In 2013, two researchers at the Oxford Martin School, Carl Benedikt Frey and Michael Osborne, published a paper estimating that 47 percent of total US employment was at high risk of computerisation over the following two decades. [1] The number traveled fast. It appeared in headlines, policy speeches, and roughly every article about robots taking jobs written in the years after. It was also widely misunderstood, including by people quoting it approvingly.

Frey and Osborne did not predict that 47 percent of jobs would be automated. They estimated the share of jobs that had a high probability of being technically automatable at some point, given foreseeable technology. Those are different claims, and the gap between them is where most of the ethics lives. This final part of the series is about that gap: the questions AI raises not as a technology but as a concentration of power, labor, money, energy, and legal authority. These are the questions that outlast any particular model.

The number that scared everyone

Three years after Frey and Osborne, a team at the OECD reran the analysis with a different unit of measurement. Instead of asking whether whole occupations could be automated, Melanie Arntz, Terry Gregory, and Ulrich Zierahn asked how many of the individual tasks inside each job could be. Most jobs are bundles of tasks, and even highly exposed jobs contain work that resists automation. On that task-based approach, the share of jobs at high risk across OECD countries dropped to 9 percent. [2] Same question, different lens, a fivefold difference in the answer. That spread should make anyone quoting a single automation percentage nervous.

The generative AI wave produced its own version of the estimate. In 2023, researchers at OpenAI and the University of Pennsylvania published “GPTs are GPTs,” which found that around 80 percent of the US workforce could have at least 10 percent of their work tasks affected by large language models, and roughly 19 percent could see at least half of their tasks affected. [3] Notice the framing again: tasks affected, not jobs eliminated. And unlike earlier waves of automation, the exposure skewed toward higher-income, more educated work. The International Monetary Fund reached a similar shape in 2024, estimating that about 40 percent of jobs worldwide are exposed to AI, rising to roughly 60 percent in advanced economies, with about half of those exposed jobs potentially helped rather than displaced. [4]

The honest reading of all this is that exposure is not destiny. A task being technically automatable does not mean it will be automated, that the automation will be cheaper than the human, or that the freed-up time will not create new work. Daron Acemoglu, one of the economists who has studied automation most carefully, modeled the actual macroeconomic effect and came out notably cool: he estimated that AI would raise total factor productivity by no more than about 0.66 percent over ten years, with the direct GDP effect similarly modest. [5] That is a real effect. It is not the civilizational rupture the 47 percent number was taken to imply.

The ethical point is not that displacement is a myth. It is that “AI will take the jobs” is the wrong abstraction. The right questions are narrower and harder: which specific workers lose bargaining power, whether the productivity gains flow to wages or to capital, and whether the people whose tasks are automated are the same people who capture the upside. Acemoglu’s larger body of work is precisely about that distribution, and history is not reassuring: automation tends to raise the returns to whoever owns the automating technology. Which brings us to who that is.

Who owns the machines

The 2024 Stanford AI Index put a price tag on the frontier. Training OpenAI’s GPT-4 used an estimated 78 million dollars’ worth of compute. Google’s Gemini Ultra used an estimated 191 million. [6] For comparison, the report notes the original Transformer model from 2017 cost around 900 dollars of compute to train. In seven years, the cost of a state-of-the-art model rose by roughly five orders of magnitude. Private investment moved in step: funding for generative AI alone reached 25.2 billion dollars in 2023, nearly eight times the previous year. [6]

Those numbers are the whole ballgame for one specific ethical question, which is concentration. When the ticket to build a frontier model costs nine figures and a data center full of scarce chips, the set of organizations that can build one shrinks to a handful of large firms and their partners. This is not speculation. Nur Ahmed and Muntasir Wahed documented the trend early, in a 2020 study analyzing more than 170,000 AI research papers. They found a widening gap they called the “compute divide”: large technology firms and a small number of elite universities were increasingly dominant in deep-learning research, while mid-tier and less-resourced institutions were pushed out, precisely because modern AI research had become gated by access to expensive compute. [7]

