AI as Normal Technology

# What is Normal?

not to understate impact or potential for disruption: examples electricity and the internet are "normal"
contrasts both the utopian and dystopian visions of the future with AI that treat it as a separate autonomous and super-intelligent species.

"AI is normal technology" means:

description of current AI
prediction about foreseeable future of AI
a prescription for how we should treat it

Focus on the relationship between technology and society. Reject technological determinism, specifically AI itself as an agent in determining its future. Draw from historical record for technological revolutions, including the slow and uncertain adoption and diffusion of technology. Assume a continuity between the past and future trajectories in terms of their societal impact and the role of institutions in shaping these trajectories.

Pace of play: if we reject "fast takeoff" scenarios it's not valuable to imagine "what comes after" until we achieve some level of the next evolution.

# Part I: A Slow Economic and Societal Transformation

Timescale of decades a critical distinction between AI methods, AI applications, and AI adoption, arguing that the three happen at different timescales.

The impact of general-purpose technologies is not measured when they are introduced or improved but rather applied and diffused through productive sectors of the economy.

safety-critical industries have been slow to adopt AI.
rely on old, proven, deterministic tools like statistics and regression analysis
doing a poor job of developing and testing new models in complex and safety-critical areas, like healthcare
diffusion speed is limited by rate of human, organizational and institutional change
rate of diffusion is about use not availability
complex, real-world uses that cannot be simulated easily have not seen the same type of cost decreases as (examples) language translation or game-playing as we climb towards general computing applications. The "capability-reliability" gap
a lot of organizational information is tribal and must be actively pursued, vs. explicit and passive.

The innovation-diffusion feedback loop is too slow for the rapid economic impact some predict.
Even if we get rapid adoption, the goalposts for "economically valuable" work (that AGI is supposed to be able to complete) will continually move further away as low-hanging fruit is automated and humans adapt by working on harder problems, making them now valuable.
Historically AI research has seen huge peaks and values and significant "herding" around relatively few ideas. With accelerated publishing are we supplementing existing ideas or turning over established ideas? In fields in which the volume of papers is highest, it appears harder, not easier, for new ideas to break through.

# Part II: Division of Labor between Humans and AI

a potential division of labor between humans and AI in a world with advanced AI (but not “superintelligent” AI, which we view as incoherent as usually conceptualized).

control is primarily in the hands of people and organizations; indeed, a greater and greater proportion of what people do in their jobs is AI control.

Issues:

Intelligence is not well defined, and a continuum of species flawed.
Fails to focus on the power to modify the environment.

We are powerful not because of our intelligence, but because technology is used to increase our capabilities.

Humans have not historically been limited by their biology and physiology (i.e. evolution)

Two key task areas:

Forecasting - events such as geopolitical outcomes
Persuasion - convince people to act against their self-interest

If we presume superintelligence:

build a galaxy brain
keep it in a box

With human intelligence that leverages AI the control problem is much more tractable.

Look to other fields for control, example security (least privilege).

As more physical and cognitive tasks become amenable to automation, an increasing percentage of human jobs and tasks will be related to AI control.

Example: truck drivers don't just drive long haul, they monitor freight, conduct safety checks, safeguard goods, maintain the truck, process paperwork and talk to customers. They also do a lot of short hauls & yard moves. AI may do some of these with varying levels of success but unlikely to do them all.

# Part III: AI Risks with AI as Normal Technology

analyze accidents, arms races, misuse, and misalignment, and argue that viewing AI as normal technology leads to fundamentally different conclusions about mitigations compared to viewing AI as being humanlike.

Risk Types:

accidents
arms races
misuse
misalignment
non-catastrophic system risks

Arms races: safer, more responsible developers get "out competed". Races to the bottom in terms of safety are historically common. Costs are externalized until market failure requires regulation, at which point the leaders then use as a competitive moat.

In aviation AI has been held to existing standards, which has limited adoption and slowed diffusion.

Arms races are likely, but sector-specific and should be addressed as such. Need to identify the impact & harm ahead of time - example: social media recommendation

Regulation will likely move much slower than AI, even without rapid adoption.

Primary defense against misuse must be located downstream of models. Aligning models to prevent misuse has proven both challenging and brittle, so defense will be elsewhere. The potential harm of a capability is contextual, so it is very hard for a model to incorporate the entire context.

Developing an AI model that cannot be misused is like trying to build a computer that cannot be used for bad things.

AI applications for defense, like cybersecurity

Larger, broader defenses like hardening infrastructure with protect against things like AI model misalignment AND existing threats like cybersecurity that adopt AI.

This makes sense when viewing AI as normal technology. Improvements in AI will have both positive and negative use-cases, so focusing on only one is unrealistic. Even if they develop asymmetrically, trying to predict what & how is improbable.

If AI is normal technology it's the smaller - but still systemic - risks that are more important. The accumulative risks - socio, economic, environmental, political - are likely to be more impactful than the threat of super-intelligence take-over.

# Part IV: AI Policy

reducing uncertainty as a first-rate policy goal resilience as the overarching approach to catastrophic risks. drastic interventions premised on the difficulty of controlling superintelligent AI will, in fact, make things much worse if AI turns out to be normal technology

policy making under uncertainty - AI safety has polarized views

reducing uncertainty with policy should be the overall goal, vs. being overly prescriptive in terms of what you can and cannot do.

Focus on hardening defenses for unpredicted outcomes - like infrastructure

whistle-blower protection
transparency reporting requirements
incident reporting
government use inventories
product and deployment registration
safe harbour provisions to promote responsible research and vulnerability testing
funding for resilience research over safety & harm reduction
evidence gathering
liability assignment
promote open model development & disclosure
restrictions until sufficient evidence of safety
interventions that are AI-independent, and might use AI (example: general-purpose monitoring and security)

Nonproliferation is not feasible and leads to single points of failure

export controls
licensing
open-weight models

can lead to draconian and heavy-handed surveillance and monitoring, policing and enforcement reduce competition in key areas reduce cooperation is common areas (shared vulnerabilities, safety research) crates brittleness at potential failure points. ex: if model weights are "secret" then leaking them is a huge risk limits advancement to small professional class

Regulation must be tailored to domain, with higher risk (medical, insurance, transport) having tighter regulation.

Focus on understanding and potentially regulating the AI supply chain: limited set of foundational models and large downstream integrators and application builders

Nuanced view of automation; not binary but degrees of human oversight and approval

Invest in the complements of automation, things that become more attractive and valuable as AI adoption increases.

AI literacy & workforce training
digitization and open data (ex: government data sets)
improved energy infrastructure and transmission

public goods are likely to be under-invested by the private sector. Accelerate adoption and diffusion

Fix public systems before turning to AI as a panacea for broken process - looking at you CRA

# Conclusion

AI as normal technology allows us to

apply historical experiences to try and predict how adoption and diffusion will proceed
focus on the likely risks, benefits and areas that require attention
gain comfort with the transition
build a nuanced and non-binary understanding and associated policies and regulations
de-emphasize the need for predictability and dis-aggregate the potential outcomes that are unhelpful in the vast generalization
mutual understanding, not completely aligned beliefs

We need to counter the super-intelligence view with a well-articulated alternative that communicates AI is an exciting and potentially disruptive transformation, not an immediate change to all aspects of human activity.