AI Risks vs. Benefits

 

AI Risks vs. Benefits: My Honest Assessment

When someone asks me whether AI is a threat or an opportunity, I've learned not to give a one-word answer. The truth is messier—and more interesting—than that.

I've spent years thinking about artificial intelligence, its trajectory, and what it means for society. I've watched the technology evolve from academic curiosity to central strategic concern. And I've noticed something: the conversation has become too polarized. 

The doomists say AI will destroy us. The utopians say it will liberate us. Both are

partly right, and both are missing something crucial.

Let me share what I actually believe.

Why I Now Say AI Risks Exceed Nuclear Threats

This might sound extreme, but hear me out. At a recent international summit I followed closely, a striking consensus emerged: 

- the near-term dangers from advanced artificial intelligence now dwarf the familiar fears we associate with nuclear weapons.

I don't say that lightly. Nuclear weapons are genuinely catastrophic. But AI is different—qualitatively different—in ways we haven't fully reckoned with.

The Four Factors That Make AI Uniquely Dangerous

Technical Speed. 

AI capabilities are improving at a pace that frankly terrifies me. Training times, model sizes, and performance gains are compounding rapidly. Recent studies show development cycles measured in months rather than years. That matters because it compresses windows for oversight. Dangerous capabilities can emerge before governance even catches up.

Diffusion and Accessibility. 

Here's the stark difference from nuclear weapons: they require rare fissile materials, specialized infrastructure, sustained programs. AI is software-native. Once an architecture or dataset is published, variants can be reproduced and adapted widely—by states, criminal networks, irresponsible firms. You can't exactly hide a nuclear program. You can hide an AI training run in a distributed compute cluster.

Dual-Use Pathways. 

The same techniques that accelerate scientific discovery and improve healthcare can be repurposed for disinformation campaigns, automated cyberattacks, biological design assistance, or destabilizing economic manipulations. This multiplies the pathways to harm in ways we're only beginning to understand.

Economic Incentives and Misaligned Competition. 

Here's what keeps me up at night: venture capital, market valuations, and national prestige all reward rapid capability deployment. This creates intense competitive pressure to prioritize performance over safety. It's an environment where corners get cut—and not in the tightly controlled way they did in the nuclear age.

Put together, these dynamics make AI a rapidly proliferating, multi-vector risk. Not a single, centralized threat like a nuclear arsenal. Instead, diffuse, cascading effects across societies and economies.

What I Heard at the Summit

I followed the conversations closely, and several realistic scenarios stood out:

In one plenary, regulators described near-miss incidents where generative systems inadvertently enabled sophisticated fraud or automated spear-phishing campaigns that bypassed standard defenses. The scale was terrifying—thousands of tailored, convincing messages per minute. Containment was nearly impossible.

A panel of cybersecurity experts walked through how AI-powered malware can adapt in real time, undermining the old patch-and-patch-again defense model. The consensus: attribution and mitigation become exponentially harder when attacks self-evolve.

Economic ministers focused on labour-market shocks from AI-driven automation, noting short windows for retraining and policy adaptation. Here's the key insight: unlike nuclear risk, which concentrates in geopolitics and deterrence, AI affects employment, civic discourse, and critical infrastructure simultaneously.

These aren't theoretical concerns. They're frontline policy problems—decentralized, fast-moving, and consequential across multiple domains.

But Here's Where I Differ From the Doomists

I'm not a pessimist about AI's future. In fact, I think we're asking the wrong question when we ask "Is AI good or bad?" It's like asking "Is fire good or bad?" Fire can warm your home or burn it down. The question is how we use it.

When it comes to jobs and productivity, I've studied the history carefully, and the evidence actually cuts against the doomist narrative.

What History Actually Teaches Us

I've heard the chorus: "AI will steal our jobs. Millions will be permanently unemployed. We're doomed."

These fears are real, and they deserve to be taken seriously. But let me offer a different perspective grounded in historical fact.

The Industrial Revolution mechanized many manual trades and transformed labour markets radically. Yes, certain jobs disappeared. But entirely new industries, urban professions, and mass employment emerged over decades. The net effect was higher living standards for most people.

Take a more recent example: ATMs. When banks introduced automated teller machines, everyone predicted the end of bank-teller employment. It was logical, right? Machines that dispense cash should eliminate tellers.

