The Machine That Builds the Machine

The chatbot is the smallest part of the story. AI is starting to improve the chips, energy systems, and science that produce the next AI.

I was in LA this weekend for my cousin’s wedding. Afterward I spent time with my best friend and his family, and the conversation turned to AI, the way it does now.

Two of his sons are recently out of college, and they had strong opinions. Not cynical ones. They were asking the questions people should be asking: what happens to jobs, whether these systems can be trusted, what to do about hallucinations, how much water and electricity this all takes, who controls the technology, and whether we’re moving faster than people, companies, schools, and governments can absorb.

I understood the concern.

I’m an optimist by nature, but optimism without facts is just make-believe, and fear without numbers isn’t much better. And the loudest voices in this debate are the worst place to get the numbers. The talking heads and media outlets have axes to grind, eyeballs to capture, and, in more than a few cases, IPOs to manage. Their incentives are not your incentives. So, I kept coming back to: what does the evidence say?

Full disclosure before I start: I’m not the technical guy. I’m an investor/advisor/author who keeps trying to learn this layer of the stack because it now touches every company I look at. So, this piece leans on primary sources: peer-reviewed papers, agency reports, trial results, and company disclosures. Where a claim is company-reported, I say so. Where the research runs deeper than I can independently judge, I’m reporting what the researchers published and where the numbers point. The precise figures all live in the references at the end, so you can check the math yourself.

I think we’re still talking about AI too narrowly. We talk about chatbots, apps, cheating, job loss, productivity. All of that matters. But underneath it there’s a bigger story, and it’s the one I want to walk through here.

AI is becoming part of the machinery that builds the next generation of technology. Energy, data centers, cooling, chips, memory, networks, software, medicine, materials, education, work. AI used to be only the thing the technology stack produced. Now it’s starting to improve the stack itself.

Electricity turns into compute, compute turns into intelligence, and that intelligence is now being pointed back at the energy systems, chips, and software it depends on. Each turn of that loop creates more capability, which feeds the next turn. That’s the shift worth understanding, and a serious conversation must cover both sides of it: the cost and the upside.

AI Is Not Floating in the Cloud

Every AI answer starts somewhere physical: a power plant, a transmission line, a transformer, a data center, a chip, a memory stack, a network, a software layer. We experience it as a prompt and a response, but underneath it’s an industrial system converting electricity into intelligence.

The International Energy Agency estimates data centers used about 1.5% of the world’s electricity in 2024, and projects that share could nearly double by 2030. [1]

The United States is further along that curve. The Department of Energy reported data centers used about 4.4% of U.S. electricity in 2023, and that share could reach somewhere between 7% and 12% by 2028. [2]

So, the energy concern is real. But “does AI use energy” is the wrong question, because of course it does. The better question is what we’re getting for that energy, how efficiently we’re using it, and whether AI can help improve the same systems it runs on.

Power Is the First Constraint

AI doesn’t scale on ambition. It scales on electricity. And power means more than generation: grid connection, transmission, substations, transformers, cooling, land, permitting, local politics, and time. A megawatt announced is not a megawatt delivered.

The IEA estimates around 20% of planned data-center projects could be delayed if grid constraints aren’t addressed. New transmission lines in advanced economies can take four to eight years to build and wait times for transformers and cables have doubled in the past three years. [3]

That reframes the race. The competition is happening at utilities, semiconductor fabs, data-center campuses, permitting boards, and water authorities as much as inside model labs. This stopped being a software story a while ago. It’s an infrastructure story.

Water Is Real, But It Isn’t One Story

Water is where the global AI story becomes local. A data center in a water-stressed region is not the same as one using recycled water, air cooling, or closed-loop liquid cooling in a cooler climate. Location, cooling design, and power source all matter.

U.S. data centers directly consumed about 17.4 billion gallons of water in 2023, and that could double or even quadruple by 2028. Google’s reported data-center water use grew by more than 40% between 2021 and 2024. [4]

That’s not trivial. But “AI uses water, therefore AI is bad” is too simple, because the engineering tradeoff is genuine. Cooling methods that save electricity tend to use more water, and methods that save water tend to use more electricity. There’s no free lunch, only design choices.

