Note: neither author has a background in history, so please take this with a lot of salt. Sam thinks this is more likely than not to contain an important error. This was written in April 2022 and we’re posting now as a draft, because the alternative is to never post.
Like electricity, AI is argued to be a general purpose technology, which will significantly shape the global economic, military and political landscapes, attracting considerable media attention and public concern. Also like electricity, AI technology has the property that whilst some use cases are innocuous, others pose varying risks of harm.
Due to these similarities, one might wonder if there are any lessons for AI governance today to be learned from the development of early electricity regulation and standards. We looked into this question for about two weeks, focusing on early electrification in the US from the late 1800s to the early 1900s, and on the UK’s nationalisation of the electricity sector during the 20th century.
This post identifies and examines lessons we found particularly interesting and relevant to AI governance. We imagine many of them will be fairly obvious to many readers, but we found that having concrete historical examples was helpful for understanding the lessons in more depth and grounding them in some empirical evidence.
In brief, the lessons we found interesting and relevant are:
- Accidents can galvanise regulation
- People co-opt accidents for their own (policy) agendas (to various degrees of success)
- Technology experts can have significant influence in dictating the direction of early standards and regulation
- Technology regulation is not inherently anti-innovation
- The optimal amount and shape of regulation can change as a technology matures
- The need for interoperability of electrical devices presented a window of opportunity for setting global standards
- The development of safety regulation can be driven by unexpected stakeholders
- Pervasive monitoring and hard constraints on individual consumption of technology is an existing and already used governance tool
There’s a lot more that could be investigated here—if you’re interested in this topic, and especially if you’re a historian interested in electricity or the early development of technology standards and regulations, we think there are a number of threads of inquiry that could be worth picking up.
Accidents can galvanise regulation
In the early days of electrification, there were several high-profile accidents resulting in deaths and economic damage:
- A lineman being electrocuted in a tangle of overhead electrical wires, above a busy lunchtime crowd in Manhattan, which included many influential New York aldermen.
- There were a number of other deaths for similar reasons, which occurred somewhat less publicly and so were less influential but still important.
- Pearl Street Station—the first commercial central power plant in the United States—burned down in 1890.
- The 1888 blizzard in New York City tore down many power lines and led to a power blackout.
Despite electric companies like Western Union and US Illuminating Company protesting regulation with court injunctions, [Hargadon & Doglas 2021] these accidents spurred government and corporate regulation around electrical safety, including:
- Various governments began to require high voltage electrical lines to be buried underground, one of the first (if not the first) governmental regulations on electricity to be introduced [Stross 2007].
- Thomson-Houston electric company developed lighting arrestors for power lines and blowout switches to shut down systems in case of a power surge [Davis 2012].
- Concerned about risks of installing AC electricity (which permitted higher voltage distribution) in residential homes, Thomson-Houston declined to do so until they could develop a more effective transformer to address the risks. Other AC companies, like Westinghouse, did not incorporate the same safety processes [Higonnet et al. 1998].
These regulations were prompted chiefly by a strong outpouring of public concern in response to the accidents. There was widespread fear that no one was safe from death by electricity, magnified by misinformation and disinformation about the hazards of electricity spread via public debate and the media. Newspaper titles like “Electric Wire Slaughter” and “Electric Murder” further fanned the flames, encouraging rapidly growing public pressure to address these safety issues [Hargadon & Doglas 2021]. As noted, one of the first government regulations regarding electricity was created in response - burying power lines in NYC to reduce the dangers of having high voltage wires out in the open streets. For a short period in time in 1889 and 1890, a majority of the city’s electricity infrastructure was actually simultaneously switched off to accomplish this, with most of NYC’s New Year’s celebrations in 1889 lit by candles and gas lamps instead.
