizzas waiting and a man pedaling as though he owes money to someone dangerous.
The kitchen is faster. The customers are not eating any sooner.
That is a bottleneck: the part that limits what the whole operation can do.
Something similar would happen if Laura prepared a hundred quotes a day, but the company could only fulfill five orders.
Delivery can improve too, of course. Bottlenecks do not protect jobs forever. But they show why an entire business does not transform just because one task becomes incredibly fast.
Reality has an annoying habit of not fitting into a demo.
Four People and Ten People's Responsibilities
Now imagine a company that can do with four AI-assisted developers what it previously did with ten.
This is a hypothetical, not a prediction for every team.
Those four might write less code by hand while still being accountable for more systems, more changes, and more decisions.
The machine produces.
The phone that rings when something breaks may still be yours.
There is a negotiating argument here: not just “I work longer hours,” but “I am delivering more results and taking on more responsibility.”
The company can save money and the professional can earn more at the same time. That is not a mathematical contradiction. It is also not something that happens automatically: it would have to be negotiated.
Your peace of mind should not come free with the enterprise subscription.
And what about the six people who left the team? They do not disappear from the story just because the numbers look good. They may need training, opportunities, and support. Firing them and recommending a tutorial is not enough.
Potential new jobs are not an automatic, immediate, or equivalent replacement for those lost, either. Economists disagree on how much new activities will offset displaced ones and on the eventual effects on employment and wages. (International AI Safety Report)
Promising that everything will sort itself out would be another way of selling hot air, this time with relaxing background music.
The Cheap Version Does Not Include Taking Responsibility
Back to Laura.
A tool that drafts quotes might be inexpensive. But letting it send them on its own requires more: current prices, accurate stock information, discount limits, and a way to handle exceptions.
In this example, the calculation would not be “Laura's salary versus the chat subscription.”
It would mean comparing the full cost of two ways of working.
Suppose the agent offers a hundred products at the wrong price. It could write an excellent apology. It might even show plenty of empathy.
The zeros in the bank account, however, do not reappear because the message ends with “we apologize for the inconvenience.”
That does not prove automation is always expensive. It proves that doing something cheaply and taking responsibility for doing it well are two different problems.
Automating some of the reviews would also be reasonable. But adding more software to review software does not remove the need to check how the whole system works and prepare responses to failures. (NIST AI Resource Center)
“Another AI approved it” might explain how an error happened.
It does not make it correct.
Do Not Bury Your Experience Yet
In our restaurant, the AI suggests new campaigns, promotions, and menus.
Meanwhile, Carmen, the waitress, notices several customers checking their watches and leaving food on their plates. She asks what is wrong.
They only have twenty minutes for lunch.
She proposes a small menu they can pre-order before arriving. Then she uses AI to develop the idea, compare options, and prepare the promotion.
The owner was looking for a revolutionary dining experience.
The customers were looking to get back to work without being reprimanded.
Could an AI have suggested something similar? Yes. This story does not prove that human creativity has supernatural powers.
It shows something simpler: Carmen discovered a need through a conversation and used a tool to address it.
Your experience can help you do more than perform a task. It can help you recognize which task is worth doing.
The machine does not make that ability useless. It can help you take it further.
Your Survival Kit
You do not need to become an AI engineer tomorrow.
Start with one small task in your profession. Try a tool with non-confidential information or sample data. Look at how much time it saves and what errors it introduces. Learn to tell the difference between a convincing answer and a result that actually works.
Do not start by handing it the company credit card and saying, “Surprise me.”
And do not throw your experience out the window because an app does something faster. Use it to frame better problems, spot mistakes, and check results.
Companies have to do their part too. Training people, designing safeguards, and assigning responsibilities should not be replaced by an email saying “we need to adapt.”
That is like throwing someone into the water and congratulating them on their new opportunity in the maritime sector.
Real changes are coming. Some will be difficult. But an uncertain future is not the same as a death sentence.
You can learn to use AI without worshipping it, and demand precautions without living in terror.
For now, leave the bunker for later. Stay curious, practice, and do not outsource all your judgment.
And you can ask who sings “Despacito.”
What you should check is why someone connected the music app to the Ministry of Defense.
