The AI Civilization Balance Sheet, Closed: Drug Discovery, Space, and the Verdict Part 5

In December 2025, a generative AI system planned a driving route across Jezero Crater on Mars, and NASA's Perseverance rover followed it without a single human correction — 140 million miles from the nearest person who could have caught a mistake. It's one of the clearest demonstrations yet of what AI can do when the stakes are real, and the distance is too great for anything else to work.

This closing chapter tests AI's case at its strongest — drug discovery and space science — and then follows the thread this series has traced from mine to model all the way to its hardest question: does the value add up, and for whom?

AI's two strongest cases: accelerating medicine on Earth, and navigating terrain no human can reach in real time. Source credit to be added on upload.

The Hardest Test: Drug Discovery's Real Numbers

If AI has one case that should silence the skeptics, it's drug discovery — and the honest numbers are genuinely impressive without needing to be exaggerated. As of mid-2026, well over 170 AI-originated drug programs are in clinical development, up from roughly two dozen in late 2023, and AI-designed molecules are clearing Phase I safety trials at 80–90%, nearly double the historical industry average of around 50%.

That's a real result, not a marketing claim: AI appears genuinely better at designing molecules with favorable safety and drug-like properties. But Phase I is the easy phase. Phase II success rates for AI-originated drugs — the stage that tests whether a drug actually works, not just whether it's safe — sit at roughly 40%, statistically indistinguishable from the traditional industry average. No AI-discovered drug has yet received full regulatory approval anywhere in the world, though the most advanced candidate, Insilico Medicine's rentosertib, has reached Phase III, with a decision expected in 2026 or 2027.

That gap between Phase I and Phase II success tells you exactly where AI's real contribution sits. It's excellent at molecular design and screening — the chemistry problem. It hasn't yet solved the biology problem: whether a molecule that's safe in humans is also effective against the disease it targets, which depends on target validation and disease mechanisms that no amount of computational power can shortcut. The honest measure of AI's value here isn't "AI cures disease" — it's the years and hundreds of millions of dollars in failed candidates it can help pharmaceutical companies avoid spending on molecules that were never going to work in the first place.

↑ Back to Table of Contents

140 Million Miles From the Nearest Correction

Space exploration has a problem no other industry shares: light itself is too slow. A round-trip radio signal between Earth and Mars takes anywhere from eight to forty-eight minutes depending on orbital position, which means a rover encountering a hazard has to make its own call, because by the time a human command arrives, the moment that mattered has already passed.

NASA's answer arrived in December 2025, when engineers at the Jet Propulsion Laboratory let a vision-capable generative AI plan an entire driving route for the Perseverance rover across Jezero Crater — a dried lakebed scientists believe once held the ingredients for microbial life. The AI analysed the same orbital imagery and terrain data human planners use, identified safe waypoints around boulders and sand traps, and Perseverance executed the route across two drives totaling 456 metres, entirely without a human correcting its path. By August 2026, the rover was completing roughly 90% of its total distance under this kind of autonomous navigation.

NASA Administrator Jared Isaacman framed the achievement in exactly the terms this series has been circling from the start: autonomous technology helping missions "operate more efficiently, respond to challenging terrain, and increase science return" as exploration pushes farther from Earth. That's not automation for its own sake — it changes what's possible. A rover that can investigate an anomaly the moment it spots one, rather than waiting nearly an hour for permission, isn't doing the same mission faster. It's doing a different kind of mission altogether.

↑ Back to Table of Contents

The Rebound Problem and the Fossil Fuel Paradox

Here's where the balance sheet gets genuinely uncomfortable. A peer-reviewed study published in npj Climate Action in August 2026 modelled AI's productivity effects across both the fossil fuel and renewable energy sectors simultaneously, using a global economic model built specifically to capture these cross-sector ripple effects. Its finding: across 64 modelled scenarios, AI's contribution to fossil fuel extraction efficiency — cheaper drilling, better geological targeting, lower operating costs — enabled more emissions than AI's contribution to renewable energy optimisation avoided.

The net effect, under parallel adoption in both sectors, was an estimated increase of 0.47 to 1.8 gigatonnes of CO₂ annually — as much as 4.8% of all global energy-related emissions in 2024. The researchers found that renewable-sector gains from AI would need to be four to five times larger than fossil-sector gains just to break even, and across their scenarios, net emissions only fell in the case where AI's fossil fuel productivity gains were zero. Separate industry estimates suggest AI-driven efficiency could expand US recoverable oil reserves by 8% to 20% — 10 to 30 billion barrels — by cutting drilling costs by roughly 30%.