Earlier in the decade, first efforts towards electrical regulation were occurring across the Atlantic in the UK as well, and for similar reasons. Like in the US, initial electrical infrastructure installation in the UK was completely unregulated, with private electrical providers tearing up roads and stringing up wires based on their own methods and systems. While the same level of high profile accidents didn’t occur, public concerns about the fire hazards that many of these early wirings posed, and buoyed by widespread fears around the dangers of electricity more generally, the UK Parliament in 1882 requested that the Society of Telegraph Engineers and of Electricians develop “Rules and Regulations for the Prevention of Fire Risks Arising from Electrical Lighting”, which later became the foundation of today’s UK electrical wiring code [Freeberg 2014].
Why should we care: Electrical accidents and hazards led to heightened public concern about electricity which prompted regulatory responses in both the corporate and government spaces, some of which might not have otherwise occurred without that push. Likewise, the first high-profile cases of AI systems causing death or economic or societal damage will likely present regulatory windows of opportunity, and potentially also reduce the barriers to more comprehensive or stronger regulation than would ordinarily be possible. It’s also worth noting that while this could facilitate governance interventions that reduce AI risk, it could also lead to uninformed or imprecise regulations that are detrimental. For instance, the regulatory response to alignment “warning shots” could be focused on the particular company, use case or kind of model that causes harm—rather than the general causes and risk factors of alignment failure. This is especially worrying in light of the EU AI Act regulating AI systems based on their use case. Therefore, it could be valuable to anticipate and prepare for crises in advance, to help ensure that whatever regulation ensues is as sensitive as possible to humanity’s long-term prospects.
People co-opt accidents for their own (policy) agendas (to various degrees of success)
Whilst accidents present a window of regulatory opportunity for actors who have public welfare at heart, they do the same for those who have private interests.
This was especially true for electricity: a new and mysterious technology where the sources of risk (and therefore ideal governance interventions) were not well-understood. This created an environment where various actors could make arguments about sources of risk and for appropriate interventions that favoured their personal interests, generating more public attention and interest than they might have had otherwise.
For instance, the deaths and fires which occurred in New York were used by Edison Lighting Company to argue for the superiority of direct current (DC) over alternating current (AC) as the safer electrical current type —when in reality neither AC nor DC is inherently more dangerous and the “correct” conclusion is that both forms of high voltage power lines should be undergrounded. Yet Edison didn’t argue for undergrounding at all: "Burying these wires will result only in the transfer of deaths to man-holes, houses, stores, and offices, through the agency of the telephone, the low-pressure systems, and the apparatus of the high-tension current itself … My personal desire would be to prohibit entirely the use of alternating currents. They are as unnecessary as they are dangerous" [Essig 2005]. It seems that Edison’s motivation was to use the accidents to argue for regulation that would force competitor companies using AC out of the market (Edison Lighting Company used DC).
In addition, gas utility companies, suffering from the surge in popularity of electricity, were also vocal about the perceived risks of the “mystic nerve energy of electric wires” [Wallace 2018]. They helped to stoke public fears, encouraging a return to gas lighting by citing it as the safer alternative, despite gas power lamps leading to higher fire and fatality rates than electric ones.
In this case, neither Edison nor the gas utility companies was successful in their goals to stymie their competition over the long term. However, in the short term, their encouragement of fears around electricity were distracting and unhelpful, generating fear and misunderstandings around electrical hazards. For example, the Tribute magazine warned that "Mr. Edison has since declared that any metallic object—a doorknob, a railing, a gas fixture, the most common and necessary appliance of life—might at any moment become the medium of death". Accordingly, some New Yorkers refused to have doorbell wires in their homes, and the Evening Post remarked that "One scarcely ventures to put a latch key into his own door." Another source argued that the only solution was not to insulate electrical wires, but to have universal limits on voltage. One could easily imagine that if these parties had more momentum and support, it could have led to misguided regulations that made developing safe and effective electrical technology less likely.
Why should we care: It’s worth remembering that others will also seek to advance their goals during windows of opportunity resulting from accidents or other incidents that heighten public concern, and that their regulatory proposals may be irrelevant or detrimental to AI existential safety. Candidates for these other parties include those who stand to gain or lose significant economic, political or military advantage on the basis of AI development and deployment decisions. Being aware of these other potential parties and understanding how they may seek to co-opt AI accidents to further their own concerns would be helpful if you wanted to counter their proposals, or adapt them in a way that increases existential safety.
