Is AI the “new electricity”?
“Just as electricity transformed almost everything almost 100 years ago, today, I actually have a hard time thinking of an industry that I don't think AI will transform in the next several years.” – Andrew Ng, 2017
“The governance of AI will not be easy. We can see this by thinking about the nature of AI as a general-purpose technology like electricity, the printing press, the combustion engine. These general-purpose technologies transform society, the economy, military in a deep fundamental way that's often hard to anticipate and very hard to govern.” – Allan Dafoe, 2019
“The way you produce the software is (…) in data centers generating tokens producing floating point numbers at very large scale. As if in the beginning of this last Industrial Revolution when people realized that you would set up factories apply energy to it and this invisible valuable thing called electricity came out AC generators. 100 years later, 200 years later, we are now creating new types of electrons, tokens, using infrastructure we call factories AI factories to generate this new incredibly valuable thing called artificial intelligence.” – Jensen Huang, 2024
1. Background
The AI-electricity analogy is quite straightforward. However, users of the analogy have put the emphasis on slightly different aspects.
The first to use the AI-electricity analogy was Kevin Kelly in his 2016 book The Inevitable. Just as almost everything was “electrified”, the AI revolution would “cognify” the world around us. His analogy specifically focuses on the grid and the electrification of various tools:
“This common utility will serve you as much IQ as you want but no more than you need. You’ll simply plug into the grid and get AI as if it was electricity. (…) The entrepreneurs didn’t need to generate the electricity; they bought it from the grid and used it to automate the previously manual (…) the business plans of the next 10’000 start-ups are easy to forecast: Take X and add AI.”
The AI researcher Andrew Ng popularized this analogy through a series of talks that argued that “AI is the new electricity” (e.g., 2016, 2017, 2017, 2017, 2017, 2017, 2018, 2018, 2018, 2019, 2019, 2020, 2021, 2023, 2023, 2023, 2023, 2023, 2023, 2024, 2024).
The high watermark of the electricity-AI analogy was around 2017 and 2018. An example of a more adventurous use from that time is Kai Fu-Lee’s 2018 book AI Superpowers in which he argues that “If AI is the new electricity, big data is the oil that powers the generators.”1 (AI:data; electricity:oil). The electricity analogy still persists; however, it is more often tied to the more general concepts of general-purpose technologies or industrial revolutions. The initial framing of AI as a utility by Kelly has gained less traction within the tech sector.
1.1 General-purpose technology
Allan Dafoe founded the Centre for the Governance of AI at the University of Oxford and is now Head of Long-Term AI Strategy and Governance at Google Deepmind. He has consistently framed AI as a general-purpose technology, and he has used electrification as a prime example of such an economy-wide transformation (e.g., 2017, 2018, 2019).
General-purpose technology is a category coined by economists. It describes technologies with the following characteristics:
Pervasiveness: They are used widely across many sectors of the economy, not limited to one industry or field.
Innovational complementarities: General purpose technologies often enable significant complementary innovations with new types of artifacts or organizational structures.
Economic growth: General-purpose technologies are also referred to as “engines of growth” because they have the potential to significantly improve productivity in many sectors. Their adoption can lead to gains in output, efficiency, and long-run economic growth.
Delayed impact: The impact of general-purpose technologies on productivity is often delayed due to lock-in factors that delay the emergence of business processes that are designed around the new possibilities. In the short run, they can even have contractionary effects due to a diversion of resources from manufacturing to R&D.
The following is a list of technologies that have been described as general-purpose technologies:
Biology: Domestication of plants, domestication of animals, biotechnology
Materials: smelting of ore, bronze, iron
Information and communication technology: money, writing, printing, computer, Internet, AI (e.g. Trajtenberg 2019)
Energy: water wheel, steam engine, internal combustion engine, electricity
Transport: wheel, three masted sailing ship, railways, iron steamship, automobile, airplane
Organization: Factory system, mass production, lean production
In his 2018 AI Governance: A research agenda Dafoe specifically highlights the potential value of exploring the electrification analogy with regards to political economy and military AI: “Historical precedents and analogies can provide insight, such as consideration of the arms race for and with nuclear weapons, other arms races, and patent and economic technology races. What about analogies to other strategic general purpose technologies and more gradual technological transformations, like industrialization, electrification, and computerization? In what ways do each of these fail as analogies?” In 2023 Dafoe published an academic analysis of the AI-electrification analogy in a military context together with Jeffrey Ding. They make the case that the impact of electricity on military effectiveness were broad, delayed, and shaped by indirect productivity spillovers, and argued that this may also apply to AI.
