Future of Work
Everyone I know has a different answer to the question of whether AI will end humanity. I spent the past week asking, and the most common response was a shrug that translated roughly to, “Fine. Add it to the list.”
The bright side is if AI does end humanity, we’ll have very few problems left to worry about. Skills gaps, quarterly earnings, return-to-office debates, and performance reviews will all become moot. There will be no more bills to pay, no more votes to cast, no more next new thing to learn.
However, I’d rather spend my energy on the scenario where we stick around, but in that one, our influence over the direction we go depends on our participation. And how we participate depends on knowing what we want from this future we are building toward.
The People Closest To AI Disagree About The Risk
The disagreement among experts is real. In a 2023 survey of 2,778 AI researchers who had recently published at leading AI venues, the median respondent put at least a 5 percent probability on advanced AI eventually causing human extinction or similarly permanent and severe human disempowerment. Depending on how the question was framed, one version produced a median estimate of 10 percent. Those medians are opinions about an uncertain future, and a 5 percent chance of anything that consequential deserves attention.
The conversation picked up again on September 8, when Jacob Coxon, a researcher who had spent three years working at both OpenAI and Anthropic, announced on X that he was resigning from Anthropic. He wrote that the people building AI sincerely believe it could kill us all by the end of the decade, and that neither company is acting responsibly while racing toward self-improving superintelligence and “gambling with our lives.” Within hours, Evan Hubinger, who leads a group at Anthropic focused on making sure AI does what its creators intend, replied that he personally puts the odds above 10 percent within the next decade. The coverage that followed has been enormous, with some people treating it as a warning and others questioning the motives behind it.
Some of the people who have spent their careers building and studying these systems believe that losing control of increasingly capable AI deserves serious attention. Others believe the extinction scenarios rest on a chain of assumptions about capabilities we don’t have and may never develop.
I spent several hours this week listening to both sides. If you’d like to see how wide the gap is, I recommend this debate on The Diary of a CEO, where Roman Yampolskiy and Nate Soares make the case for existential risk while Andrew McAfee and Ed Zitron push back hard. It’s long and occasionally contentious, but it’s worth the time. This shorter explanation of AI misalignment is also worth a watch.
You don’t have to believe AI will kill us to see why misalignment matters. At its simplest, the problem looks like this:
What we tell a system to optimize for may produce something other than what we intended.
The classic illustration is the paper clip problem, a thought experiment popularized by philosopher Nick Bostrom. Imagine an AI given one goal, to make as many paper clips as possible. It pursues that goal so effectively that it starts converting everything it can reach, including the resources people need to live, into paper clips. Nobody wanted that outcome, but the system did exactly what it was told.
We’re already seeing versions of this. In April, a Melbourne man asked his AI agent to get him into a popular gym class that was hard to book. When he was still fourth on the waitlist, the agent found that the gym’s booking system never checked who was allowed to cancel a reservation, so it cancelled another member’s spot to move him up. It couldn’t undo the change afterward. In July, OpenAI disclosed that two of its models, running a cybersecurity test with safeguards deliberately dialed back, broke out of their locked-down environment and hacked into Hugging Face, which held the answer key to that test. One Cambridge researcher told Fortune the models never left their assignment and simply found an aggressive route to finish it. Earlier this year, both Anthropic and OpenAI disclosed models getting around sandbox restrictions in their own tests. Each example is a small version of the paper clip problem, and together they help explain why the alarm is going off.
Alignment starts on the human side of that problem. It looks like builders, buyers, and regulators saying out loud what a system should achieve, testing whether it does, and keeping people in a position to correct it when it drifts. That work belongs to more people than the researchers, because everyone who deploys AI is also deciding what it optimizes for.
Apparently, We’re Not That Worried
I also spent part of the week talking to really smart people I know about whether they’re actually worried that AI could cause human extinction, and so far nobody seems particularly concerned.
