Can a dish of brain cells be intelligent? Organoid intelligence, every experiment counted

01 Key findings

1. We found 23 experiments, 13 of them since 2025. 19 are in journals and 4 are preprints; 15 involve cultured rodent or human neurons, 7 brain organoids and 1 a mixed set.IQ Metrics ledger
2. No dish is large. Since 2015 each started with 70,000 to 1,000,000 cells, between half a fruit-fly brain and 7.2, and under 1.5% of a mouse brain (67,870,000 neurons).Papers; Herculano-Houzel 2011; IQ Metrics calculation
3. The CL1’s neuron count is not on the maker’s page. The press numbers of about 800,000 and 200,000 match cells plated in two Cortical Labs papers about other chips; two 2026 preprints on CL1 chips print 500,000 cells per chip.Kagan 2022; Watmuff 2025; Loeffler 2026; Abu-Bonsrah 2026
4. The Pong result is real but narrow. In 9 mouse and 11 human cultures rally length rose within 20 minutes and did not in media-only, rest or in-silico controls; the word “sentience” is used in a formal sense, and a 2026 Cortical Labs preprint warns that stimulation artifacts can inflate accuracy.Kagan 2022; Loeffler 2026
5. “A million times less energy” is a ratio, not a measurement. It is 21 million watts of a supercomputer against a brain, for different tasks; we found 0 measured per-task comparisons.Smirnova 2023; Elsworth 2025; IQ Metrics
6. 10 of 20 graded experiments pass all five checks; we found none repeated by another lab. Grades: A 10, B 1, C 6, D 3, plus 3 old papers read only as abstracts.IQ Metrics grading rule

02 Statistics to quote

03 What is organoid intelligence?

Organoid intelligence is the idea of computing and learning with living human brain cells grown in a dish; we found 23 experiments on living neurons between 2001 and 2026. They put cells through a task, from steering a virtual animal in 2001 to a humanoid robot in 2026 ^{[3] [25]}. The field names its own goal “organoid intelligence”, and its roadmap paper calls today’s organoids small: under 500 micrometres across and under 100,000 cells, with a target of about 10 million neural cells ^[28]. A related company-led line calls the same idea “synthetic biological intelligence” ^[1].

Brain organoid

A pea-sized ball of brain-like tissue grown from stem cells. It has several cell types and some layering, but no body, no blood supply and no senses; the ones used for computing are about 500 micrometres across ^{[28] [2]}.

Cultured neurons (in vitro neurons)

Brain cells taken from a rodent embryo, or made from human stem cells, and spread flat on a chip. They wire themselves into a network within days ^[7].

Multielectrode array (MEA)

A glass or silicon plate with tens to thousands of tiny electrodes under the cells. It records their electrical spikes and delivers small pulses. The CL1 chips have 64 electrodes; the DishBrain chip had 26,400 ^{[1] [24]}.

Closed loop

The cells’ activity changes a simulated world (a paddle, a cart, a robot) and the world’s state is sent back as stimulation, many times a second ^[3].

Reservoir computing

Treating a network as a black box that turns an input pulse pattern into a rich response, then training only a simple read-out on top. It tests whether the network transforms information, not whether it “learns” ^{[11] [22]}.

Wetware, biocomputing

Computing with living cells instead of silicon. “Wetware” is the lab slang; “biocomputing” is the term used by the platform companies ^[2].

Brain-computer interface (BCI)

A device that links a person’s brain to a computer. It is not organoid computing: the brain belongs to a patient and the aim is to restore a function. Our count of people with implants is in the brain implant study.

Synthetic biological intelligence (SBI)

Cortical Labs’ name for neurons on a chip embodied in a simulated world ^[1].

04 How many experiments have used living neurons to compute?

23 experiments between 2001 and 2026: 19 in journals and 4 preprints, from 8 countries, and 13 of them since the start of 2025. The first wave (2001 to 2008) showed that a culture on an electrode array can drive a virtual animal ^[3] and shift its response with repeated stimulation ^{[4] [5]}, and that cultures can filter and recognise images ^[6]. Then came a gap of seven years in our data, a 2015 Tokyo experiment on sorting mixed signals ^[7], and the 2022 DishBrain Pong paper ^[1], which started the present wave. Organoids first enter the list in 2023 with Brainoware ^[22].

Figure 1

Twenty-five years of living neurons that compute: 23 experiments, 13 of them since 2025

Each dot is a paper or a product date; hover for the lab and the result. Cortical Labs and FinalSpark have their own rows because they sell platforms. The first 4 experiments span 2001 to 2008, then nothing in our data until 2015. As of 3 November 2026.

