General-purpose AI models with systemic risk - a threshold of 10^25 floating-point operations (FLOP) used for training the model
A revisión of AI ACT
Why do we need rules for general-purpose AI models
Artificial Intelligence promises huge benefits to our economy and society.
General-purpose AI models play an important role in that regard, as they
can be used for a variety of tasks and therefore form the basis for a range of
downstream AI systems, used worldwide.
The AI act aims to ensure that general-purpose AI models are safe and
trustworthy.
To achieve that aim, it is crucial that providers of general-purpose AI
models ensure a good understanding of their models along the entire AI value
chain, so as to allow both to integrate such models into AI systems and to
fulfil their own obligations under the AI Act.
More specifically AI models must draw up technical documentation of
their models to make available to downstream providers and to provide upon
request to the AI Office and national competent authorities, put in place a
copyright policy, and publish a training content summary.
In addition, providers of general-purpose AI models posing systemic
risks, which may be the case either because the models are very capable or
because they have a significant impact on the internal market for other
reasons, must notify those models to the Commission, assess and mitigate systemic
risks stemming from those models, including via performing model evaluations,
reporting serious incidents, and ensuring adequate cybersecurity of these
models.
General-purpose AI models
The AI Act defines a general-purpose AI model
as “an AI model, including where such an AI model is trained with a large
amount of data using self-supervision at scale, that displays significant
generality and is capable of competently performing a wide range of distinct
tasks regardless of the way the model is placed on the market and that can be
integrated into a variety of downstream systems or applications” (Article 3(63)
AI Act).
The recitals to the AI Act further clarify
which models should be deemed to display significant generality and to be
capable of performing a wide range of distinct tasks.
According to recital 98 AI Act, “whereas the
generality of a model could, inter alia, also be determined by a number of parameters,
models with at least a billion of parameters and trained with a large amount of
data using self-supervision at scale should be considered to display
significant generality and to competently perform a wide range of distinctive
tasks.”
Recital 99 AI Act adds that “large generative
AI models are a typical example for a general-purpose AI model, given that they
allow for flexible generation of content, such as in the form of text, audio,
images or video, that can readily accommodate a wide range of distinctive
tasks.”
Note that significant generality and ability
to competently perform a wide range of distinctive tasks may be achieved by
models within a single modality, such as text, audio, images, or video, if the
modality is flexible enough.
This may also be achieved by models that were developed,
fine-tuned, or otherwise modified to be particularly good at a specific task or
at a number of tasks in a specific domain.
Different stages of the development of a
model do not constitute different models.
General-purpose AI models with systemic risk
Systemic risks are risks of large-scale harm
from the most advanced (i.e. state-of-the-art) models at any given point in
time or from other models that have an equivalent impact (see Article 3(65) AI
Act).
Such risks can manifest themselves, for example, through the
lowering of barriers for chemical or biological weapons development, or
unintended issues of control over autonomous general-purpose AI models (recital
110 AI Act).
The most advanced models at any given point in time may pose
systemic risks, including novel risks, as they are pushing the state of the
art. At the same time, some models below the threshold reflecting the state of
the art may also pose systemic risks, for example through reach, scalability,
or scaffolding.
Accordingly, the AI Act classifies a
general-purpose AI model as a general-purpose AI model with systemic risk if it
is one of the most advanced models at that point in time or if it has an
equivalent impact (Article 51(1) AI Act).
Which models are considered general-purpose AI models with
systemic risk may change over time, reflecting the evolving state of the art
and potential societal adaptation to increasingly advanced models.
Currently, general-purpose AI models with systemic risk are
developed by a handful of companies, although this may change in the future.
To capture the most advanced models, i.e.
models that match or exceed the capabilities recorded in the most advanced
models so far, the AI Act lays down a threshold of 10^25
floating-point operations (FLOP) used for training the model (Article 51(1)(a)
and (2) AI Act).
Training a model that meets this threshold is currently
estimated to cost tens of millions of Euros (Epoch AI, 2024). The AI Office is continuously monitoring
technological and industrial developments, and the Commission may update the
threshold, through the adoption of a delegated act (Article 51(3) AI Act), to
ensure that it continues to single out the most advanced models as the state of
the art evolves. For example, the value of the threshold itself could be
adjusted and/or additional thresholds introduced.
To capture models with an impact equivalent
to the most advanced models, the AI Act empowers the Commission to designate
additional models as posing systemic risk, based on criteria such as the
evaluations of capabilities of the model, number of users, scalability, or
access to tools (Article 51(1)(b) and Annex XIII AI Act).

Comments
Post a Comment