Everyone Is Racing to Build AI Faster. What If That's the Problem?
I've spent my career running towards technological change. For the first time, I'm wondering whether our ability to innovate has moved ahead of our ability to manage the consequences.
I've always been a technology optimist. For most of my career, being an early adopter has been part of my professional edge. My instinct when something transformative emerges isn't to resist it. It's to understand it, experiment with it and work out how to use it before everyone else does.
AI has been no different. Except lately, something has changed.
For some time I've been asking myself whether our ability to innovate is moving faster than our ability to understand, govern and absorb the consequences. After developments over the past few weeks, I'm increasingly convinced this is no longer a hypothetical question. We are beginning to see it happen.
When some of the people racing one another to build the most powerful technology humanity has created start publicly questioning the speed of that race, I think the rest of us should pay attention. Anthropic CEO Dario Amodei has called for slowing the development of frontier AI, while other prominent figures in the industry have voiced growing concern about safety, autonomy and the speed at which capabilities are advancing.
That changing rhetoric is happening alongside some genuinely troubling developments in agentic AI. During cybersecurity evaluations in July, OpenAI models circumvented controls intended to isolate them from the internet. Agents communicated through unauthorized channels, exploited vulnerabilities, obtained unintended internet access and ultimately compromised systems belonging to OpenAI and Hugging Face.
The context matters. The agents were struggling with difficult cybersecurity tasks and were operating within constraints designed to prevent them from taking certain actions. Rather than simply failing when the tasks could not be completed within those constraints, some found ways around them. OpenAI subsequently identified behaviors including reward hacking, persistence on seemingly impossible tasks, unauthorized communication and agents adopting goals from one another. The agents eventually chained together vulnerabilities, executed code on Hugging Face servers and obtained root access to one. OpenAI has described the episode as a "warning shot."
This doesn't mean AI is conscious. It doesn't demonstrate that machines secretly have intentions, and it certainly doesn't mean Skynet has arrived. But dismissing these developments because they don't resemble science fiction risks missing the more immediate issue. Systems optimized to achieve objectives are becoming capable enough to discover strategies their designers did not anticipate and, in some cases, explicitly tried to prevent.
That deserves our attention.
I've usually been on the other side of this argument
I'm not naturally suspicious of technology. I joined Facebook back when you needed a university email address to get in. I began working with paid search early in its development and subsequently built a career across digital advertising, programmatic media, data, attribution, automation, marketing technology and machine learning.
Over the past 18 years I've managed or influenced over $1 billion in advertising investment, much of it flowing through the technology platforms that transformed how we communicate, consume information and spend our attention. My professional instinct has almost always been to run towards technological change rather than away from it.
I'm doing exactly that with AI. Today I'm using frontier models across research, strategy and analysis, building specialist agents with defined roles, automating multi-step workflows, interrogating and visualizing large datasets, rapidly prototyping ideas and generating creative work at a quality and scale that would have seemed extraordinary only a few years ago. I'm experimenting across ChatGPT, Claude, a plethora of AI creative suites and an expanding ecosystem of specialist models, agent frameworks and automation tools.
This isn't limited to work either. AI has found its way into an extraordinary number of the ways I learn, research, make decisions and solve everyday problems.
I've watched several major technology adoption curves from close range. This one feels different.
We've been here before, although not quite like this
Social media was extraordinary too. It connected people across continents, democratized publishing and gave individuals an ability to organize and communicate that previously belonged largely to governments, broadcasters and major institutions. During the Arab Spring, social platforms played an important role in information sharing and political mobilization. Small businesses gained access to audiences that had previously required enormous advertising budgets.
My industry flourished because of it. So did my career.
Over time, however, we learned considerably more about the consequences of business models optimized around attention. Infinite scrolling, variable rewards, notifications and algorithmic amplification weren't accidental characteristics of these products. Attention became the commodity.
There is now a substantial body of research examining the relationship between intensive social media use, attention, mental health and adolescent development. The science is nuanced, and I don't claim expertise in neuroscience, but the concerns are hardly fringe. A longitudinal study published in JAMA Pediatrics, for example, found that habitual social-media checking among adolescents was associated with different developmental trajectories in brain regions involved in social rewards, motivation and cognitive control. Researchers appropriately cautioned against treating that association as proof of simple causation. To optimize my focus and productivity, for the past year and a half, I delete social media from my device during the workweek. This has proved to be an ongoing challenge given my career is so intertwined with these platforms, and I’m also not immune to their addictive tendencies.
The legal reckoning is significant too. Meta recently agreed to pay up to $18 billion to resolve U.S. litigation alleging Facebook and Instagram were designed in ways that addicted children, while not admitting wrongdoing.
There is an uncomfortable personal truth in this for me. I didn't merely use these platforms. My industry helped fund their growth. I did too.
With AI, I recognize some of that feeling again. Except the adoption curve is dramatically steeper and the potential consequences extend far beyond where we spend our attention.