The AI-electricity analogy has occasionally been picked up by policymakers and if so, often in conjunction with other general-purpose technologies. For example, the 2018 Coordinated Plan on Artificial Intelligence from the European Commission opened with the sentence “Like electricity in the past, artificial intelligence (AI) is transforming our world”. However, it is particularly UK policymakers that have adopted this analogy:
“Yet just as electricity, or steam power before it, started out with particular, often somewhat niche, uses prior to gradually becoming fundamental to almost all aspects of economic and social activity, AI may well grow to become a pervasive technology which underpins our daily existence. Electrification had many consequences: unprecedented opportunities for economic development, new risks of injury and death by electrocution, and debates over models of control, ownership and access, to name just a few. We can expect a similar process as AI technology continues to spread through our societies.”
- UK House of Lords. (2017). AI in the UK: ready, willing and able?
This is not an isolated use. Many major UK AI governance documents explicitly frame AI in line with electricity as a general-purpose technology (Growing the AI Industry in the UK, 2017; Establishing a pro-innovation approach to regulating AI, 2022; National AI Strategy – AI Action Plan, 2022; A pro-innovation approach to AI regulation, 2023; Future of Compute Review, 2023).
1.2 Second Industrial Revolution
The Second Industrial Revolution is generally dated between 1870 and 1914, and the rise of electricity as a general-purpose technology was an important aspect of it. As such, the electricity-AI analogy can also be used as an example of a technology as the engine of an Industrial Revolution. This is closer to how NVIDIA CEO Jensen Huang uses the electricity-analogy. For a full analysis of parallels with the Industrial Revolution, please see AI Revolution vs. Industrial Revolution.
1.3 On electrification
First, one element that is a bit peculiar about the AI-electricity analogy is that AI is also literally electricity. On some level, AI is just electrons moving on chips. So, we could say case closed, AI is electricity. However, reducing complex systems to such a low level of analysis is not very insightful. Humans are literally 60% water by weight, but studying water does not provide a lot of insights on humans. What we are interested in is comparing the societal, economical, and political structures of electrification to that of the AI industry.
Second, the term electrification can be ambiguous. Google defines electrification as “the conversion of a machine or system to the use of electrical power”. If we accept that definition, then electrification is a process that has started as early as 18402 but that has by no means finished yet. The share of electricity in world energy consumption has increased from about 0.1% in 1900 to 4% by 1950 to about 19% in 2022.
Even if we look at a rich country like the United States, electrification is far from over and it is easy to think of energy-intense sectors which are not yet electrified. Most notably, cars and heating.
However, when you listen or read how the electricity-AI analogy is used, it never refers to a still ongoing process. Rather it refers to a historical period in which the United States and other Western countries worked towards and eventually achieved universal access to electricity. Meaning that every company and every household is connected to the electricity grid and can use electricity for artificial light and other purposes.
For Western countries this roughly corresponds to the period from the invention of incandescent light bulbs around 1880 to about 1950. However, note that globally universal electricity access has not been achieved yet. Universal electricity access is indicator 7.1.1 of the UN Sustainable Development Goals for 2030. As of 2022, there were still about 700 million people without electricity access, mostly in Subsaharan Africa.
2. Policy implications
What policy implications could be deduced if we accept the electricity-AI analogy as a mental heuristic for the future of AI?
Nature of the situation
A multi-decade build-up of a new infrastructure layer that transforms the production in factories and empower a vast set of domestic appliances.
Stakes
Long-run economic growth through increased productivity with indirect effects on military competitiveness.
Policy prescriptions
Infrastructure investment: Electrification required billions of dollars of upfront investments in power plants, transmission lines, and distribution networks.
Some might intuitively think of electric current from their individual perspective as users rather than from a systems perspective:
Don’t regulate the technology, regulate the use cases: The electric current that comes out of the socket in our homes is standardized, but the bulk of product safety regulation falls on electricity applications. Electricity is adaptable across numerous applications, from lighting and heating to industrial processes, which have their own regulations. Andrew Ng has been consistent in opposing any level of cross-cutting AI regulation and instead argued that the full burden should be on AI applications (e.g., 2017, 2023).