The researchers' own conclusion is worth sitting with: AI doesn't inherently decarbonise or intensify emissions on its own — its productivity gains simply reinforce whatever economic structure it's deployed inside. Applied to a solar developer, it lowers the cost of clean power. Applied to an oil driller, it lowers the cost of extraction. The technology is the same. The outcome depends entirely on which industry is holding it, which is precisely the framing this whole series has been building toward.

↑ Back to Table of Contents

Who Owns the Machine

Every resource this series has traced — electricity, water, minerals, land — has an owner, and increasingly, the same small set of countries and companies own most of them. The United States and China dominate AI compute and chip design; a handful of countries, as Part 3 covered, control the refining capacity for the minerals that make any of it possible; and the communities in Part 4 absorbing land and electricity-price costs are rarely the same communities capturing the economic upside.

That concentration turns an environmental story into a geopolitical one. Whoever controls the chips, the compute and the data centres controls a growing share of scientific capability, economic productivity and — increasingly — the diplomatic leverage that comes with mineral and energy dependency. Export controls on advanced AI models are no longer a hypothetical policy tool; they're already being used, which means AI capability, unlike oil or gas, can be switched off for an entire country with a regulatory decision rather than a supply disruption.

None of this is evenly distributed, and it was never going to be. The countries and companies that own the compute, the minerals and the capital are the ones best positioned to capture AI's economic upside — which means the "AI inequality" question isn't a side issue to the environmental one. It's the same question, asked about wealth and power instead of electrons and litres.

↑ Back to Table of Contents

The AI Civilization Balance Sheet, Closed

Five chapters ago, this series opened by rejecting two easy answers: that AI is simply good, or simply bad. Nothing in the five chapters since has made either of those simpler stories true. AI is genuinely accelerating drug discovery, even if the biology problem remains unsolved. It's genuinely enabling autonomous exploration 140 million miles from home. It's also genuinely making oil and gas extraction cheaper, at a climate cost that outweighs its renewable-energy benefits in the same models. It genuinely displaces farmland and raises electricity bills in specific, named communities, while genuinely improving the grids, crops and climate forecasts those same communities depend on.

The framework this series built — AI creates against AI consumes, run through infrastructure, resources, environment, society, science, economics and geopolitics — was never going to resolve into a single number. What it resolves into instead is a sharper question than the one most people are asking. Not "is AI good or bad," but: in this specific application, in this specific location, with this specific electricity source, mineral supply chain and community bearing the cost — does the value clear the bar?

That question has to be asked sector by sector, region by region, and year by year, because the answer keeps moving. A gas turbine built to skip a grid queue looks different in a country building nuclear capacity than one that isn't. A data centre's water footprint looks different beside a Nordic river than in an Arizona drought. An AI-designed drug that never reaches Phase III still isn't wasted work, if it eliminates candidates that would have failed anyway. Civilization has built new industrial systems before — railroads, electrification, the internet — and each one arrived with exactly this kind of ledger: real costs, run up against real gains, whose final balance took decades and a lot of hard governance choices to settle. AI's ledger is being written now, faster than any of the others, and how it settles depends less on how intelligent the models get than on whether the choices made in the next several years — about where the electricity comes from, whose water is used, whose land is taken, and who captures the value — are made deliberately, or simply left to happen.

↑ Back to Table of Contents

K M Manohar’s Insight

The AI Civilization Balance Sheet should not end with a verdict that AI is good or bad. That would be too simple for a technology this powerful. My view is that AI has crossed a point where its benefits and costs can no longer be separated from the systems in which we deploy it.

The most important lesson from this five-part journey is that AI does not consume resources equally, and it does not create value equally. A drug-discovery model can potentially save years of research. An autonomous rover can expand human knowledge where direct control is impossible. But the same intelligence can make fossil-fuel extraction cheaper, intensify pressure on electricity grids, consume scarce water, and increase the value of land around data centres.

So the real question is not how intelligent AI becomes.

It is how intelligently we choose where to apply that intelligence.

For me, the next phase of the AI story should move beyond model capability and benchmark competition. We need to measure AI by the value it creates against the physical, environmental, economic and social cost required to create that value.

That means asking harder questions: What electricity powers this AI? Whose water is being consumed? Where did the minerals come from? Who bears the land and infrastructure cost? Who captures the economic benefit? And ultimately, does the application create more long-term value than the system it puts under pressure?

That is the balance sheet that matters.

AI may become one of humanity’s most powerful tools. But intelligence alone does not guarantee wisdom.

The future of AI will depend not only on building smarter machines, but on making smarter choices about what those machines are used for.