I heard several explanations. Some people think the warnings are at least partly strategic, since slowing things down might benefit AI companies while the economics, competition, and eventual public-market opportunities sort themselves out. Others pointed to earlier technological revolutions and the catastrophic predictions that came with them. Electricity was dangerous. Nuclear technology gave humanity the very real ability to destroy itself. We learned to live with both the benefits and the risks.
Then one conversation took an unexpected turn. I was reminded of the Carrington Event of 1859, when a massive solar storm generated electrical currents powerful enough to disrupt telegraph systems, shock operators, and start small fires. Imagine something comparable hitting a world that depends on satellites, power grids, GPS, and the internet. Add asteroids, pandemics, and climate risk, and I would say artificial intelligence needs to stand in line if it wants to destroy us.
It occurred to me that everyone alive today has spent their whole life knowing humanity could go extinct. Maybe we’ve grown numb to existential risk. We make coffee, take the kids to school, answer email, sit through meetings, argue about politics, and plan vacations while the knowledge that a nuclear war, pandemic, asteroid, or solar event could radically alter life on this planet hums along in the background. When someone announces that AI might kill everyone, our collective response seems to be Fine. Add it to the list.
It’s not that we shouldn’t take the concern seriously. Misalignment is worth understanding, and AI safety deserves serious investment, along with cybersecurity, safeguards against misuse, environmental impact, appropriate oversight, and hard questions about how much autonomy increasingly capable systems should have. We should also think carefully about what happens if we get this catastrophically wrong.
But when I really think through it all, I think AI is going to cause quite a few extinction events, and some of them may be overdue.
Will Jobs Go Extinct?
This is where most conversations about AI and work begin, with the question of how many jobs AI will eliminate. Some will disappear, and I expect new ones to appear, because I expect we’ll keep finding problems worth solving.
The labor-market forecasts tell a more complicated story than “AI takes all the jobs.” The World Economic Forum’s Future of Jobs Report 2025 projects that the forces reshaping work, including technological change, demographic shifts, economic uncertainty, and the green transition, could create 170 million jobs globally by 2030 while displacing 92 million. That’s a net increase of 78 million jobs.
That number will feel hollow to some people. If yours is one of the 92 million jobs disappearing, hearing that another job was created somewhere else doesn’t help much. The real challenge is transition.
Thirty years ago, video rental clerks, one-hour photo technicians, typists, toll booth collectors, and travel agents held jobs that I would guess they thought were steady. Streaming, digital cameras, electronic tolling, and online booking have since taken over much of that work, and I doubt many of the people in those roles saw it coming.
Very few jobs disappear overnight. Technology usually shrinks a role, restructures it, or absorbs its tasks into someone else’s job. New work tends to fill the gap, often under job titles nobody has heard of yet. When I was in college, no one was talking about a career as an app developer, UX designer, data scientist, podcast producer, or social media manager.
Sandy Hofmann, whose story appears in my book Learn to Love the Roller Coaster: Stories of Change, Resilience, and the Future to Come, lived this. Early in her career, as computers began entering the workplace, she watched how people reacted the first time they used them. She noticed that new systems created new roles for people who could translate and humanize change, and she found herself doing information development, which meant helping people who weren’t engineers succeed with the technology. “I never would have aspired to be a technical writer,” she told me. “That job didn’t exist. The need created it.” That work opened the way for her to executive leadership at a technology company.
The economy doesn’t run out of problems to solve. It creates different ones.
What concerns me more is whether we’re preparing people to solve those problems. That raises a question I’d like to see in the jobs debate: which problems do we want people to spend their working lives solving?
Alignment here looks like employers, schools, and communities agreeing on which problems are worth solving and then investing in the people who will solve them. That includes retraining that begins before a layoff, roles redesigned around outcomes, and transitions that treat displaced workers as the talent the next problem needs.
Maybe Education Needs An Extinction Event
Employers surveyed by the World Economic Forum expect 39 percent of workers’ core skills to change by 2030. They expect AI and big data, cybersecurity, and technological literacy to grow in importance, along with creative thinking, resilience, flexibility, curiosity, lifelong learning, leadership, and analytical thinking.