The designs are improving, though. Microsoft says more than 90% of its new Fairwater facility uses closed-loop liquid cooling: the system is filled with water once during construction and then recirculates it, with outside air handling most of the year and fresh water used only on the hottest days. Satya Nadella has compared the annual water draw to a typical restaurant. That doesn’t erase water concerns across the industry, but it shows how AI-specific designs can change the footprint. [5][6]

The honest position is that water impact is uneven, design-dependent, and local, not that the fears are overblown.

The Constraint Doesn’t Disappear. It Moves.

Technology rarely eliminates a constraint. It relocates it. Think Whack-a-Mole.

Better cooling moves the bottleneck from water to power density. Faster chips can shift it from raw compute to memory. Smarter software tends to push it from hardware toward data quality and workflow design, and orbital compute would move it again, this time from land and water to heat, launch mass, radiation, repairability, latency, debris, and regulation.

That last one sounds strange until you look at what SpaceX is now discussing. Reuters reported Elon Musk describing AI data-center satellites powered by solar energy and cooled by radiating heat into space, with a first proposed satellite carrying roughly one high-end AI rack’s worth of computing. That’s not a hyperscale data center in orbit. It’s closer to a rack-scale experiment. [7]

And SpaceX itself has warned that orbital AI compute is early stage, unproven, and may not be commercially viable. Space is harsh, launch economics are unforgiving, and thermal management, radiation, repairability, and regulation are all open problems. [8]

So, I wouldn’t frame this as “Elon’s answer is to leave Earth.” The useful read is that AI infrastructure isn’t escaping physics. It’s being redesigned around physics. Maybe some compute eventually moves to orbit and maybe most doesn’t, but the fact that serious people are looking tells you the industry is already hunting for places to put the bottleneck next.

Now the Loop Starts

AI consumes power, water, chips, memory, software, and human attention. It’s also starting to improve power systems, data centers, chips, software, science, medicine, and work. Both of those are true at the same time, and that’s the loop.

DeepMind reported back in 2016 that machine learning cut the cooling energy at a Google data center by up to 40%. The claim was never independently verified at scale, and it doesn’t erase the growth in total data-center electricity demand, but it’s a concrete case of AI improving the infrastructure AI runs on. [9]

The grid example may matter more. The IEA says AI-based fault detection can cut power outage durations by 30% to 50%, and that smart sensors paired with AI management could free up large amounts of capacity on existing transmission lines, without building a single new one. [3]

No magic in any of that, just optimization. AI uses energy, and AI can help make energy systems more efficient, flexible, and reliable.

Chips Are Where Electricity Becomes Intelligence

If electricity is the fuel, chips are the engine, and better chips mean more intelligence per watt rather than just faster models.

ASML, the Dutch company that builds the machines used to print the world’s most advanced chips, says its newest platform can create features 1.7 times smaller than its current systems, allowing nearly three times as many transistors in the same space. That’s an equipment-maker’s claim, but it reflects where advanced manufacturing is heading. [10]

TSMC, the world’s largest contract chipmaker, says its next-generation process should cut power use by 25% to 30% at the same speed compared with today’s most advanced chips. A roadmap, not independent proof, but the direction is clear: more compute for the same energy or less. [11]

Chips aren’t the only gate, either. These systems don’t just need to calculate, they need to move enormous amounts of data fast, and memory makers say their newest parts do that with about 30% less power than competing products. [12] Compute without memory bandwidth is stranded intelligence.

And here’s where the loop turns recursive: AI is now helping design the chips themselves. Google says AI-generated layouts have been used in every generation of its in-house AI chip since 2020. That’s company-reported, and chip-design claims are hard to verify from outside, but the pattern holds. [13] AI helps design better chips, better chips run better AI, and better AI helps design the next generation. That’s the machine building the machine. The caution is obvious, because a chip isn’t a blog post and mistakes cost real money and time. But AI is clearly moving upstream into the hardware that will run future AI.

This matters beyond AI, too. More efficient chips help medical imaging, robotics, genomics, wearables, autonomous systems, defense, industrial automation, and edge devices. When compute gets cheaper per watt, more fields can afford to use it.

Software Can Matter as Much as Hardware

Sometimes the breakthrough is software that makes the chips you already have more useful.

In the past few years, researchers have published software techniques that squeeze two to four times more output from the same AI hardware, simply by managing how data moves through the machine. [14][15] I can’t independently evaluate the underlying engineering. What I can do is track what the researchers publish and what the deployment numbers show, and both point the same direction.

This matters because software cycles move faster than chip cycles. A new chip can take years. A software improvement can spread in months. Lower the cost of running these models and more people can use AI. Improve utilization and the same data center produces more useful work. Software efficiency is one of the fastest ways to bring down the cost of intelligence.

