Field guide

DishBrain: brain cells that play Pong

DishBrain is the 2022 experiment in which neurons grown on a chip controlled the paddle in a simple game of Pong. They got better within a few minutes. What made the difference, as far as I can tell from the paper, was what the game did to them after a miss.

Peer-reviewed sourcesUpdated 13 November 2026

The feedback rule

This is what the chip sent to the cells, as described in the methods of the paper you'll find in the sources below.

Ball position

Eight sensory electrodes, pulsed 4 to 40 times a second as the ball gets closer to the paddle.

Hit

Every electrode at once, 100 pulses a second for 100 ms, and then the game carries on as you'd expect.

Miss

Random pulses at random sites for four seconds, and then the ball starts again in a random direction.

Part 01

Who ran it

You'll find the study in Neuron, published in December 2022 after it went online on 12 October. The authors work at Cortical Labs and at universities in Melbourne and London, and one of them is Karl Friston of University College London. It's worth knowing who is behind it, because the paper says the first author works for Cortical Labs and that two of the authors own shares in the company and have an interest in related patents. In March 2025 the same company announced the CL1, a commercial unit built on the same idea, at a reported price of about $35,000 (I've compared it with FinalSpark's remote lab in FinalSpark vs Cortical Labs).

Part 02

The setup

The neurons grew on a MaxOne chip, which is a high-density multielectrode array. It has 26,400 electrodes, and the team's software read up to 1,024 of them at a time (20,000 samples a second). They used human cortical neurons grown from stem cells and also cortical cells from mouse embryos. For the mouse cultures about 800,000 cells went onto each chip.

Eight electrodes made up a sensory area. Which ones were pulsed told the cells where the ball was relative to the paddle. The rate went from 4 pulses a second when the ball was at the far wall up to 40 as it got to the paddle. Then two motor regions on the chip were read, and their activity moved the paddle up or down, so you have a closed loop between the game and the dish.

Part 03

How the cells were taught

You can't give a dish of cells a treat, so there was no reward as such. Instead the design comes from the free energy principle. That's Friston's idea that living systems act so as to keep what they sense predictable. I've put the two kinds of feedback below.

EventWhat the cells received
The paddle hit the ballOne short burst on all electrodes together, 100 pulses a second for 100 milliseconds, and the game carried on as expected
The paddle missedRandom pulses at random sites on the eight sensory electrodes, 5 a second for four seconds, then the ball restarted in a random direction

If the theory is right, the cells should drift towards whatever avoids the random noise. In this game that means hitting the ball more.

Part 04

What the results show

The authors compared the first 5 minutes of play with minutes 6 to 20. Both the human and the mouse cultures kept rallies going longer later on, while the control groups didn't improve. In the abstract they call it apparent learning within five minutes of real-time play, and I'd keep the word apparent.

Then they ran a second experiment on the feedback itself. It went on for three days with three sessions a day, 486 sessions in all.

ConditionCulturesAfter a missResult
Stimulus27Random pulses, then a restartSignificant learning, longer rallies than the other two
Silent17All input stopped for the same time, then a restartSome gain over no feedback
No feedback15The ball bounced back and play went onNo learning

So it was the consequence of a miss that mattered. When the cells only got the ball position and nothing happened after a miss, they didn't learn.

Part 05

Where the claims went too far

The title says the neurons learn and exhibit sentience. The abstract is more careful. It talks about cells that organise their activity towards a goal after sparse feedback, and calls that synthetic biological intelligence. In 2023 a group of scientists wrote to Neuron to challenge the claims of intelligence and sentience, and I think they had a point. My reading is that the cultures adapted to structured stimulation within a session. That's a real result and an interesting one, but I don't see anything in the data that shows awareness.

If you want the wider field (its limits and ethics too), read organoid intelligence explained. The energy question is on brain vs GPU power.

Part 06

Questions

What is DishBrain?

It's a system from Cortical Labs that connects living neurons on a multielectrode array to a computer game, so the cells' activity moves a paddle and the game talks back to them through electrical pulses.

Did brain cells really learn to play Pong?

Yes, in a limited sense: cultures of human and mouse neurons kept rallies going longer after the first five minutes of play while the controls didn't, though I'd treat the claim of sentience with care because other scientists challenged it in the same journal.

How were the neurons rewarded?

They weren't rewarded in the usual sense, because a hit gave them a short predictable burst of pulses and a miss gave them four seconds of random ones, and when you take that feedback away the cells don't improve.