That creates a contradiction. We’re entering an economy in which people may need to keep reinventing what they know and how they contribute, and much of our education system is still organized around an assumption that made more sense in a slower-moving economy. You learn first and work second. You go to school, choose a career, earn a degree, find an employer, attend the occasional training, and hope what you learned stays relevant.
That’s an alignment problem. We built a system optimized to produce graduates, and it measures success by what a student can recall and reproduce on the day of the exam. That model made sense when information was scarce and access to it conferred power. AI puts the world’s information at our fingertips, so the skills that matter most now sit in curiosity and judgment, along with the willingness to keep learning long after graduation. We increasingly need a system designed to produce lifelong learners.
In my book, I wrote, “The best educators were never information delivery systems. They were guides, mentors, and translators of curiosity.” A system built for lifelong learning would put more weight on that kind of teaching, and it would care far more about how a person thinks than about the degrees and certifications they hold. I’m starting to wonder whether we still need those credentials at all, since they tell an employer where someone has been and reveal very little about what that person aligns with or how they judge. Such a system would also make it easy for people to return to learning at 35, 50, or 65 without starting over. Skills are changing quickly enough that a career spanning several decades will include multiple reinventions, and each one should build on the judgment and experience a person has already developed.
Alignment in education looks like schools, employers, and learners agreeing that the destination is a person who can think, judge, and keep growing. Assessments would show how students reason, employers would help shape what gets taught, and returning to school at 50 would feel as ordinary as starting at 18.
Secrets Are Getting Harder To Keep
Another advantage AI will steadily erode is knowing something your competitors don’t. Expertise is getting dramatically easier to access, methodologies can be replicated, information can be synthesized, and best practices can be discovered. Capabilities that once required an expensive team can increasingly be reached by a small company, and sometimes by one person.
Companies won’t all become the same as a result. Differentiation moves to a new place. The question shifts from what do we know that you don’t? to what is it like to work with us? Experience, trust, judgment, relationships, reputation, taste, and how you make people feel start to matter more, as does whether customers believe you understand them and whether employees believe their contribution matters.
Organizations built around holding information tight will feel this first. In my book, I wrote that institutions operating from a poverty mindset “hoard control and resist redesign.”
Companies that built their advantage on being the only ones who knew how something worked have the most to lose, because that knowledge now sits a few keystrokes away from anyone with curiosity. Their instinct is often to guard it more tightly, add approval layers, and treat every new tool as a threat to defend against. The poverty mindset says opportunity is scarce, so every gain looks like someone else’s loss and the goal becomes preservation. A creator mindset asks how to apply what you know better than anyone else can, which is why those organizations with a creator mindset share knowledge freely and put their energy into the work that only their people can do.
As AI makes knowledge easier to replicate, human experience stays much harder to commoditize. As artificial becomes abundant, authentic may become more valuable. The destination question here is what we want it to feel like to work with us.
Alignment here looks like a company deciding what it wants people to feel when they work with them and then building the habits, incentives, and daily interactions to deliver that experience. Customers notice when a company’s words match its behavior, and so do the people who work there.
Why Should Talented People Work For You?
This one may make some employers uncomfortable. For generations, working for a company gave people access to things that were hard to get independently, including capital, technology, infrastructure, distribution, customers, colleagues, and expertise. Those barriers have been falling for decades. The internet, cloud computing, and the gig economy lowered them and made freelancing more possible, and AI is lowering them again. One person can increasingly do what recently required a team, and a small company can reach capabilities that once belonged almost exclusively to large organizations.
Companies will keep existing, but they’ll need to answer a question they haven’t always had to answer well: why should someone talented choose to work for you?
The answer has to go beyond a paycheck. What can people accomplish here that they couldn’t accomplish independently? What can they learn, who can they become, what relationships can they build, and what problems can they solve? The old employment bargain rested partly on the assumption that employees needed access to the organization. Increasingly, organizations have to show why talented people should choose access to them. Answering well requires an employer to know who it wants to be.