Don’t Freeze AI at Today’s Model

There’s another reason to be careful judging AI by today’s systems: the models themselves are changing.

Most people meet AI through today’s chatbots, which generate text by predicting one small piece at a time. They’re impressive, but they aren’t the final form. The field is already branching in several directions: models that think longer on harder problems instead of answering immediately, models that learn to reason through trial and error rather than only from human examples, models built to handle book-length documents and genomes without the cost exploding, models that carry memory instead of being reminded of everything each time, models small and efficient enough to run on your phone, and models learning physics from video, which early results suggest could help robots handle objects in places they’ve never seen. [16][17][18][19][20][21][22]

Not all these win. The simpler takeaway is that we’re looking at one frame of a moving film, not the finished product. That matters for the infrastructure debate, because if models get more efficient, run locally, or understand the physical world better, the economics and the constraints shift again.

Science Is Where This Gets Bigger Than AI

Forget better chatbots for a minute. The prize is faster science. Science is a loop of observe, hypothesize, test, fail, refine, repeat, and AI can compress parts of that loop.

AlphaFold’s database offers open access to more than 200 million predicted protein structures, work that used to take a graduate student years per protein. [23] DeepMind reported millions of new candidate crystal structures for materials science. [24] An autonomous laboratory in Berkeley synthesized 36 of 57 target materials in 17 days; I’m using the corrected Nature figure rather than the higher numbers in early coverage, because the corrected one is the safer number to stand behind. [25] And new generative models point toward inverse design, where you ask for a material with the properties you want instead of screening what already exists. [26]

None of this replaces scientists. It helps them search larger spaces, generate better hypotheses, and decide which experiments are worth running. That reaches drugs, batteries, catalysts, carbon capture, semiconductors, agriculture, diagnostics, and industrial materials. AI doesn’t replace the scientific method. It can make parts of it move faster.

Medicine Shows the Promise and the Caution

Healthcare is where the upside is easiest to see and the risks are hardest to ignore.

Insilico’s rentosertib is an AI-designed drug for idiopathic pulmonary fibrosis, a progressive and often fatal lung-scarring disease. In a mid-stage trial published in Nature Medicine, patients on the highest dose gained lung capacity over 12 weeks while patients on placebo lost it. That’s early, and it isn’t approval or proof that AI has solved drug development. But it’s a real signal. [27]

In breast cancer screening, a large, randomized trial in Sweden showed AI-supported mammography cut radiologists’ screen-reading workload by 44%, [28] and a separate analysis reported a 29% increase in cancer detection without raising false alarms. [35]

Promising, but healthcare AI won’t be judged by demos. It’ll be judged by outcomes, safety, equity, cost, liability, and trust, and that bar should be high.

What This Means for Workers

The job conversation is usually too binary. AI replaces everyone, or AI helps everyone. Neither is serious. The better model is that tasks change first, jobs change second, and institutions change last.

A National Bureau of Economic Research study of more than 5,000 customer-support agents found that an AI assistant raised productivity by 14% on average, and by 34% for newer, less-skilled workers. [29] A Harvard Business School and Boston Consulting Group experiment with 758 consultants found that workers using AI finished more tasks, worked about 25% faster, and produced better work on tasks the AI handled well. But on a task designed to sit just outside the AI’s ability, the people using it were 19% more likely to get the answer wrong. [30]

That gap is the whole issue. AI can help, and AI can mislead, and the difference is knowing when the tool is inside its capability boundary and when it isn’t.

The International Monetary Fund estimates about 40% of jobs worldwide are exposed to AI, rising to about 60% in advanced economies, with roughly half of those jobs likely to benefit and the other half facing pressure on demand and wages. [31] The World Economic Forum projects that by 2030, the labor market could create 170 million new roles while displacing 92 million. That’s a forecast, not a guarantee, but it makes the point: transition is the real problem to manage. [32]

One more thing the exposure rankings miss: the structure of a job matters more than its exposure score. A job built around a single task can be eliminated whole. A job built around ten tasks gets reshaped instead, and the person doing it often becomes more valuable on what’s left. That’s why a long-haul truck driver is at more risk over the next five years than a management consultant, even though the popular rankings will tell you the opposite. I published a four-question screen earlier this year for scoring specific roles on exactly this, and operators are now running it on their own industries. [36]

Workers who learn to use AI well gain real advantage; the ones who blindly trust it create risk. The same split runs through companies: the ones that redesign their workflows capture value, and the ones that buy licenses and wait for productivity to appear waste money.