The reason this is an ethics problem and not just an industrial-organization problem is that these systems increasingly mediate how people find information, get hired, receive medical triage, and interact with the state. When the capacity to build them is concentrated in a few firms, so is the power to decide what they refuse to do, whose values they encode, and what the defaults are for hundreds of millions of users. Market concentration in a soap company affects the price of soap. Concentration in the infrastructure that shapes what people read, write, and believe is a different kind of concern. It does not require anyone to be a villain. It just requires the ownership of a general-purpose capability to sit in very few hands.

The power bill

The environmental case against AI is real, and it is also the area where numbers get abused most freely, so it is worth being careful.

The paper that launched the conversation, by Emma Strubell and colleagues in 2019, reported a headline figure that a full neural architecture search for a Transformer model could emit around 626,000 pounds of carbon dioxide, which they compared to about five cars over their lifetimes. [8] That number got quoted everywhere. It was also, it turned out, a substantial overestimate for that particular case. A 2021 analysis by David Patterson and colleagues at Google recalculated the same architecture search and found the real figure was smaller by a large factor, once you accounted for how the search was actually run and the efficiency of the hardware and data center. [9] I bring this up not to dismiss the concern but because it is a clean example of the brief’s warning: environmental figures for AI are routinely misquoted, and the splashiest ones often come from back-of-envelope estimates that do not survive scrutiny.

The better-grounded numbers are still substantial. When Sasha Luccioni and colleagues did a careful lifecycle estimate of BLOOM, a 176-billion-parameter model, they found training it emitted about 25 tonnes of carbon dioxide equivalent counting only the electricity used, and about 50 tonnes counting manufacturing and infrastructure. [10] That is one model, trained once, on a relatively low-carbon grid.

The figure that actually matters is not any single model but the aggregate. The International Energy Agency estimated that data centers consumed around 460 terawatt-hours of electricity globally in 2022, and projected that data centers, AI, and cryptocurrency together could push that past 1,000 terawatt-hours by 2026, roughly the annual electricity consumption of Japan. [11] Water is the quieter cost. Researchers led by Pengfei Li estimated that training GPT-3 in Microsoft’s US data centers could directly evaporate around 700,000 liters of clean freshwater for cooling, and that a short exchange of a few dozen queries consumes on the order of a half-liter bottle, depending heavily on where and when it runs. [12]

The ethics here is not “computing uses energy,” which is trivially true of everything. It is about who bears the cost and who decides. Data centers get sited in specific communities, draw on specific water tables, and load specific electrical grids, and the people who absorb those local effects are usually not the people capturing the value. That is a distributional question, and distributional questions are exactly the ones markets handle badly on their own.

Writing the rules

On August 1, 2024, the European Union’s AI Act entered into force, the first comprehensive horizontal law aimed at regulating AI by a major jurisdiction. [13] Its core design is a risk pyramid. A small set of uses is simply prohibited as posing unacceptable risk, such as government social scoring and certain manipulative or exploitative systems. A larger “high-risk” category, covering AI used in things like hiring, credit, education, and critical infrastructure, is allowed but subject to heavy obligations around data quality, transparency, human oversight, and documentation. Most systems fall into lower tiers with light or no obligations. The bans took effect in early 2025 and the rules for general-purpose models in mid-2025, while the heaviest high-risk obligations phase in over the following years, on a timeline that has itself been the subject of proposed delays. [13]

The contrast with the United States is instructive, because it shows how contingent all of this is. In October 2023, the Biden administration issued a sweeping executive order on AI safety and oversight. On his first days back in office in January 2025, President Trump reversed course, revoking that order and setting new policy in one titled “Removing Barriers to American Leadership in Artificial Intelligence,” which reoriented federal policy away from mandated oversight and toward deregulation and speed. [14] Within about fifteen months, the same country reversed its posture entirely, by executive fiat, with no change in the underlying technology. That is worth sitting with. It means the governance layer, the thing that is supposed to hold the technology accountable, is at least as unstable as the technology itself, and often more so.