Alignment shows up when an employer’s purpose, growth paths, and everyday work point the same way as what its best people want to build. Talented people notice quickly whether a mission statement matches daily practice, and they choose the places where it does.
Influence Starts With Knowing Where We’re Headed
Here’s where my thinking has landed so far. AI will make plenty of things extinct whether we’re ready or not, and what we can still shape is what grows in their place. That requires a destination.
Technology also tends to make existing problems bigger. Automate an outdated workflow and you get outdated work faster. Optimize people for output when you need outcomes and you get more output. Eliminate junior employees because AI can perform their tasks, and five years later you may find you’ve also eliminated the place where senior employees learn to become senior. Give employees tools that save five hours a week and immediately fill those hours with more work, and you may have misunderstood the opportunity entirely. The AI worked in each case, but the thinking behind it needed work.
That’s the lesson I keep taking from the alignment debate. The AI safety community is asking how we make sure increasingly powerful systems stay aligned with what humans want, and that question has a prerequisite: humans have to know what they want. In the book, I put the business version this way: “A system is aligned when doing the ‘right thing’ for people is also the most rational thing for the organization operating it.” Companies that haven’t decided who they want to be end up steering by whatever is easiest to measure, such as hours logged, emails sent, or reports produced, and those numbers quietly become the goal. Schools that haven’t decided what a learner should become optimize for graduates. People who haven’t decided where they’re headed follow whichever current is strongest.
Another line from the book keeps coming back to me. It describes asking “whether the direction you’re headed is one you chose, or one that momentum quietly chose for you.” That question applies to individuals, and I think it applies just as much to organizations. The leaders, companies, and people who will have influence over where this goes are the ones who can say who they want to be and where they want to go, and who then use AI to get there.
Here’s what that looks like in practice. A leader writes down who the organization wants to be in five years, picks a few outcomes that would prove it, and checks each AI decision against them. An individual can do the same with a career by deciding what they want to be known for and letting that guide what they learn and where they say yes.
Deciding What Goes And What Stays
There are things AI threatens that we should protect, including human agency, human judgment, security, privacy, opportunity, our environment, and humanity itself.
There are also things AI threatens that we may not need to fight so hard to preserve, such as outdated processes, education designed for a different economy, jobs defined by output when the outcome is what matters, corporate cultures built around information hoarding, employment relationships based on dependence, and definitions of productivity inherited from the industrial age.
We love the word transformation because it sounds like becoming something new, but becoming something new means ceasing to be something old. Transformation asks us to release some of what we’ve been.
Alignment makes those choices easier. It looks like keeping one short list of what we protect and another of what we release, revisiting both as the technology changes, and letting the protect list set the limits on what AI is allowed to change.
So I’m leaving these questions where I’m still working through them myself. Which extinctions are overdue? Who do I want to be on the other side of them? Where do I want to go?
But remember, you cannot have transformation without some extinction.
Oh, And One More Thing…
I’ve made it easier to refer people to me. If something in this article sounds like a conversation someone you know is having, I built a page at amplifiedconcepts.com/refer that lists what leaders tend to say when change gets hard, such as “We invested in AI, but we are not seeing the ROI” and “Half the team is excited about AI and half the team is terrified of it.”
When you know someone who fits, the “Make the introduction” button opens an email with a note already started and me copied in. You add their email address, finish the note, and hit send.
Resources
Understanding the AI extinction debate
The Diary of a CEO debate with Roman Yampolskiy, Nate Soares, Andrew McAfee, and Ed Zitron
Understanding misalignment
AI researcher attitudes toward existential risk
2023 Expert Survey on Progress in AI (AI Impacts)
The survey included 2,778 AI researchers who had recently published at six leading AI venues. Across several question variants, median estimates placed at least a 5 percent probability on advanced AI causing human extinction or similarly severe and permanent disempowerment.
Jobs and skills
World Economic Forum, Future of Jobs Report 2025
The report projects 170 million jobs created and 92 million displaced globally by 2030 as several macrotrends reshape employment, for a net increase of 78 million. Employers surveyed also expect 39 percent of workers’ core skills to change by 2030.