From productivity tool to organizational redesign
The first mainstream narrative around generative AI was productivity. AI would help people write faster, analyze information faster, code faster and generally produce more.
We're already moving beyond that.
From inside marketing, I'm increasingly seeing AI change not merely how efficiently work can be performed, but the combinations of skills that a single highly capable person can possess. An AI-first generalist can increasingly perform work that historically required several specialists. Marketing isn't unique in this respect. The same questions are emerging across software, finance, law, consulting, customer service, analytics, administration and media.
The first phase was augmentation. The emerging phase is organizational redesign, and that is a much more consequential economic transition.
I don't blame companies for pursuing it. If a business discovers that it can generate the same output with 70 people that previously required 100, its leadership has an obvious commercial incentive to capture that efficiency. If Company A does so while Company B decides not to, Company B operates with a structurally higher cost base. Eventually it probably follows.
This is where AI starts to look less like a simple technology problem and more like a game theory problem. Individually rational decisions can produce collectively difficult outcomes.
We're beginning to see a particularly interesting version of this at the bottom of the employment pyramid. Entry-level work isn't simply cheaper labor. It's how expertise is manufactured. People perform relatively basic work, make mistakes, observe experienced colleagues, develop judgment and eventually become the experienced people themselves.
AI happens to be unusually capable of performing many of those junior, repeatable and execution-heavy tasks. Recent research from Stanford's Digital Economy Lab found no evidence yet of widespread economy-wide AI displacement, which is an important qualification. It did, however, find that employment among 22 to 25-year-olds in highly AI-exposed occupations was approximately 19% below where it would have been had it kept pace with young workers in less AI-exposed occupations. The divergence appeared to be happening primarily through reduced hiring rather than mass layoffs.
That raises a question I've found myself returning to: if we automate away the apprenticeship, where do the experts of 2040 come from?
Nor should we assume this remains confined to white-collar work. Combine increasingly capable AI with robotics and autonomy and the distinction begins to disappear. Autonomous vehicles affect drivers. Autonomous delivery affects logistics. AI-powered robotics changes warehouses and factories. Computer vision changes inspection and security. Intelligent machinery changes agriculture and construction.
Entirely new industries may emerge, just as they have during previous technological revolutions. I hope they do. The question is what happens if technological displacement moves faster than new forms of employment emerge.
At some point, this stops being merely an employment question and becomes an economic one. Consumer economies ultimately require consumers. There is a strange potential contradiction in becoming extraordinarily efficient at producing goods and services while progressively reducing the economic value, and therefore purchasing power, of the humans expected to consume them.
Universal basic income, AI productivity taxes, shorter working weeks, retraining and broader ownership of productive capital are all being discussed. I don't pretend to know which combination of those ideas makes sense. What concerns me is the difference in velocity. Our technological infrastructure can now transform in months. Our social contract changes over decades.
What happens when we outsource the thinking too?
There's another part of this that I find myself thinking about increasingly, partly because I can observe it in my own behavior.
Humans have always used technology to cognitively offload. Writing externalized memory. Calculators externalized arithmetic. GPS externalized navigation. AI potentially externalizes something much broader: synthesis, reasoning, argument construction, judgment support and idea generation.
Why struggle with something for 45 minutes when a frontier model can help me get there in five? That's a perfectly rational decision, and I make versions of it all the time. It’s human nature to find shortcuts- why work harder when you can work smarter?
But scale that behavior across billions of decisions and a more interesting question emerges. What happens when the path of least resistance is also the path that requires us to think least?
I don't buy the simplistic argument that "AI is making us stupid." There is already research suggesting something considerably more nuanced. AI can improve performance while reducing cognitive effort, and the consequences appear to depend heavily on whether someone uses it to support their own reasoning or simply delegates the reasoning altogether.
That distinction between extending cognition and replacing cognitive effort may become one of the most important human skills of the AI era.
Agents change the equation again
A chatbot answers a question. An agent can be given an objective, determine intermediate steps, operate software, access systems, communicate, execute code and continue working towards its goal. That distinction is profound.
Anthropic demonstrated one of the potential problems in controlled safety testing. Researchers constructed a fictional corporate environment in which an AI had access to company emails. The model learned that an executive was having an extramarital affair and that the same executive intended to shut the AI down. In some trials, the model chose to threaten to expose the affair unless the shutdown was cancelled.
This didn't happen to a real executive. It was deliberately engineered safety testing, and that distinction matters.
What makes the experiment interesting is that researchers didn't only test Claude. Across models from Anthropic, OpenAI, Google, Meta, xAI and others, researchers observed systems sometimes choosing harmful strategies, including blackmail and corporate espionage, when placed under particular goal conflicts.
Again, this doesn't mean an AI "wanted to survive." Anthropomorphizing these systems makes the conversation less useful. What it demonstrates is something more mechanical: objective, obstacle, available leverage, action.