No export controls on electric current: Electricity was not seen as a primarily military technology, nor is it traditionally framed as a dual-use technology. It is fair to say that the users of electricity-AI analogy are not exactly China hawks.
“[on training “electrical engineers” for AI] I think the US-China competition is a false dichotomy. Really, I think the US learns a lot from China, more and more. China has learned a lot from the US, and this is one of those things where the more people do it, the more we're all better off.” – Andrew Ng, 2018
Of course, the cross-cutting electricity production industry that enables the electric current to come to our homes is in fact heavily regulated:
Public utility regulation: Implementing price controls and policies aimed at universal service provision were strategies used to ensure that electricity reached not just urban but also rural and underserved areas, enhancing equity and social welfare. Furthermore, there are substantial environmental requirements, and mandatory safety and security assessments for power plants. This could imply mandatory safety and security audits for AI producers.
Export controls on power generation and distribution equipment: If you look at power plant equipment and specific sectors, such as nuclear energy, there are of course longstanding and extensive efforts to limit proliferation to unfriendly actors. This could roughly correspond to existing US export controls on AI hardware.
Chances of success of policy options
The chances of success of public vs private ownership of this new infrastructure will be contested.
Moral rightness of policy options
Universal access: Electricity is widely seen as a public utility to which all citizens of a country deserve access.
Dangers associated with a policy option
Slow adoption: Don’t be scared. “Implement AI all over the place”. Adoption brings local economic benefits that outweigh the risks, so the real risk is slow adoption.
Some might intuitively think of risks of electricity from their personal perspective as users:
Limited risk: We might think of individual-level risks of electricity from electrocution, such as a hair dryer falling into a bathtub or a child playing with a fork and an electricity plug. These are naturally quite limited risks.
If the users of the analogy have a systems perspective:
Monopoly risk: There need to be some price controls to limit rent seeking.
Dependency risk: Most other critical infrastructures depend on electricity. To reduce risks from blackouts, there is a need for regulated reliability and redundancies.
Environmental risk: The cost of externalities such as air pollution and greenhouse gases need to be internalized.
“Civilization altering technologies tend to scare many people. Consider electric current - an invisible force that can kill a human on the spot, demonstrated to kill an elephant. Are you willing to have it embedded in your house walls, next to your children?” - Wojciech Zaremba, 2023
“If you think of AI, if you think of superintelligence in particular, as just another technology, like electricity, you're probably not very worried. But you see, Turing thinks of superintelligence more like a new species. Think of it, we are building creepy, super capable, amoral psychopaths that don't sleep and think much faster than us, can make copies of themselves and have nothing human about them at all. So what could possibly go wrong?” – Max Tegmark, 2023
3. Key Communalities and Differences
Key Commonalities
1. General-purpose technology
Both electricity and AI share some of the key characteristics of general-purpose technologies.
Cross-industry applications: Both have applications in nearly all significant industries.
Innovational complements: For electricity this might be the electric motor, the incandescent lightbulb, the grid, the electrical telegraph, the telephone, and radiotelegraphy. For AI this might be things such as autonomous vehicles, and general robotics.
Productivity: Electrification is widely accepted as having increased productivity across a range of sectors. Similarly, AI's expected impact on productivity is substantial, driving cost reductions and innovation across sectors (e.g. pwc 2016, McKinsey 2023, Accenture 2024).
2. Switch from in-house capacity to an outsourced service?
Before electrification, power could generally only be transmitted mechanically over small distances using line shafts and belt drives. In practice, this meant that manufacturers had their own in-house power plants. For example, wind power was used for some processes such as grain milling in Europe. Factories were built near rivers or streams where water wheels or later, more efficient turbines, could be installed to harness the kinetic energy of flowing water. This was especially common in industries such as textiles, flour milling, and sawmilling. Finally, fossil fuels could be transported to the manufacturer and then turned into power by burning in their steam engines, which increased location flexibility. However, industries that could locate closer to coal mines to minimize the cost of coal transportation still had an advantage. This led to the growth of industrial centers in coal-rich regions, such as the Ruhr Valley in Germany.
Electricity self-production remained the norm in factories in the earliest years of electrification. After 1900 the cheaper prices of centrally produced power led them to switch to outsource power production and adopt an electricity-as-a-service model.3
Similarly, we can see some evidence for economies of scale and a service-based model in AI. The biggest owners of AI hardware in the world are the large cloud providers, and they offer AI compute as a service both for training AI and for running inference on trained models. In other words, the cloud providers will tell you, you don’t need to buy your own GPU’s, you can use GPUs-as-a-service when you need the computing power. Similarly, the frontier AI companies will tell you, you don’t need to train your own foundation model – just use their AI model running on a cloud.