There is no neutral choice here. Heavy regulation can entrench incumbents, since only large firms can afford large compliance departments, which quietly reinforces the concentration problem from earlier. Light regulation leaves the externalities, on labor, environment, and information, to be absorbed by whoever is least able to refuse them. The interesting policy work is in the details of which specific harms get named and who carries the burden of proof, not in the abstract question of more rules versus fewer.

Open or closed

The last fault line runs through the technology itself, and it is a genuine ethical dilemma rather than a case with an obvious good side. Should powerful models be released openly, with their weights downloadable by anyone, or kept closed behind an API the developer controls?

The case for open models is accountability and distribution of power. You cannot independently audit a system you cannot inspect, and closed models concentrate control in exactly the few firms discussed above. The case against is misuse: weights that anyone can download are weights that anyone can strip of safety guardrails and repurpose. Both concerns are real, which is why serious researchers land in different places.

Two papers frame the tension well. In 2024, a large group led by Sayash Kapoor and Rishi Bommasani argued for taking open models seriously by assessing their marginal risk, meaning the additional harm they enable beyond what existing tools already allow. Across misuse vectors like cyberattacks and biological weapons, they found the current evidence insufficient to show that open models meaningfully raise the risk over what is already possible, while the benefits for innovation, competition, and transparency are concrete. [15] Pointing the other way, David Gray Widder, Meredith Whittaker, and Sarah Myers West published a paper in Nature the same year arguing that “open” in AI is often marketing. Many systems branded as open still depend entirely on the data, compute, and infrastructure of a few large firms, so the label can obscure concentration rather than counter it, and openness by itself guarantees neither accountability nor a shift in power. [16]

Both can be right at once. Openness is a genuine check on concentrated power and a genuine vector for misuse, and calling a model “open” does not automatically deliver the benefits people associate with the word. The ethical move is to stop treating “open versus closed” as a slogan and start asking the specific questions underneath it: open in what respect, released to whom, with what documented risks, and auditable by which independent parties.

The through-line

Every thread in this part is the same question wearing different clothes. Who benefits, who bears the cost, and who decides. Labor exposure is about whether the gains from automating tasks flow to workers or to the owners of the automation. Compute concentration is about whether the capacity to build these systems, and therefore to set their defaults, sits in many hands or few. The energy and water bills are about which communities absorb costs for value captured elsewhere. Regulation is about who gets to name the harms and assign the burden. Open versus closed is about whether the power to inspect and to misuse should be widely held or tightly kept.

None of these are questions the models can answer, and none of them get easier as the models get better. They get harder, because more capable systems raise the stakes on every one of them. The technical progress is genuinely impressive. But the parts of this that will matter most in ten years are not in the architecture. They are in the ordinary, unglamorous, deeply political work of deciding how the power gets distributed. That work is ours, not the machine’s, and it is the part no benchmark measures.

