Idea 10/Neuro & BCI/Basic science + hardware/3 min

A Wet Benchmark for Credit Assignment

Cortical organoids learn without backpropagation. That makes them the only physical test of which learning rule actually scales.

If a few hundred thousand cultured neurons can learn a goal-directed task with no global error signal, they are a decisive experimental arbiter between learning rules that no simulation can settle.

01What changed

Cortical organoids have been reported to show goal-directed learning in 2026, the first credible claim that cultured human neural tissue can be trained on an external objective rather than merely exhibiting spontaneous activity.1 Organoid learning, 2026
Biocomputing entered a more sober phase at the same time: reviews now separate demonstrated plasticity from speculative computation claims, which makes a falsifiable benchmark both possible and welcome.2 Biocomputing review, 2026
Simulated architectures for organoid intelligence have been proposed, but they encode an assumed learning rule rather than testing which one the tissue uses.3 IEEE, 2026

02The gap

Learning-rule debates are settled by simulation, and simulations are built on the assumption being tested. Biological tissue is the only system where the rule is fixed by evolution rather than by the modeller, yet nobody uses it as an arbiter.

03The idea

Design tasks that separate the candidate rules. Backpropagation needs a global error signal and symmetric weights; three-factor rules need a neuromodulatory broadcast; Hebbian rules need only correlation. Choose tasks where these make different predictions about sample efficiency, interference and the effect of disrupting specific modulators.

The key manipulation is pharmacological, not architectural. Block a neuromodulatory pathway and see which tasks break. A rule that requires a global third factor should collapse when the broadcast is interrupted, while a purely local rule should degrade gracefully. That experiment is impossible in silico and routine in a dish.

Use organoids on multi-electrode arrays with a closed-loop stimulator, and score learning as the change in a task-relevant decoding metric per unit stimulation, comparing against matched random stimulation.

The deliverable is not a computer. It is a table: for each task class, which rule family reproduces the tissue's learning curve, and where each rule fails. That table would constrain neuromorphic hardware design more than any benchmark on silicon.

04Why now

Organoids now learn, arrays are dense enough to observe population dynamics, and pharmacology gives clean perturbations. The three capabilities have not been combined in one experiment.

05The decisive test

Do this

Three task classes, four pharmacological conditions, ten organoids per cell, with matched random-stimulation controls and a pre-registered decoding metric.

Success looks like

At least one task shows the predicted dissociation, where blocking a modulator selectively destroys learning in the task that requires a global error signal.

Kills it

If learning is explained entirely by passive adaptation of the recording interface, the tissue is not learning and the benchmark measures nothing.

06What would break it

Organoid variability is large and sourcing is not standardised.
Stimulation artefacts can masquerade as plasticity.
Ethical review for goal-directed training of human tissue is still being written and may slow the work.

07If it works

Learning-rule theory gets an experimental court of appeal, and neuromorphic engineering gets a rule validated on real tissue rather than on a hypothesis.