If we project this forward in the long-term, this could mean a lower stock of fix-employed in-house intellectual labor in companies and an increasing share of flexible intelligence consumed from the hyperscalers. Or, to phrase it more prosaically an “exocortex” for companies (see exocortex analogy). Having said that, computer technology has seen different waves of decentralization and centralization (mainframe - centralized, personal computer - decentralized, cloud - centralized). So, it would be premature to declare the triumph of centralized AI hardware (see also “Steve Jobs and the computer as a bicycle for the mind”).
Key Differences
1. No new transmission infrastructure
Typically, households in developed economies are connected to several major utility and service networks. These are:
Electricity: Essential for powering appliances, lighting, heating and cooling systems, and more.
Water supply: Clean water for drinking, cooking, cleaning, and other domestic uses.
Wastewater removal: Removal of wastewater from homes and subsequent treatment.
Telecommunications: This includes multiple generations of wired and wireless communication technology, which are essential for communication and access to information in the modern world. Wired networks have included the telegraph, the telephone, telex, cable TV, and ultimately today’s fiber networks. Wireless networks have included radio broadcasting, TV broadcasting, satellite, and ultimately the different generations of mobile data networks (1G, 2G, 3G, 4G, 5G, etc.).
Natural gas: Exists in many developed economies for heating and cooking.
Additionally, there are mobile public networks that offer periodic services with a more active participation of consumers:
Waste collection: Regular waste and recyclable materials.
Postal services: Distribution and receipt of physical mail and packages.
As of now, there are no plans for any new major transmission and distribution network for AI, neither as a fixed infrastructure, nor as a periodic public service. Rather AI is distributed over the existing data networks as part of Internet traffic.
There is massive investment into new AI infrastructure, but this is an investment into data centers, not into a transmission network. So, if anything, from an infrastructure point of view it makes more sense to compare power plants and AI datacenters, as Jensen Huang does.
2. Local vs global market
Electricity has less transmission losses than mechanical power transmission, but it certainly still has transmission losses. This means the electricity industry is still substantially location dependent and shaped by a trade-off between economies of scale and transmission losses. In short, there are no cross-ocean electricity transmission lines4 and there is no globally integrated market for electricity. There is no global market price for a kilowatthour of electricity. As of March 2023, electricity costed an average of 14 cents per kilowatthour in Iceland, whereas it costs an average of 52 cents per kilowatthour in Germany.
Note that a map of synchronized grids overstates regional integration of electricity markets. Most of it is still national and in fact subnational. In France in which the 85% state-owned company EDF provides about 85% of electricity to households from its nuclear power plants, there is essentially a national price of electricity. However, in many countries there is zonal pricing for subnational zones. For example, Sweden has pricing based on four subnational zones. In Switzerland these zones are even more granular and largely correspond to municipalities. Despite being a small country, the electricity can still differ up to 5x by municipality.
There is no inherent information loss per distance, and the cost of transport for the photons and electrons is negligible. So, from a purely technological perspective and for non-time sensitive applications, there can be a global integrated market for AI. With the obvious caveat that AI still needs to comply with a diversity of local laws (e.g. copyright, hate speech, political censorship, adult content, data localization). In that sense AI will be much closer to the Internet, where there has been a 30+ year clash on “Internet fragmentation” or “digital sovereignty”.
3. Public utility regulation
Electric power production is an essential service, but it has characteristics of a natural monopoly because it has high barriers to entry (high infrastructure investment costs for power plants & transmission lines), and economies of scale. Hence, most governments have introduced regulation to ensure that they do not exploit their monopoly position (including price controls), and that they serve the broader public interest. This includes universal service obligations (offering service to all households within their service area, including those in remote or less economically attractive regions), reliability standards (electricity outages cause much more economy-wide costs than the immediate costs felt by electricity producers, meaning the socially optimal investment in reliability is higher than one based on narrow monetary incentives of power companies), and environmental considerations.
The broadband Internet infrastructure in the United States has also regulated as a public utility under Title II of the Communications Act of 1934 (2015-2018, 2024?) This includes net neutrality rules, which prohibit Internet service providers from blocking, throttling, or engaging in paid prioritization of traffic.