References

1. Frey, C.B., & Osborne, M.A. (2013). The Future of Employment: How Susceptible Are Jobs to Computerisation? Oxford Martin School. The paper estimates that 47% of total US employment is at high risk of computerisation over roughly the next two decades. https://oms-www.files.svdcdn.com/production/downloads/academic/The_Future_of_Employment.pdf
2. Arntz, M., Gregory, T., & Zierahn, U. (2016). The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis. OECD Social, Employment and Migration Working Papers No. 189. Using a task-based approach, only about 9% of jobs across OECD countries are found to be highly automatable, versus the 47% from occupation-based estimates. https://www.oecd.org/en/publications/the-risk-of-automation-for-jobs-in-oecd-countries_5jlz9h56dvq7-en.html
3. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. Around 80% of the US workforce could have at least 10% of their work tasks affected by LLMs, and about 19% could see at least 50% of tasks affected. https://arxiv.org/abs/2303.10130
4. Cazzaniga, M., et al. (2024). Gen-AI: Artificial Intelligence and the Future of Work. IMF Staff Discussion Note. About 40% of jobs globally are exposed to AI, rising to roughly 60% in advanced economies, with about half of exposed jobs potentially complemented rather than displaced. https://www.imf.org/en/blogs/articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity
5. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper No. 32487. Estimates the macroeconomic effect of AI as modest: no more than about a 0.66% increase in total factor productivity over ten years. https://www.nber.org/papers/w32487
6. Stanford Institute for Human-Centered AI (2024). The 2024 AI Index Report. GPT-4’s training used an estimated $78 million of compute and Gemini Ultra an estimated $191 million; generative AI private investment reached $25.2 billion in 2023. https://hai.stanford.edu/ai-index/2024-ai-index-report
7. Ahmed, N., & Wahed, M. (2020). The De-democratization of AI: Deep Learning and the Compute Divide in Artificial Intelligence Research. Documents a widening “compute divide” favoring large firms and elite universities in AI research, driven by unequal access to compute. https://arxiv.org/abs/2010.15581
8. Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. Proceedings of ACL 2019. Reports an estimate of roughly 626,000 lbs of CO2 for a Transformer neural architecture search, compared to about five cars over their lifetimes. https://aclanthology.org/P19-1355/
9. Patterson, D., et al. (2021). Carbon Emissions and Large Neural Network Training. Recalculates the Evolved Transformer neural architecture search emissions and finds the earlier estimate was substantially too high once search process and hardware/data-center efficiency are accounted for. https://arxiv.org/abs/2104.10350
10. Luccioni, A.S., Viguier, S., & Ligozat, A.-L. (2022). Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. Training BLOOM emitted about 24.7 tonnes CO2e counting dynamic power consumption, and about 50.5 tonnes counting the full lifecycle including manufacturing. https://arxiv.org/abs/2211.02001
11. International Energy Agency (2024). Electricity 2024. Data centers consumed an estimated 460 TWh globally in 2022; data centers, AI, and cryptocurrency together could exceed 1,000 TWh by 2026, roughly the electricity consumption of Japan. https://www.iea.org/reports/electricity-2024/executive-summary
12. Li, P., Yang, J., Islam, M.A., & Ren, S. (2023). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. Training GPT-3 in Microsoft’s US data centers could directly evaporate about 700,000 liters of freshwater; a short set of queries consumes on the order of a 500ml bottle. https://arxiv.org/abs/2304.03271
13. European Commission. Regulatory framework for AI (AI Act, Regulation (EU) 2024/1689). Entered into force 1 August 2024, using a four-tier risk structure (unacceptable, high, limited, minimal) with staggered application dates for prohibitions, general-purpose model rules, and high-risk obligations. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
14. The White House (2025). Removing Barriers to American Leadership in Artificial Intelligence (Executive Order 14179, January 23, 2025). Rescinds the prior 2023 executive order on AI and reorients federal policy toward deregulation. https://www.whitehouse.gov/presidential-actions/2025/01/removing-barriers-to-american-leadership-in-artificial-intelligence/
15. Kapoor, S., Bommasani, R., Klyman, K., et al. (2024). On the Societal Impact of Open Foundation Models. ICML 2024. Proposes a marginal-risk framework and finds current evidence insufficient to show open models meaningfully increase misuse risk over existing technologies, while their benefits are concrete. https://arxiv.org/abs/2403.07918
16. Widder, D.G., Whittaker, M., & Myers West, S. (2024). Why ‘open’ AI systems are actually closed, and why this matters. Nature, 635, 827–833. Argues that “open” AI often remains dependent on the resources of a few large firms, so openness alone does not guarantee accountability or a shift in power. https://www.nature.com/articles/s41586-024-08141-1