Tech experts can have significant influence in dictating the direction of early standards and regulation
Some of the earliest electricity best practices (about e.g. voltage levels, energy distribution models, current type) were disseminated via training schools set up by the first electrical companies, who also wrote the first installation and maintenance documents and distributed them throughout the industry. Edison and his electric company devoted a particular amount of attention to this knowledge transfer effort, resulting in the professionals they trained later becoming highly influential figures throughout the industry. In addition, Edison’s manual on electricity became popular to refer to, and the best practices he recommended were widely emulated by other companies. These also heavily influenced many of the later, more formalised educational efforts, such as the first electrical engineering degrees offered by academic institutions, most of which were in the US and Germany.
Early professional expertise was also directly sought when developing governmental and corporate regulations and standards, such as deciding on the preferred method for burying electrical wires and identifying the safest devices and processes in operating electrical lighting. Letters between Edison and his associates show that their opinions were sought after for developing regulations relating to electrical standards by other industries as well, such as the insurance sector. Meanwhile, as mentioned previously, the first version of the UK wiring code was directly solicited from professional organisations by Parliament in 1882. As the electrical industry began to expand, early professional organisations such as the AIEE and NELA were created, having substantial influence in determining standards and regulations. They and other experts were strongly influential in many of the first all-encompassing industry wide standards and regulations, such as the world’s first corporate safety codes. For instance, the 1897 National Electric Code, which later became the foundation for the first national US governmental electric regulations, was developed together by professional organisations and experts from around the globe.
Why should we care: Early electrical experts had a huge influence on the early regulatory environment, which in turn had a big influence on later regulations up to and including the present. This suggests that with similar technologies, not only can experts dictate a significant part of the direction of both technological regulation and subsequently its development, but that these impacts are likely to be higher before the governance landscape matures. If this is equally the case in AI, wherein not only can AI experts make a significant contribution towards shaping the technology’s regulations and standards, but also they can contribute much more early on, then it becomes especially crucial for them to both encourage the sharing of best practices and become involved with governance now, but to also think about how standards and regulations can be as agile as possible in the future.
Technology regulation is not inherently anti-innovation
The structure of the electricity industry in the UK saw two major shifts: from a decentralised, uncoordinated early period (up until 1926), to increasing nationalisation of electricity transmission and then distribution (between 1926 and 1990), to complete privatisation (from 1990 onwards).
The reasons for nationalisation are interesting. In the early period, the UK fell behind best practice in the US, where the price of electricity was lower thanks to economies of scale from larger generating stations. But to benefit from larger generating stations, you need an electricity grid—which also allows you to cope with region-specific supply fluctuations.
However, in a decentralised, uncoordinated market, electricity grids don’t tend to get built. So a statutory corporation, called the Central Electricity Board (CEB), was set up to build the grid. It was modelled on BBC, acting more like a commercial enterprise than a nationalised industry: it had considerable autonomy, paid high salaries, and was financed by fixed-interest loans that were not guaranteed by the government. It also promoted national competition/innovation in generation by operating a "merit order" (taking electricity preferentially from the more efficient plants). The UK had caught up with best practice in the US 9 years after the CEB was founded. Around 1928, central public ownership was extended to the distribution of electricity as well as its transmission, for similar reasons.
One lesson here is that technology regulation is not inherently anti-innovation. The nationalisation of electricity generation enabled the grid to be built, which increased the UK’s national competitiveness, allowing economies of scale in generation and helping to better cope with region-specific electricity supply fluctuations.
Why should we care: One narrative that’s been floating around in the AI policy world for a while is that “we can’t regulate because our competitors won’t; it would put us out of the running for being a leader in AI”. The argument is that regulation is a restriction that diminishes the chances to not only innovate but ‘win’ against competition. However, it’s important to note that there are plausible regulatory proposals that have the opposite effect. More monitoring of AI systems on safety-relevant metrics can help to gamify and hence accelerate innovation in AI safety, for example (cf. prestige races). And in general, monitoring can give policymakers more information, allowing them to design more targeted but lighter touch regulation.
