The Kaplansky direct-finiteness conjecture asserts that the group algebra of any group
over a field
is directly finite, meaning that
implies
for
.
Unlike the Kaplansky zero-divisor
conjecture, it does not assume that
is a torsion-free group
(Bunina et al. 2026).
The conjecture holds for sofic groups over every field (Elek and Szabó 2004). OpenAI (2026, p. 1)
constructed a counterexample in one specified odd field
characteristic. For ,
let
be the smallest prime factor
of
. Then
is odd, and there are a finite
field
with
elements, a finitely generated group
with torsion, and
with
but
.
MathVet (2026) published a successful Lean-kernel check of this construction and
its defining identities, using only Lean's standard axioms.
This example also refutes the Gottschalk
surjunctivity conjecture by giving a cellular
automaton that is injective but not surjective
on
(OpenAI 2026).
Bunina et al. (2026) report that for every odd prime number there is a finitely
presented group
with
satisfying
,
,
and
. These relations imply
, since
would give
.
The same counterexample persists over any field of field
characteristic
.
The claimed construction would also provide a group that
is not a sofic group.
This broader argument depends on substantial geometric and probabilistic inputs from earlier OpenAI manuscripts. The formal check of OpenAI's specified-prime example does not establish this broader claim.
Mghirbi and Douglas (2026) showed that for each prime number there is a finitely
presented, torsion-free hyperbolic
group
with a finite two-dimensional classifying
space whose group algebra
contains elements
with
over every field
of field characteristic
. Their characteristic-2 result establishes hyperbolicity
for groups from an earlier OpenAI construction. For odd
they construct new groups by adapting
its cancellation argument. The authors report that AI models worked out the exact
color-balance solution and transferred the random-graph estimates to odd field
characteristic. The authors report a Lean formalization of the existence theorem
using only standard axioms. The formalization uses the
random-graph estimates in existence form and does not establish all of their probabilistic
results.