What actually happened? Banks redeployed human staff. Tellers moved into customer service, sales, branch expansion. The nature of teller work changed, but employment in banking actually rose as the banks reached more customers with lower-cost operations.

I see a pattern repeated throughout economic history: automation removes tasks. It rarely removes the need for human judgment, empathy, creativity, entrepreneurship, and the social processes that make markets function.

The Venture Capitalist Who Got It Right

I've cited venture capitalist Vinod Khosla on this. He made a point that stuck with me: "Slowing down could kick up a different set of problems."

That's not naive techno-optimism. It's an empirical observation about how economies actually work. Rapid automation increases productivity and lowers costs, which expands markets and creates demand for entirely new lines of work. History shows that if you let markets work, new roles emerge—often in ways that are complementary to machines rather than directly competitive with them.

But—and this is crucial—that only happens if we manage the transition well.

Where Optimism Without Policy Becomes Reckless

Here's the part where I diverge sharply from Silicon Valley cheerleaders: optimism without policy is hollow. Skepticism without solutions is paralyzing.

I accept Vinod Khosla's core argument that halting progress has real costs. But I add this: we must distribute the gains fairly and invest in human capital.

Treating automation as purely an economic opportunity, with no thought to fairness or resilience, is how you get widespread suffering even as productivity rises. 

Companies benefit from higher productivity; 

society must ensure those gains support transitions, not just concentrate wealth.

My Policy Prescriptions

For the Jobs Question

We need a multi-pronged response:

Retraining and lifelong learning. Invest in scalable, employer-aligned retraining programs. Short, modular credentials tied to local demand can help workers transition more quickly than broad, academic retraining.

Education reform. Move fast on curricula that build digital literacy, critical thinking, and adaptability. Not narrow vocational skills that become obsolete in two years.

Stronger safety nets. Enhance unemployment support, portable benefits, and targeted wage insurance during transitions. Give people room to retrain without falling into poverty.

Tax policy and corporate incentives. Consider tax credits for firms that invest in worker retraining or that create new roles requiring human-AI collaboration. The "robot tax" idea—levying taxes on automation gains and using proceeds for transition programs—deserves serious debate, but we must design carefully to avoid stifling innovation.

Universal Basic Income—think carefully. UBI is an important conversation, but it's not a silver bullet. A modest, well-targeted basic support could complement retraining and active labour-market policies. Large, unconditional UBI would reshape fiscal priorities and deserves pilots before full deployment.

For the Systemic Risks

The risks I outlined earlier require different medicine:

Treat AI as a systemic risk. Governments should adopt threat models that account for cascade effects across information ecosystems, critical infrastructure, and the economy—not just narrowly defined rogue systems.

Build modular international norms. Create interoperable standards for safety testing, red-teaming, and transparency that countries can adopt incrementally. Think of it as arms-control verification adapted for software and compute.

Control critical inputs. Regulate and monitor specialized compute, high-value training datasets, and supply chains for components that materially accelerate capability growth.

Incentivize safety research. Public funding and procurement should prioritize alignment, interpretability, and robust verification research. Reduce the market incentive to race on capability alone.

Strengthen incident-response capabilities. Establish rapid, cross-border channels for sharing threat intelligence, forensic tools, and coordinated mitigation strategies when emergent harms appear.

Foster public-private compacts. Since most frontier capability sits in private firms, governments must negotiate enforceable commitments on testing, disclosure, and safe deployment.

My Honest Bottom Line

I don't side with doomsayers who predict mass, permanent joblessness. But I also reject unfettered techno-utopians who promise effortless benefits for all.

AI will change what work looks like. 

History suggests new jobs will emerge even as tasks disappear. But—and this is the critical part—the real question is whether our institutions evolve fast enough to help people move from displaced tasks into meaningful, compensated roles.

I believe we can do that. 

But only if we combine the innovation that creates wealth with robust public

policy aimed at fairness and resilience. 

Only if we treat AI's economic benefits as a public responsibility, not just a private windfall.

The window to build durable safeguards is narrow. 

Governments need to act quickly on international coordination, safety standards, and incident response. 

Companies need to prioritize responsible deployment alongside innovation.

Societies need to invest in education and safety nets.

But if we do this right—if we marry technological progress with genuine stewardship—then AI can be what it promises: a tool that expands human capability rather than diminishes it.

That's what I'm working toward. That's what I believe is possible.

---

Hemen Parekh October 2026