One might also want to make a normative point here: devising regulatory proposals which are, or at least are framed as, pro-innovation likely helps increase their political feasibility.
This is particularly relevant in the UK at the moment, as its national AI strategy is heavily pro-innovation. This attitude might be encouraging deregulation in other areas, such as personal digital data access and privacy regulations, for example by walking back GDPR.
The optimal amount and shape of regulation can change as a technology matures
As of 1990, the publicly owned corporations involved in UK electricity generation ended up being divided and listed on the market. Why? Whilst public ownership had a comparative advantage in mobilising investment to ensure coordinated expansion during the early period, once the grid was built and private entities experienced the advantages of coordination, other considerations became more salient. For instance: classic bureaucratic inefficiencies led to overly high running costs, attempts to effectively steer the nationalised transmission/distribution companies were hampered by lobbying, the centrally planned responses to demand shocks in the 1970s were inefficient, and an unreasonable number of internationally uncompetitive British firms were kept alive due to pressure from industrial policy.
Early in the new century however, as the nation experienced some of the negative consequences that came from embracing an industrial economy, the federal government showed more interest in lighting policy. That included broadly based policies such as antitrust actions, and lighting specific policies such as blackouts and other use restrictions when the nation went to war.
A loose alliance of reformers united by a desire to rid the government of cronyism and inefficiency, Progressives “created a code of professional public administration” to advance those goals. That included a rationalist approach to policy that influenced planning and laid the groundwork for interventions in areas such as resource management. Establishment of a National Bureau of Standards (1901) brought professionals into government to promote economic efficiency in a way that would assist, not compete with, the private sector. At the same time, another area of Progressive concern centred on the emergence of large corporations perceived as a threat to basic American values. Progressives pursued antitrust actions as one way to control large private sector actors. Lighting industry consolidation drew their attention as General Electric’s ability to control the market and influence the political process grew.
Why should we care: As it was with electricity, the development and deployment of AI technology will (hopefully) not unfold overnight, though it’s likely to move much faster than electricity. Use cases will grow. Safety challenges may change. Different resources may become bottlenecks to progress. Policymakers should be wary that there is unlikely to be a one-size-fits-all regulation that is optimal over all these development/deployment phases. This highlights the importance of adaptable/agile regulation, especially one that scales with AI capabilities. Throwing out our previous understanding of the best way to regulate is not without precedent, and it might not be because the former regulation was wrong, merely that the technology has matured. Instead, to identify the optimum shape of regulation, it’s crucial to examine and understand the broader geopolitical and economic landscape.
The need for interoperability of electrical devices presented a window of opportunity for setting global standards
Initially, electrical devices and machinery developed in different operating regions used different voltages, frequencies, fittings, and so on. The need for standardisation, to facilitate global trade, became recognised, and so the International Electrotechnical Commission (IEC) was founded in 1906. The IEC seems like a surprisingly long-lasting institution: it is still active today and has 207 committees for developing standards for everything from switches to lasers.
Why should we care: With AI, it’s unclear exactly what kind of standardisation will be called for, but it might require various classes of foundation models to have standardised input and output spaces. This could present an opportunity for baking certain kinds of safety features into the standards.
The development of safety regulation can be driven by unexpected stakeholders
Another interesting finding was regarding how influential the insurance sector was on the early development of US electrical safety codes and standards. In fact, one of the earliest installation best practices published was issued in 1881, by the New York Board of Fire Underwriters, which became the foundation for many later state and federal electrical codes. They incentivised adoption of these standards through selling fire insurance premiums at a more inexpensive rate when the building in question adhered to proposed safety codes and used approved products.
Soon after, the newly created Underwriters Laboratories (UL) was one of the first places to specifically engage in safety testing of electric installations, including panels, sockets, wires, switches, bulbs, and other parts. In 1898, UL published a list of “approved” devices and fittings, which received a UL Mark to denote their safety. Today, the UL Mark is the most recognisable certification mark on products across the US.