As of now, there are no public utility regulations that would put AI companies under a responsibility to serve the public interest. However, some parts of the AI supply chain do share high investment barriers to entry and economies of scale. In the EU there have been some preliminary efforts to evaluate anti-trust investigations into the AI industry (1,2).
At the same time, there is arguably less overlap between the companies that build AI models and the owners of the Internet transmission network than there was between power plants and transmission network during electrification. And, where there is, Google Fiber (ca. 1% of US households) could still be forced by net neutrality to not slow down ChatGPT and Anthropic. Combined with the absence of a transmission loss, this makes for a more competitive dynamics as Sam Altman argues:
“In electricity (…) you kind of have one choice of who to buy electricity from and you know one person that has a water hook up to your home and I think a lot of the regulation there is important and you wouldn't want that company maybe doing certain things with AI models. There will be many models to choose from and so if you're unhappy with the behavior of us, you go to some other company, and I think in that sense it looks more like a competitive marketplace.”
4. Universal electricity access vs. universal AI access
The idea of “democratization” or universal access has been present both in electrification as well as in AI. However, upon closer inspection AI has advanced far more quickly towards widespread access because it can be distributed over existing infrastructure. I personally dislike the use of the term “AI democratization”. The way this term is being used by tech companies, has nothing to do with democratic oversight over technology, strengthening the trust in democratic elections, or empowering democratic countries vis-à-vis techno-authoritarianism. The people that use the word mostly seem to mean increased or universal access, so that’s what this section focuses on.
Universal access has been a (national) political issue for all the major utility and service networks. The reason is simple, the higher the population density in an area, the less network infrastructure per capita is needed. On top of that, urban areas usually have a higher GDP per capita. So, if you follow market principles, the rollout of any new major utility and service network will have an urban-rural divide.
In the US there has been a conscious political effort to close this gap. Most notably, the Tennessee Valley Authority Act of 1933, and the Rural Electrification Act of 1936 under Franklin D. Roosevelt. These acts were massive federal infrastructure projects to create more power plants and extend the electricity grid in the rural areas of the US.
As discussed above, unlike electricity, AI does not require a new transmission network. It uses the regular Internet infrastructure. Furthermore, the location of a datacenter within a country does not really matter, because there are no transmission losses.
However, the logic of universal access can be applied to Internet infrastructure. In fact, the Rural Electrification Act has literally been amended to ICT-infrastructure, first to the telephone network and more recently to broadband Internet access. So, while we shouldn’t expect something like urban-rural divide in household electricity access in the US, the remaining digital divide is in some sense also an AI divide, albeit with caveats.
First, we need to distinguish between the digital divide as an absolute lack of access to the Internet and a relative digital divide between those with faster and slower Internet speeds. The first digital divide does largely not exist anymore within rich countries – based on a Pew survey urban areas in the U.S. have about a 7-9 per cent higher broadband / smartphone / laptop penetration rate. Similarly, this type of digital divide will eventually disappear between countries, but it’s still quite a bit of work with roughly a third of humanity, or around 2.6 billion people still unconnected to the Internet.
In contrast, expecting the relative bandwidth gap between rich and poor regions or countries to ever close is illusory and in fact not desirable. It is perfectly normal and efficient that the newest and most expensive communication technology (e.g. 5G) will always be first rolled out in rich urban centers rather than poor rural areas.
Therefore, an important question from an AI access perspective is: Does AI have especially high bandwidth requirements, so that only those with the absolute fastest Internet speeds can profit from it? The general answer is no. Online video-services such as Netflix, YouTube or adult sites require the most bandwidth today. AI to generate videos or entire games on the cloud will also require a lot of bandwidth and in the latter case be sensitive to any delays. However, for most productivity related purposes chatbots and text are already sufficient, and they only require very little bandwidth. So, overall, a relative digital divide is not that problematic for AI access.
In short: Yes, the world is inequal. Yes, AI has the potential to make the world more inequal, such as through its impact on labor markets. However, so far, by any measurable factor, the rollout of ChatGPT was unprecedent in terms of how fast most of the world has gained access.