What This Means for Individuals

For individuals, AI means access: a tutor, a coding assistant, a writing partner, a research assistant, a health explainer, a financial explainer, a creative collaborator. That doesn’t make everyone an expert, but it changes the starting line. A person with curiosity, judgment, and good questions can now reach capabilities that used to require a team.

That access is powerful and risky, because the same tool that helps people learn can help them avoid learning. One study found students with AI access did better while they had the tool and worse once it was taken away, when the system had no guardrails. The safer design wasn’t an answer machine. It was a tutor-like setup that kept students intellectually active. [33] Whether this access becomes real capability or quiet dependency comes down to literacy, transparency, and control.

The Downside Is Not Theoretical

Data centers can strain local power systems. Water use can become a community fight. Emissions can rise if new demand is met with fossil generation. Workers can be displaced or deskilled. Students can outsource their thinking. Medical AI can make mistakes. Companies can concentrate power. Bad actors can scale misinformation, fraud, cyberattacks, and surveillance.

The IEA projects data-center emissions could nearly triple by 2035 in its higher-growth case, while still staying below 1.5% of total energy-sector emissions. At the same time, it estimates that broad use of existing AI applications could cut global emissions by an amount equal to roughly 5% of energy-related emissions in 2035. The footprint is real and the mitigation potential is real, and neither one cancels the other. [3]

Stanford’s 2026 AI Index reports documented AI incidents rose to 362 in 2025, up from 233 the year before, and that responsible-AI testing isn’t keeping pace with capability testing. [34] The concern is governance, incentives, transparency, and whether institutions can move fast enough.

So, What Does This Mean for Us?

This is the question underneath everything else, and it’s bigger than “is AI good” or “is AI bad.” What happens when intelligence becomes cheaper, more available, and embedded into the systems that shape human life?

Workers will see many tasks change. Some roles get augmented, some get displaced, and the premium shifts toward judgment, adaptability, domain expertise, taste, ethics, and the ability to work alongside intelligent systems. Individuals gain access to capabilities that used to require experts, institutions, or money. Science moves faster from question to hypothesis to experiment. Medicine gets earlier detection, better discovery tools, more personalized support, and possibly faster therapeutics.

In companies, the winners won’t be the ones that say they use AI, because everyone will say that. They’ll be the ones that redesign work, measure quality, retrain people, protect trust, and figure out where AI creates an advantage.

At the level of society, it comes down to a choice. We can use AI to concentrate power, erode trust, strain resources, and displace workers badly. Or we can use it to expand human capability, strengthen infrastructure, speed up discovery, improve healthcare, cut waste, and make expertise more broadly available. The technology doesn’t decide that. We do.

What Might Be Possible Next

Over the next one to three years, AI becomes a standard layer inside knowledge work: writing, coding, research, customer support, legal review, marketing, operations, analytics, internal search.

Three to seven years out, it moves deeper into infrastructure: chip design, data-center optimization, grid management, scientific discovery, materials design, drug development, robotics, manufacturing.

Beyond that, on a seven-to-fifteen-year horizon, the line between digital intelligence and physical experimentation may get much thinner. AI proposes designs, robots run the experiments, software analyzes the results, models improve, and humans set the goals, constraints, ethics, and direction. None of that is guaranteed, but it’s where the early evidence points.

Better chatbots are the least interesting part of what’s coming. The bigger change is a faster loop between intelligence and the physical world.

Back to the Questions at Dinner

I didn’t give those two a clean answer that night, so here’s the version I’d give now.

On hallucinations, the question that matters is not whether the tool makes things up but where it does. These systems are reliable inside their capability boundary and unreliable outside it, and the whole skill is knowing which side of that line you’re on. The Harvard consultants showed both halves in one study: real gains on tasks inside the boundary, and a 19% higher chance of a wrong answer on the task outside it. The trap is that the tool sounds just as sure in both cases and won’t tell you which one you’re in. [30]

The control question is the one I think about most, and it has the least comforting answer. The infrastructure is expensive and consolidating. Power, chips, data centers, and the largest models are gathering in a small number of hands. Underneath that sits a quieter layer: whoever owns the default path between what you want and what gets done. Own that path and you collect a fee on every decision that runs through it. That part isn’t settled, and it’s the fight I’d tell those two to watch.