Why should we care: Levers or pathways for AI governance may be broader and more multifaceted than initially expected. While most of AI governance research today focuses on governance arising from countries or the AI industry itself, it’s worth thinking about how other intersecting fields may be able to leverage influence over AI safety standards and regulations, and what that might look like.
Pervasive monitoring and hard constraints on individual consumption of technology is an existing and already used governance tool
Moving away from history and towards more general remarks on electricity governance, it’s interesting to note that there exists pervasive monitoring of and hard constraints on the consumption of electricity. For instance, at the household level, there is a device that measures how much electricity you use, which is installed and monitored by a certified professional. Likewise, one can only draw up to a certain amount of current, as determined by a fuse box (which, again, must be installed by a certified professional). One can install solar panels to generate additional electricity, but only up to a certain power level. At the commercial and industrial levels, the constraints are higher, as is the associated level of responsibility.
Why should we care: We’re far from having monitoring or hard constraints on resources aimed at AI risk reduction—for instance compute monitoring. But the example of electricity suggests that such a thing isn’t fully and permanently outside the Overton window, and there already exist methods in the world for implementing this. Drawing an analogy to electricity may help to make the case for e.g. compute monitoring being a reasonable governance intervention.
Additional things we thought were neat, but didn’t explore further
- Visibility of technology has a significant influence in shaping public perception of said technology: Electric lighting served as a visual representation of the opportunities and dangers of electric technology. Exhibitions, fairs, and luxury uses of lighting showcased the benefits and attractiveness of lighting, playing an instrumental role in encouraging its adoption. Conversely, the highly visible hazards it posed, such as fires and deaths, lead to fears around electricity.
- With enough public pressure strong anti-technology responses are possible, even when inconvenient: Negative public reactions in New York City to the electrical accidents in the late 19th century lead to temporarily turning off the electricity lighting in much of the city.
- Early ‘legacy systems’ can be sticky, even when inconvenient: Even after AC ‘won’ the war of the currents and electrical infrastructure developed and evolved, some early adopters of DC continued to use it, even when it was less safe compared to newer alternatives and required additional operational and economic costs to run.
- It always comes back to money: In the early era of US electrical governance, most of the standards and regulations developed were rooted in financial concerns. Energy suppliers (and their board of directors), insurance companies, expert associations - many of their decisions were made in consideration of operational costs and existing or speculative financial returns. Even when regulations were developed explicitly in response to safety concerns, such as US fire underwriters’ wiring codes for example, the underlying motivation for this was to reduce their potential monetary burden.
Thanks to Shahar Avin for raising this as a research area and for useful discussion, Julian Hazell for feedback, and Justis Mills for line editing.
These periods were part of the first era of electrification, when electricity generation and distribution systems were gradually deployed in the US, Britain, and other high-income countries, between the mid-1880s and ~1950.
Another aspect we hoped to focus on, but didn’t end up having time for, was the early development of international regulation around electricity, such as plug standardisation or international energy trading.
Fun examples of misinformation: "Death does not stop at the door," one expert said, "but comes right into the house, and perhaps as you are closing a door or turning on the gas you are killed. It is likely that many of the cases of sudden death we hear of from heart disease may come about this way."
"There is no safety, and death lurks all around us," another expert warned. "A man ringing a door-bell or leaning up against a lamp post might be struck dead any instant." [Essig 2005, ch. 17]
One reason for this is roughly: if you’re a private operator considering sinking significant capital into a utility whose prices you expect to be regulated, you want a strong and legally enforceable guarantee of a fair rate of return on your investment. But if you’re guaranteed this, then you can’t lose, giving you little incentive to invest prudently and otherwise be efficient—which will make regulators leery. Given this tricky regulatory problem, it seems that just nationalising electricity transmission was a good fix.
Two handwavy reasons to expect AI development and deployment to move faster than electricity: inference is much cheaper than training, so once trained, models can diffuse quickly; advanced AI will likely be useful in accelerating AI capabilities---whereas such R&D feedback loops that weren't present in the case of electricity.