However, my suspicion is that not all that call for “AI democratization” care that much about universal AI access. For example, open-sourcing foundation models which are fundamentally giant CSV files that not even AI PhDs fully understand is objectively not what empowers a farmer in some remote Indian village. Actors which may5 be empowered by open-source foundation models are:
academics that would like to examine and test AI models,
companies/militaries that want to run a LLM locally, and have the means to fine-tune a foundation model on proprietary data but lack the resources to train their own foundation model,
groups in authoritarian countries that lack access to large scale AI compute due to US export control restrictions,
groups that want to use foundation models without guardrails or in violation of terms of use.
In contrast, what enables more people to use AI is improved Internet access, a user-friendly interface, and a free tier of AI models. If we want to further improve access to AI, the logical way forward would be investment in electricity and Internet infrastructure in rural areas of the world’s poorest countries.
5. There is no free tier of electricity
Business model: The business models of electricity providers and those providing AI services have some overlap. Specifically, they both offer “metered” access to business customers, measured in kWh and tokens. However, LLM companies tend to also offer a free tier of service for less capable models, and a flatrate premium subscription for access to their most capable models. The first element of this is more remarkable than it may sound. The costs for LLM-inference are substantially higher than the cost for providing an Internet search. Yet, LLM companies do not just offer this service for free, they also offer it without advertisements. This stands in stark and positive contrast to advertisement-driven business models of companies like Facebook and Google. Free electricity can happen for short periods of time (e.g. solar overproduction), but it remains a very rare exception 150 years into electrification.
Early adopters: Household electricity started out as a luxury good in the homes of the superrich, such as J.P. Morgan (1882)6, that was subsequently “democratized”. I don’t think the same pattern holds for AI – at least at current performance levels. Part of it is a demographic effect (AI users: young, capital owners: old). Part of it is that at the current stage AI labor is much cheaper than human labor but not yet at the level of a human top performer in his or her field (with access to AI). I haven’t found reliable statistics on this. However, I cannot see a super-rich handing over their legal matters directly to GPT-5/Harvey anytime soon, but I could see it being used by many that could barely afford a human lawyer. You can already order by tablet in McDonald’s, but I doubt that there is a Michelin star restaurant without human waiters, etc.
6. Intelligence is not an interchangeable commodity
There are no “specialized electrons” that are good at running medical appliances to detect cancer but not useful at all in almost all other contexts. Your electric car doesn’t run faster, more energy-efficient, or safer depending on whether the electrons used to charge its battery were produced by burning coal, nuclear fission, or solar power. For practical purposes all 1051 or so electrons on Earth are interchangeable copies of each other, hence, electricity is a commodity. The key characteristics of commodities are that they are standardized and undifferentiated - the quality of the product is considered uniform or within a very narrow range of variation, so it is not a significant factor in purchasing decisions. Other examples of commodities include crude oil, wheat, copper, or gold.
Computing power is a commodity. Neural networks are not. Or, at least, not to a remotely similar degree. In an artificial neural network, each neuron is unique and so is every large neural network. To be clear the recent development towards AGI, means that the same AI can do an increasing set of tasks (rather than having 1’000 smaller, more specialized AIs), and AGI companies can “meter” the output of their AGI in tokens. So, in some sense there is a standardized output unit for intelligence.
However, AI will always encompass both narrow and general AIs, and an AGI still has input factors, such as proprietary training data sets and reinforcement learning from human feedback that will differ between two AGI companies. So, while intelligence is in some ways becoming more commoditized, it will never be as commoditized as electricity. No one will ever run a medical device on electrons from one power plant and then from another power plant just to see if it reaches the same conclusion. In contrast, it is very reasonable to ask for a second or even third opinion on medical diagnosis from both human and AI doctors.
7. Electricity was not a substitute for human labor
Humans have historically mostly been hired for their muscles and are now mostly hired for their brains. No humans were ever hired for their natural light. Electrification in its first phase was about replacing candles and gaslights with electric lights. Electrification in factories was a transition from one artificial form of energy to another. The replacement of line shafts and belt drives using steam engines and waterpower by electric motors created more energy efficiency, provided more flexibility in factory layout, reduced noise levels, and improved air quality. It shifted some of the skills needed for machine maintenance but didn’t replace human muscles or human brains. If we look at the fundamental production factors, electrification was primarily capital substitution. As such, it has never caused any significant worries about massive job losses or actual massive job turnover. Of course, there were still labor disputes, but these were directed directly against company owners.
This stands in stark contrast to the First Industrial Revolution in which many human laborers did lose their jobs, and which led to much more significant backlash directed against the technology itself, such as the Luddites, the Swing Riots, or the “Maschinenstürmer” that destroyed labor-replacing machines.
