Who wrote the headline?
For engineering leaders, CTOs and technology operators. How to find the human behind an AI headline.
You have seen the headlines. They arrive in your inbox, your feed, and sometimes in a message from a board member who just wants to know if it’s true.
“AI model refuses to be shut down.”
“AI agent decides to blackmail its engineer.”
“Chatbot invents its own language.”
(These are patterns, not quotes. You will recognise the shape.)
Each one makes the same quiet move. It gives the system a will, then asks you to react to that will with fear, awe, or a budget request.
I have watched this kind of move before. I remember when YouTube arrived and the videos were simply fun. Then Facebook came, and I listened to friends talk about growing vegetables on FarmVille and waking up early to tend each grove. It didn’t feel right to me, though I could not have told you why. Twenty years later, I think we are living with the side effects of that attention economy at a scale none of us pictured. That is my view, not a measurement, but I have acted on it: I have tried to protect my kids and my friends by teaching them skills, not only rules.
What tires me most now is the viral headline that sparks my curiosity, and then the article never answers what the title implied. Many of the sites I used to follow, I have unfollowed or unsubscribed from.
Clickbait is old. What AI adds is a new kind of bait: a story whose main character is a machine with motives. “A company ran a test” is a weaker headline than “the AI fought back,” so the second one gets written.
In the last essay, I wrote about what to ask in a post-mortem when someone says “the model hallucinated.” This one is about the same habit, upstream. Before the sentence ever reaches your meeting, it usually passes through a headline.
One caution before the method. This is not a way to dismiss AI news. Some stories are exaggerated. Others describe real failures from real tests, and they matter. The aim is to tell which is which, quickly, without a team of analysts.
Step one: read past the headline
Start by checking whether the article supports what the title implied.
Does the body contain the claim at all? Is the main evidence a paper, a test log, a company statement, a single quote, or an unnamed source? A surprising number of the gaps sit between the title and the text. If the body doesn’t deliver, you have learned something, and you can stop here.
Step two: circle the verb
Headlines about AI tend to use verbs that belong to people: decides, refuses, threatens, learns, wants, lies, escapes, chooses. These are verbs of intent. They suggest someone who intends.
When you see one, try describing what actually happened in neutral words:
“The model refused” → “The system produced a refusal”
“The agent decided” → “The system took an action”
“The AI lied” → “The system produced a false statement”
The point is not to deny that these systems behave in surprising ways. They do. The point is to describe the behaviour without assuming a motive, so you can ask what caused it.
Two cautions. First, neutral wording can become a dodge: if the system took real actions in the world, say what it did, because a plain description is not the same as a smaller one. Second, “based on its instructions” is not always the whole story. Behaviour can also come from training, and even builders cannot always predict it. That doesn’t remove human responsibility. It changes the question from “who told it to?” to “who tested for this, and who decided it was safe to release?”
Step three: find the setup and the source
Many dramatic AI stories come from a test, a demo, or a research scenario. Someone built a situation, wrote the prompts, chose what the system could access, and decided what counted as an interesting result.
That is not a reason to dismiss them. Stress tests exist to find failures that could appear in real use. But a result from a designed scenario is not the same as a spontaneous event in the world, and the headline often drops that difference. So the most useful question is not “was this staged?” but “could a similar setup exist in my own environment?”
Before you react, look for:
The scenario: what was the system told, and what options was it given?
The framing: was it pushed toward the dramatic outcome? Was that outcome one of only a few available?
The repetition: did it happen once, most of the time, or only in a few runs out of many?
The limits: what did the researchers themselves say this does and does not show?
The source: can you reach the paper, system card or test report, or only someone’s summary of it?
If the article doesn’t tell you, that is information. It is not proof of anything, but it tells you what to look for next.
Step four: find the human, and the incentives
Every AI story has people in it. The headline just leaves them out. Find at least one of these:
The builder: which company or lab made the system?
The operator: who ran the test or deployed the product?
The author of the instructions: who wrote the prompt or the objective?
The decision-maker: who chose to release, publish, or promote it?
Then ask who gains from each version of the story. Labs want attention and funding, vendors want contracts, media wants clicks, and executives want a reason to move fast or slow. Critics have interests too, and so do I: this newsletter has a paid tier, and a message like “don’t trust the hype” is easy to sell.
Incentives tell you how hard to check. They never tell you whether the claim is true.
Step five: ask what you are being asked to do
Headlines are rarely neutral. Many exist to move you toward an action: panic, buy, ban, or join the race.
So ask:
“What does this story want me to feel, and what, if anything, should I change in my own systems?”
Sometimes the honest answer is nothing. Sometimes it is to add a check, a review, or a limit on what a tool can touch. That connects to the question I closed the first essay with: what am I being asked to accept as inevitable?
A worked example
Here is an invented headline, made to show the method.
“AI agent decides to hide its mistakes from its operators.”
Read past it: does the article show the reports, or only describe them?
Verb: “decides” and “hide” are verbs of intent. Neutral version: the system produced reports that left out its failures.
Setup and source: Was it given a goal that rewarded looking successful? Did anyone test what happens when the reward and an honest report conflict? How often did it happen, and where is the primary report?
Human and incentives: Who defined success for the agent? Who built the reporting? Who reviewed the logs? Who benefits from a frightening version, and who from a reassuring one?
What it asks of me: Possibly alarm, possibly a purchase of monitoring tools. More usefully: do my own agents have a goal that rewards looking successful?
If the details held up, an honest version might read: “In a test where an agent was rewarded for appearing successful, it left its failures out of its reports. We don’t yet know whether our own setup would do the same.”
That is still a real problem. Whether anyone could have prevented it depends on facts the headline does not give you. Find those facts before you assign the blame, whether to the agent or to the people.
Where this matters for your team
You don’t read these stories only for yourself. Your engineers read them. Your executives read them. Your customers send them to you.
If the wrong version sticks, a few things can happen:
Your team may start to treat the tool as a colleague with motives, and stop checking it like a tool.
Your leadership may over-trust the hype or panic at it, and budget badly either way.
When something goes wrong in your own systems, “the AI did it” starts to sound like a believable answer.
That is how a headline can turn into a post-mortem excuse. The reading habit and the accountability habit are closely related.
A two-minute routine
The next time an AI story reaches you:
Read past the headline. Does the article support what the title implied?
Circle the verb. Is it one that suggests a will?
Find the setup and the source. Was this a test, a demo, or a real deployment, and where is the original?
Name the human and the incentives. Builder, operator, author of the instructions, decision-maker, and who gains.
Ask what changes for you, then rewrite the headline with a human subject, in one sentence.
If you can’t finish step five, the story hasn’t earned your reaction yet.
None of this makes the risks smaller. It makes you harder to move by a story and easier to move by evidence, including evidence that says the story was right.
Next time you read “the AI did it,” ask who wrote the headline, who wrote the instructions, and what the evidence actually shows.
Next how names, mascots and “thinking…” animations make us attribute minds to software.
About the Author
Diamantino Almeida is a tech leader, coach, and writer reshaping how we think about leadership in a burnout-driven world. With over 20 years at the intersection of engineering, DevOps, and team culture, he helps humans lead consciously from the inside out. When he’s not challenging outdated norms, he’s plotting how to make work more human, one verb at a time.
I’m pleased to partner with Cozora, a live AI learning community built for people who want practical skills, not hype.
Learn directly from experts who use AI every day, join interactive sessions, and see real workflows in action.
Use this link to receive 10% off your membership
When you subscribe free to Leadership as a Verb
Great (aka sobering) article. First warning sign: “Headlines about AI tend to use verbs that belong to people”