Small print.
Wide open.

[Embedded graphic 1 — OCR text]
There are four broad categories of risk: misuse, loss of
control, authoritarian abuse, and economic disruption.

e Alcan be misused by those who wish to do harm,
for example to conduct cyberattacks or to design
biological weapons. We have strict guardrails and
trusted access programs that are designed to block
misuse of our technology while still enabling
positive applications. We regularly shut down
attempts to misuse our models, report them to law
enforcement when appropriate, and endeavor to
transparently disclose them to the public.

e AI may also behave autonomously in ways we (and
our users) do not intend—as illustrated by recent
high-profile cybersecurity incidents. Losing
control of sufficiently powerful AI systems could
be a global security threat of the first order.
Anthropic has pioneered a number of methods for
improving the control of our AI systems. These
include mechanistic interpretability, which allows
us to look inside AI models and try to understand
how they behave, and constitutional training,
which allows us to train AI models to act in line
with a specific written document—Claude’s
constitution. We also publish model system cards
that run to hundreds of pages, designed to give the
public visibility into our progress on model safety.
It has long been our practice to disclose incidents
and publish research findings that help mitigate
against this risk across the industry.

e Authoritarian governments may use AI to create
high-tech surveillance states, deepen repression,
and make themselves much more difficult to
dislodge. We are committed to working with free
societies to provide a counterweight to authoritarian
abuse and to promote freedom and democracy.

e Finally, the incredible economic value of AI may
come with economic disruption, similar to what
has happened with previous technologies but
greater in magnitude. We are committed to
studying this disruption—we maintain an
economic index that provides aggregated
information about Claude’s usage across the
economy, and our researchers build

macroeconomic models to predict the possible
future impact of AI. We also suspect more radical
actions will be needed and have already
committed $200 million to partner with
governments to run studies on different ways of
responding to changes in the economy. We do this
to find evidence about which types of
interventions are effective, so that if society does
encounter economic disruption, we will have
helped build and de-risk the right tools to use in
the response. We expect to scale up this funding
significantly in the future.

We consider it our personal responsibility to help the
world understand these risks, to prevent or mitigate
them when they arise from our AI models, and to work
with governments and the rest of the industry to
address them at a societal level.

Anthropic’s unique approach

Navigating a technology with the properties we have
just described requires a very different approach from
that of a typical company. Economists use the term
externalities for effects on society that fall outside the
transactions between a company and its customers,
that the market therefore does not fully reward or
penalize: the cost of pollution produced by a factory, for
example, or the value of a research breakthrough by
one company that other companies freely build upon.
Al has unusually large externalities, both positive and
negative. Cultivating the positive externalities, and
defending against the negative ones, is to some degree
in tension with the natural incentives of a for-profit
company, and with the standard governance structure
of a publicly traded company.

As founders and business leaders, we believe deeply in
the power of entrepreneurial capitalism to deliver
massive benefits to humanity. But we also believe that,
for a technology as consequential as AI, entrepreneurial
capitalism must be balanced and augmented by
mechanisms that protect the common good and share
the upside as widely as possible. Everything about the
way we have built Anthropic, from our culture to our
corporate structure, has been designed to ensure we can
act in the public interest when the situation demands it.