Artificial intelligence is a broad umbrella term and naturally there AI applications which are not labor replacing. However, all things considered AI will replace human brains in many tasks and even in entire jobs. The impact of AI on jobs is difficult to forecast, however, there are many serious institutions that predict a massive labor substitution. For example, Goldman Sachs predicts that in the next 10 years about 300 million people will lose their jobs to AI. Furthermore, there are at least some early warning signs of a human labor substitution effect in many industries (e.g. online customer service, journalism, Hollywood (1,2), programming (1,2)).
So, in terms of labor turnover and related societal unrest and pushback, the First Industrial Revolution is a better fit to AI than the Second Revolution. In the long-run AI potentially goes much further than any previous shift in the sense that it is the declared goal of many of the world’s biggest tech companies to build AGI. AGI is an ambiguous concept, but in many formulations, it explicitly includes the aspiration to permanently replace all or at least most human labor in Earth’s economy (rather than automating some tasks/jobs but enabling more new tasks/jobs).
8. Speed of diffusion / growth
If electricity consumption would have grown at the same pace as computing power consumption in the last six decades, you personally would now consume more electricity per year than the whole world did in the 1960s. This obviously hasn’t happened. If anything, modern appliances use less electricity than old ones due to environmental concerns.
The electrification of the United States from around 1880 to 1950 was a remarkable feat and it happened at a remarkable pace. However, electricity consumption has still not grown as fast as computing hardware and neural networks.
The highest relative growth speed of installed electric motor power capacity at US manufacturing plants happened in the decade from 1889 to 1899 with an annualized growth of about 40% (1899-1909 = 25%, 1909-1919 = 13%, 1919-1929 = 8%, 1929-1939=3%). I don’t have a perfect apples-to-apples comparison; however, general computing power did grow at about 40% per year for six decades (Moore’s Law) and the computing power to train large AI models has increased at rates closer to 300% per year in the last decade.
9. Electricity is less of a military technology
Shared broad and indirect impact: Jeffrey Ding and Allan Dafoe make a convincing argument that we can learn something from electricity as a general-purpose military transformation for AI. First, we should not just focus on AI weapons, but on a broader range of applications, including military targeting, logistics management, and decryption. Second, they argue that electricity had an indirect effect on the military by significantly upgrading industrial productivity, which increases military production potential. The same is arguably true for AI.
Their third argument, which takes the multi-decades productivity lag observed in re-organizing factory floors during electrification and projects it on to all military applications of electricity and AI is less convincing.7 I would not contest that there can be delayed impacts, but quite a few narrow applications are straightforward and have little diffusion and restructuring lag.
Electric power is overwhelmingly civilian: Any general-purpose technology has some military applications. Still, if we would look at electricity as a share of primary energy consumed, the electrification rate of the armed forces would be one of the lowest rates among major organizations. Not only are there no electric death-rays, but there are also no electric troop transporters, no electric tanks, no electric battleships, no electric submarines, no electric fighter jets, and no electric missiles.
Armed forces need to be mobile and able to operate in all kinds of environments where they cannot rely on the fixed infrastructure of the electricity grid. Further, electric batteries are simply no match in energy density for fossil fuels and cannot provide the endurance requirements of armed forces. Lithium-Ion batteries provide 200 to 300 watthours per kilogram. Gasoline comes in at 12’200 watthours per kilogram, diesel at 12’700.
AI weapons, export controls, and DARPA: AI is broader than lethal autonomous weapons, but still we cannot ignore that they do exist and that they are already being deployed. For example, tech billionaire Eric Schmidt is working on AI “slaughterbots” (1,2). Based on the reporting of 972 magazine Israel appears to already have created an AI kill list, that has de facto ordered the killing of 10’000+ Palestinians, with very limited human oversight. Electric power as a commodity also doesn’t contain any classified information or provide any technological advantage to potential adversaries. In contrast, AI chips, sensitive datasets, and trained AI models may all fall under export control restrictions. Lastly, the U.S. Defense Advanced Research Project Agency has been funding key AI research for six decades now. There was no equivalent to DARPA during US electrification.
10. Electrification of specific objects was largely a one-off event
Electrification has largely grown by electrifying more objects. It’s not that electrified objects would use more electricity every year (in fact it’s the opposite, we are focusing on energy efficiency over increased power). Nor is it the case that they get better electricity every year.