Their last worry was whether we’re moving faster than people, companies, schools, and governments can absorb. The honest answer is yes. The technology compounds and the institutions around it don’t, and that mismatch is the real risk, more than any single model or data center.

And none of this is a prediction that things go smoothly. They won’t. Every question those two asked will still be a live issue five and ten years from now: jobs, trust, water, power, control. What I’m arguing against is letting the loudest voices set the frame, because the people shouting on either side mostly aren’t measuring anything. This piece is my attempt at a balanced ledger instead: real costs, real benefits, sourced so you can check the math yourself.

The Machine That Builds the Future

AI has costs. The real question is whether we’re honest enough to measure them correctly, manage them intelligently, and weigh them against the benefits this technology can deliver.

The loop is forming. Electricity becomes compute, compute becomes intelligence, and that intelligence is improving energy, chips, software, medicine, science, and industry, which in turn creates more capability. That’s powerful, and it’s dangerous if it goes unmanaged.

What happens next won’t be decided by the model alone. It’ll be decided by the systems around the model: power, water, chip supply chains, software, workers, scientists, regulators, and institutions. We’re getting good at turning electricity into intelligence. The part we haven’t figured out is turning intelligence into judgment.

Harry Glorikian is the author of The Invisible Interface: How AI Turns Intentions into Actions—and Who Wins (Ideapress Publishing / Simon & Schuster, June 2026).

Pre-order The Invisible Interface: Reserve your copy on Amazon ahead of the June 2026 release.

References

[1]: International Energy Agency, Energy Demand from AI. Data centers consumed ~415 TWh globally in 2024 (~1.5% of consumption), projected ~945 TWh (~3%) by 2030. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai

[2]: U.S. Department of Energy / Lawrence Berkeley National Laboratory. U.S. data centers used 176 TWh in 2023 (4.4% of U.S. electricity), projected 325-580 TWh (6.7%-12%) by 2028. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers

[3]: International Energy Agency, Energy and AI: Executive Summary. Includes grid-delay estimates, outage-reduction figures, the 175 GW transmission-capacity estimate, and emissions projections (180 Mt CO2 today, 300-500 Mt by 2035). https://www.iea.org/reports/energy-and-ai/executive-summary

[4]: MOST Policy Initiative, Data Center Water Use. U.S. data centers: ~17.4 billion gallons (2023), projected 38-73 billion by 2028. Google: 4.3 billion gallons (2021) to 6.1 billion (2024). https://mostpolicyinitiative.org/science-note/data-center-water-use/

[5]: Microsoft, Datacenter Sustainability: Efficiency. Company-reported FY25 efficiency figures (PUE 1.17, WUE 0.27 L/kWh). https://datacenters.microsoft.com/sustainability/efficiency/

[6]: Microsoft, Made in Wisconsin: The World’s Most Powerful AI Datacenter. https://blogs.microsoft.com/on-the-issues/2025/09/18/made-in-wisconsin-the-worlds-most-powerful-ai-datacenter/

[7]: Reuters, Ahead of SpaceX IPO, Musk Says AI Satellites Will Use Mostly Existing Technology. First proposed satellite: ~150 kW peak, ~120 kW sustained compute. https://www.reuters.com/business/media-telecom/ahead-spacex-ipo-musk-says-ai-satellites-will-use-mostly-existing-technology-2026-06-09/

[8]: Reuters, SpaceX Says Unproven AI Space Data Centers May Not Be Commercially Viable, Filing Shows. https://www.reuters.com/world/spacex-says-unproven-ai-space-data-centers-may-not-be-commercially-viable-filing-2026-04-21/

[9]: Google DeepMind, DeepMind AI Reduces Google Data Centre Cooling Bill by 40%. https://deepmind.google/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-by-40/

[10]: ASML, 5 Things You Should Know About High-NA EUV Lithography. 8 nm resolution; features 1.7x smaller; transistor density up to 2.9x higher. https://www.asml.com/news/stories/2024/5-things-high-na-euv

[11]: TSMC, A14 Technology. 10-15% speed gain at same power, or 25-30% power reduction at same speed, vs. N2; >20% logic-density improvement. https://www.tsmc.com/english/dedicatedFoundry/technology/logic/l_A14