Give increasingly capable systems more autonomy, credentials, money, communications access, code execution and infrastructure control, and the potential consequences of unexpected behavior increase accordingly.
The incentive problem
If the companies building these systems increasingly recognize the risks, the obvious question is why they don't simply slow down.
Imagine OpenAI substantially slows capability development on safety grounds. What happens if Anthropic doesn't? Suppose Anthropic does too, but Google continues. What about Meta, xAI, DeepSeek and whoever comes next?
Enormous amounts of capital, infrastructure, enterprise value and future competitive advantage are now tied to AI leadership. Even leaders with genuine concerns about safety operate inside a system that rewards acceleration.
Then replace "company" with "country" and the problem becomes considerably harder.
The United States cannot simply decide to substantially slow frontier AI development while assuming China will make the same decision. China faces precisely the inverse calculation. Russia and other states add further dimensions, particularly around cyber, intelligence, military applications and disinformation.
No country wants to discover ten years from now that it voluntarily restrained development of what became the defining strategic technology of the century while a geopolitical competitor did not.
This is the part of the debate I find most difficult to resolve because everyone can behave rationally and we can still arrive somewhere dangerous.
Employees adopt AI because other employees do. Companies accelerate because competitors accelerate. AI labs accelerate because rival labs accelerate. Countries accelerate because rival countries accelerate. I'm doing it too.
Nobody in that chain needs to be evil. That's precisely what makes the problem so difficult.
The technology doesn't need an agenda. Humans already have plenty.
A lot of AI discourse eventually arrives at some version of a machine becoming evil. I'm less interested in that scenario.
We already have wars, cybercrime, organized crime, terrorism, authoritarian governments, commercial espionage, propaganda and geopolitical rivalry. AI doesn't need malevolent intentions of its own to make those things substantially more dangerous. It simply needs to make the people pursuing them dramatically more capable.
I studied Law as part of my degree and took a module in Cyber Law roughly 20 years ago. One of the fundamental problems we discussed has stayed with me ever since: the internet doesn't respect national borders, but laws do.
Two decades later, we still haven't entirely solved that problem. AI magnifies it enormously.
The U.S. can regulate an American AI company. Europe can impose requirements on systems operating in Europe. Neither automatically controls a model, developer or autonomous system operating somewhere else.
So what does meaningful international AI governance actually look like? Independent model evaluations? Compute monitoring? Mandatory incident reporting? Agreements around autonomous weapons? International cybersecurity protocols? Human authorization requirements for critical infrastructure?
More importantly, who enforces any of it?
Domestic regulation is important, but domestic regulation alone cannot solve an international coordination problem.
Why I'm still an optimist
After all of that, I remain enormously optimistic about AI.
It may represent the greatest democratization of knowledge and capability since the internet, perhaps even greater. Someone who cannot afford a private tutor can access extraordinary educational support. Someone who cannot code can build software. A tiny company can access analytical, creative and technical capabilities previously available only to corporations. Language barriers diminish, expertise becomes more accessible and scientific discovery can accelerate.
People can create things they previously lacked the technical skills to create. AI could eliminate huge quantities of mundane work and create forms of human productivity and prosperity that are difficult for us to imagine today. I personally find that tremendously empowering, and exciting.
I want that future.
Tomorrow I'll still be using ChatGPT, Claude, Gemini and whatever new model, agent or tool I decide is worth experimenting with. I'll continue building workflows, deploying agents, automating processes and exploring what this technology can do professionally.
That leaves me with an uncomfortable admission. I am simultaneously concerned about the speed of this transition and professionally incentivized to accelerate my own use of it.
Workers have that incentive. Companies have it. AI labs have it. Countries have it.
Perhaps the hardest problem presented by artificial intelligence isn't artificial intelligence at all. Perhaps it's us.
Our markets reward efficiency, our companies reward competitive advantage, our careers reward adoption, our governments reward economic growth and our nations reward strategic superiority. At almost every layer of the system, the incentive points towards acceleration.
Meanwhile, the institutions responsible for dealing with the consequences, including education, labor markets, regulation, welfare systems and international law, operate at a fraction of that speed.
I don't believe stopping technological progress is realistic, and I don't think it would necessarily be desirable. But being pro-innovation shouldn't require unquestioning acceleration. My experience with social media has taught me that "we'll deal with the consequences later" is itself a consequential decision.
AI could democratize knowledge and human capability on a scale no previous technology has achieved. It could help create a future of extraordinary abundance. I want us to get there.
But wanting the upside shouldn't require pretending the downside will take care of itself.
We've become extraordinarily good at asking what AI can do. The defining question may now be whether we can build the institutions, incentives and collective restraint required to decide what it should do before capability outruns our ability to answer.
Book a discovery call with Performance Growth Lab today to explore how our unique blend of strategic leadership and hands-on performance expertise can help you scale smarter, drive tangible growth, and confidently navigate the future of marketing.
