By
Stuart Kerr, Technology Correspondent,
LiveAIWire
Most AI systems are impressive at
tasks with defined goals and measurable outcomes. Set a chess-playing AI
against the world champion and it wins. Ask a language model to summarise a
document and it does so accurately. Prompt a code-generation model to write a
function with specified inputs and outputs and it delivers something correct.
None of these are trivial achievements, and they have transformed industries.
But none of them require the property that defines the most consequential
aspects of human intelligence: the capacity to pursue genuinely novel goals
that have never been specified, to discover what is worth pursuing rather
than pursue what has been specified. This property, which researchers call
open-endedness, is what Google DeepMind
researchers argued in a landmark ICML 2024 paper is the essential
missing ingredient between current AI and artificial
superintelligence.
The argument is both more specific and
more consequential than the familiar “AI can’t be truly creative”
claim. Open-endedness is not simply about producing novel outputs; it is
about continuously discovering new goals, capabilities, and environments to
explore without those discoveries being specified in advance by a human
programmer. Nature achieves open-endedness through evolution: each species or
innovation becomes the environment that drives the next round of discovery.
Human culture achieves it through the accumulation of knowledge where each
generation’s discoveries become the foundation from which the next generation
can discover further. Current AI systems, including the most powerful
frontier models, do not have this property. They are trained on fixed
datasets to optimise fixed objectives, which means their capability frontier
is bounded by the distribution of their training data and the imagination of
the humans who designed their objectives.
Why
Open-Endedness Is Hard
The challenge is not primarily
technical in the sense of requiring more compute or larger models. It is
structural: goal-directed learning systems and open-ended systems require
fundamentally different architectures and reward mechanisms. A system that
maximises a specified objective cannot simultaneously discover that the
objective itself should be revised. A system trained on a fixed dataset
cannot spontaneously discover what data it needs to expand its capabilities
beyond that dataset. Building a system that does both requires solving what
the DeepMind researchers describe as the challenge of creating a
“Cambrian explosion” of emergent capabilities, behaviours, and
artefacts without those emergences being pre-specified.
The
data scarcity problem compounds this. Current frontier models have been
trained on essentially all high-quality publicly available text and image
data. Scaling this approach further runs into the fundamental constraint that
there is no more data of this kind to consume. Open-endedness, the ability to
generate new learning environments and objectives rather than requiring
pre-existing data, is one research direction that could break through that
constraint. Rather than consuming external data, an open-ended system
generates its own learning curriculum through the process of exploration,
allowing capability to accumulate beyond what any fixed dataset could support.
What
Current AI Systems Actually Do
It is worth being precise
about what the claim of open-endedness limitation is and is not. Current AI
systems produce genuinely novel outputs in the sense of combinations of
patterns that have not appeared in that exact form before. They can extend
metaphors, combine concepts from different domains, and produce text or
images that their creators would recognise as unexpected. This is a real
capability, and it is why calling current AI systems uncreative is too
simple. The limitation is more specific: they produce novel outputs within
the distribution defined by their training, but they cannot discover that
there are entirely new distributions worth exploring. They can vary within
the space of what their training suggests is valuable; they cannot discover
that the space itself should be different.
Connecting this
to the curiosity
and compression framework that underlies intrinsically motivated AI
research helps clarify the architecture question. A truly open-ended system
would need an intrinsic drive to seek out new compressible patterns rather
than a static training objective that rewards compression of a specific
dataset. That drive does not yet exist in production AI systems in a form
that enables meaningful open-ended exploration at the capability level where
it would matter most.
The Research Directions That
Matter
The most promising current approaches to
open-endedness in AI research involve quality-diversity algorithms, which
maintain populations of diverse solutions rather than converging on a single
optimum; foundation models used as components of larger systems that can
propose and evaluate new objectives; and multi-agent systems where the
interaction between agents creates emergent complexity that neither agent’s
objective specified directly. Each approach captures some aspect of the
open-ended property without yet delivering the full version that the DeepMind
researchers describe as necessary for superhuman AI.
The
practical relevance for anyone using or building current
AI systems is calibration rather than concern. The absence of
open-endedness in current AI is not a fatal limitation for the vast majority
of commercial and professional applications, which involve well-specified
tasks where optimising a defined objective is exactly what is needed. It is a
limitation for the applications that require genuine discovery: scientific
research at the frontier, design of genuinely new products rather than variations
on existing ones, and the kind of strategic creativity that identifies what
to pursue rather than how to pursue what has been identified. Understanding how curiosity
and compression underlie AI intelligence helps frame why
open-endedness requires not just more capability but a fundamentally
different relationship between AI and its objectives. Understanding where
current
AI capabilities genuinely sit relative to these frontier questions
is more useful than either dismissing the limitation as unimportant or
treating it as evidence that AI is fundamentally
stuck.
Why the Gap Between Current AI and Open-Endedness
Matters Now
The practical relevance of the open-endedness
gap is not primarily about whether AI will surpass human creativity in some
distant future. It is about accurately calibrating which tasks AI can
currently be trusted to handle with minimal oversight and which tasks require
the kind of human judgment that open-endedness enables. Tasks that require
identifying what is worth pursuing, rather than pursuing what has been
specified, sit in the second category regardless of how sophisticated the AI
system involved is. This includes strategic planning at the level where
options are being generated rather than evaluated, genuine research frontiers
where the question is what to investigate rather than how to investigate a
specified question, and design contexts where the brief itself is what needs
to be discovered. Treating current AI as capable of open-ended discovery in
these domains is not just inaccurate; it produces a specific type of failure
where the AI confidently explores well-specified territory adjacent to the
actual problem rather than identifying that the problem specification itself
needs revision. The distinction matters most in the very high-stakes contexts
where AI is most temptingly capable: it is fluent and fast and apparently
thorough, which makes it easy to mistake exploration of the specified for
discovery of what matters.
The open-endedness frontier is
not an obstacle to using current AI well; it is a map of the boundary where
human judgment must still take the lead.
About the
Author
Stuart Kerr is Technology Correspondent at
LiveAIWire, covering artificial intelligence, cybersecurity, and the social
impact of emerging technology. He publishes daily at
LiveAIWire.com.