[12]: Micron, HBM3E Product Page. >1.2 TB/s bandwidth; ~30% lower power than competing HBM3E. https://www.micron.com/products/memory/hbm/hbm3e

[13]: Google DeepMind, How AlphaChip Transformed Computer Chip Design. https://deepmind.google/blog/how-alphachip-transformed-computer-chip-design/

[14]: Tri Dao et al., FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness, NeurIPS 2022. Reported 3x speedup on GPT-2 at sequence length 1,000. https://proceedings.neurips.cc/paper_files/paper/2022/hash/67d57c32e20fd0a7a302cb81d36e40d5-Abstract-Conference.html

[15]: Kwon et al., Efficient Memory Management for Large Language Model Serving with PagedAttention / vLLM. Reported 2-4x throughput gains at similar latency. https://arxiv.org/abs/2309.06180

[16]: OpenAI, Learning to Reason with LLMs. https://openai.com/index/learning-to-reason-with-llms/

[17]: DeepSeek-R1 Incentivizes Reasoning in LLMs Through Reinforcement Learning, Nature. https://www.nature.com/articles/s41586-025-09422-z

[18]: Gu and Dao, Mamba: Linear-Time Sequence Modeling with Selective State Spaces. https://arxiv.org/abs/2312.00752

[19]: Google Research, Titans + MIRAS: Helping AI Have Long-Term Memory. https://research.google/blog/titans-miras-helping-ai-have-long-term-memory/

[20]: Inception Labs, Introducing Mercury (diffusion language model). https://www.inceptionlabs.ai/blog/introducing-mercury

[21]: Liquid AI, Introducing LFM2 (on-device models). https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models

[22]: Meta AI, Introducing the V-JEPA 2 World Model and New Benchmarks for Physical Reasoning. Reported 65-80% success on pick-and-place tasks in new environments. https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/

[23]: AlphaFold Protein Structure Database. https://alphafold.ebi.ac.uk/

[24]: Google DeepMind, Millions of New Materials Discovered with Deep Learning / GNoME. 2.2 million new crystal structures; 380,000 predicted stable. https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/

[25]: Szymanski et al., An Autonomous Laboratory for the Accelerated Synthesis of Inorganic Materials, Nature. https://www.nature.com/articles/s41586-023-06734-w

[26]: Zeni et al., A Generative Model for Inorganic Materials Design / MatterGen, Nature. https://www.nature.com/articles/s41586-025-08628-5

[27]: Xu et al., A Generative AI-Discovered TNIK Inhibitor for Idiopathic Pulmonary Fibrosis: A Randomized Phase 2a Trial, Nature Medicine. Highest dose: +98.4 mL mean forced vital capacity over 12 weeks vs. -20.3 mL on placebo. https://www.nature.com/articles/s41591-025-03743-2

[28]: Lang et al., Artificial Intelligence-Supported Screen Reading Versus Standard Double Reading in the Mammography Screening with Artificial Intelligence (MASAI) Trial, The Lancet Oncology. https://pubmed.ncbi.nlm.nih.gov/37541274/

[29]: Brynjolfsson, Li, and Raymond, Generative AI at Work, NBER. https://www.nber.org/papers/w31161

[30]: Harvard Business School, Navigating the Jagged Technological Frontier. 12.2% more tasks, 25.1% faster inside the frontier; 19% less likely to be correct outside it. https://www.hbs.edu/faculty/Pages/item.aspx?num=64700

[31]: International Monetary Fund, AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity. https://www.imf.org/en/blogs/articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity

[32]: World Economic Forum, Future of Jobs Report 2025. https://www.weforum.org/press/2025/01/future-of-jobs-report-2025-78-million-new-job-opportunities-by-2030-but-urgent-upskilling-needed-to-prepare-workforces/

[33]: Bastani et al., Generative AI Without Guardrails Can Harm Learning, PNAS. https://www.pnas.org/doi/10.1073/pnas.2422633122

[34]: Stanford HAI, 2026 AI Index Report: Responsible AI. https://hai.stanford.edu/ai-index/2026-ai-index-report/responsible-ai

[35]: Lang et al., Screening Performance and Characteristics of Breast Cancer Detected in the MASAI Trial, The Lancet Digital Health. https://www.thelancet.com/journals/landig/article/PIIS2589-7500%2824%2900267-X/fulltext

[36]: Harry Glorikian, The Four-Question Role Screen, glorikian.com, April 2026. [INSERT EXACT URL BEFORE PUBLISHING]