In contrast, AI-powered technology might profit from more regular software updates. Especially in contexts, where adversaries will adapt their technology and tactics to your technology. For example, an AI spam filter, an AI deepfake detector, an AI fraud detector, an AI malware detector, or an AI military object detector can all not be static over long periods of time. Otherwise, they will be significantly less effective. Rather there is a bit of a “cat & mouse” game, with the ability to update the intelligence of processes and objects without necessarily needing to replace or update corresponding hardware.
11. Complexity, explainability, predictability
The fundamental science of electricity and electro-magnetism was developed in the 19th century by figures such as Alessandro Volta, Michael Faraday, James Clerk Maxwell, and Heinrich Hertz pre-ceded the electricity grid. Hence, electrification was an actual engineering science, and we could calculate and correctly predict the behavior of electric infrastructure. There were still some side effects from interactions with the world at large that became clearer with deployment (e.g. overground transmission lines in cities as a hazard, vulnerability of these lines to weather).
However, this can in no way be compared to large neural networks, which are still largely unexplainable and at times unpredictable. The complexity of large neural networks is arguably also a counterargument against only regulating and auditing AI applications and not foundation models. The safety, security and legal compliance of foundation models will impact all downstream applications.
For example, if you are a health insurance provider and want to build a medical AI-assistant that understands natural language and can analyze pictures to provide preliminary medical diagnoses and triage, you will likely build on a foundation model. However, the compliance of such an application with requirements on robustness, security, explainability, fairness etc. depends on the underlying foundation model.
At the same time governments have limited capacity to audit large neural networks. Rather than 10 superficial assessments from industry-specific agencies with limited AI expertise, it would make much more sense to have one in-depth “foundation model audit” from an AI agency on which the more specialized agencies can build on for application-specific audits.
12. Agency, autonomy, and superhuman power potential
First, rather than “cognifying” 1’000 objects separately, as we electrified objects, large language models can serve as a smart universal interface to interact with the world. So, rather than having an AI-enabled “smart fridge”, an AI-enabled “smart closet”, AI-enabled “smart shoes” and an AI-enabled “smart toaster” you will much more likely have one “personal AI” that interacts with you and acts for you.
Second, electric current as it comes out of your socket is a controlled and understood physical phenomenon with no cognition, goals, or agency. You can’t talk to it and develop a relationship with it, it can’t think of a step-by-step plan and make decisions, let alone pass a university exam and beat you in chess. As such, it’s integration into society was arguably less complex than the integration of AI and the disaster risk was more locally bounded. In contrast, AI-companies are not just building general-purpose tools, but general-purpose agents that can follow instructions with many intermediate steps and use tools themselves, and we should expect these to get more and more autonomy over time.
Especially, those concerned with losing control over AI will highlight this difference. In the words of Eliezer Yudkowsky:
“It's smarter than you. That's it. Like all the other technologies in the past, people are trying to choose what to do with them. (…) AI is choosing what to do itself. It’s doing so using a more powerful ability to steer, it understands reality better than you do – not right now but in the predictable place it’s going -, makes better choices, goes outside the box better, has more of a spark of creativity, of invention, of like creative unexpected uses of the world around it, better at manipulating you. To it, you are an object whose rules it knows better than you understand yourself. Um, yeah, all of that is not something that's true of electricity.”
Kai Fu-Lee. (2018). AI Superpowers: China, Silicon Valley, and the New World Order. p.50
The electrical telegraph which replaced the optical telegraph preceded the electrical grid.
Richard Du Boff. (1967). The Introduction of Electric Power in American Manufacturing. The Economic History Review, 20(3), 509-518. p. 510
There are of course cross-ocean lines for electronic telecommunications starting with the transatlantic electric telegraph cables 1858/1866. However, these are using electricity to transport and sell information, they are not selling electric power.
“May” is important here, because many benefits of open-source AI could also be provided in other ways that do not involve circumventing export controls and terms of use.
Jill Jones. (2003). Empires of Light. Random House. Chapter 1.
For example, the invention of radiotelegraphy is conventionally dated to something like 1896 when Marconi filed his first patent. The British Navy adopted Marconi’s wireless telegraph as soon as 1899, and its military application was not a mystery that had to be figured out over decades, it was very straightforward. By 1902 this had already turned into one of the hotspots of German-British rivalry.