CL1 page live (archived 4 March 2025)Peer-reviewed paperPreprintPlatform or productEthics, roadmap or critique

Rodent neuron cultures11 experimentsShahaf and MaromBakkumIsomura 2015SumiShao

Human and mouse organoids7 experimentsBrainoware preprintBraille

Cortical Labs (Melbourne)DishBrain, CL1DishBrainCL1 digits

FinalSpark (Vevey)NeuroplatformEnergy claim: one billion times

Ethics and critiquewho says whatUS National Academies reportKosik

200120052010201520202025

Growth is steep but young. 13 of 23 experiments are from 2025 and 2026, and 4 of those are preprints that have not yet been through peer review ^{[24] [25] [23] [26]}. Six come from labs tied to one company, Cortical Labs; the others are university groups in 7 countries.

Figure 2

More than half of all experiments appeared in the last two years

Experiments on living neurons by year of the first paper, by kind of tissue. 13 of 23 are from 2025 or 2026; there are none in 2024 in our data (the FinalSpark platform paper and two preprints of that year are counted at their later journal dates). Organoids first appear in 2023.

Cultured neuronsOrganoidsMixed

2'01

'02

'03

'04

1'05

'06

'07

1'08

'09

'10

'11

'12

'13

'14

1'15

'16

'17

'18

'19

'20

'21

1'22

4'23

'24

7'25

6'26

Figure 3

Where the experiments come from: the USA, Australia and Japan lead

Experiments by country of the lead laboratory. 6 of 23 involve a company (all of them Cortical Labs); the rest are university laboratories. Australia's 5 are the Melbourne company's work with Monash, Melbourne and other universities.

USA8

Australia5

Japan4

China2

Israel1

Italy1

Switzerland1

United Kingdom1

05 How many neurons is a "mini-brain"? Brain organoids against animals

Between 70,000 and 1,000,000 cells per dish since 2015: about half a fruit-fly brain to 7.2 fruit-fly brains (a fly has 139,255 neurons ^[49]), under 1.5% of a house mouse brain (67,870,000 neurons ^[50]). The numbers are cells plated or seeded, not neurons counted afterwards: Isomura seeded 500,000 cells per array ^[7], DishBrain about 800,000 mouse or 1 million human cells ^[1], and a 2025 drug study 200,000 neurons plus 50,000 astrocytes ^[13]. Many studies print no count at all, including Brainoware ^[22], the Braille organoids ^[19] and the FinalSpark platform paper ^[2].

Figure 4

How many cells did the dish start with? Since 2015, between 70,000 and 1,000,000

Each dot is one experiment: cells plated per dish or chip (blue) or neurons read out (orange), by year of the paper, on a log scale. The bottom row (2000) is not data: it shows whole-brain neuron counts of animals for scale. Plated cells are not counted neurons; the dishes hold 0.5 to 7.2 fruit-fly brains and under 1.5% of a mouse brain.

Cells plated or seededNeurons read out or countedStated targetAnimals, for scale

Year of the paper (bottom row: animals)

20052010201520202025

DishBrain 2022Isomura 2015CL1 chips 2026Sumi 2023Target 10MWormFlyReptileMouse

101001k10k100k1M10M100M

Cells or neurons (log scale)

Figure 5

A dish of brain cells holds up to 7.2 fruit-fly brains

Cells plated per dish or chip, divided by the 139,255 neurons of one fruit-fly brain (1 = one fly). A house mouse brain would be 487 on this scale. These are cells put in the dish, not neurons counted afterwards, and organoids and cultures have no body, no organisation and no senses.

Habibollahi 2023, 1 million cells7.18 flies

DishBrain 2022, 0.8 to 1 million6.46 flies

Isomura 2015, 500,0003.59 flies

CL1 chips 2026, 500,0003.59 flies

Drug study 2025, 200,0001.44 flies

Shao 2025, 80,000 to 135,0000.77 flies

Kuchler 2025, 70,0000.50 flies

Sono 2026, 1,869 neurons0.01 flies

Compared with animals the picture is small. A worm has 302 neurons ^[48], a fruit fly 139,255 ^[49], the smallest reptile in our animal data 1.77 million ^[52], a house mouse 67,870,000 and a human 86 billion ^[51]. The roadmap target of 10 million cells would still be about a seventh of a mouse brain ^[28]. More comparisons are in our study of 222 animal brains and, for extinct animals, in how smart dinosaurs were.

Neuron count is a poor stand-in for ability. A dish has no organisation, no body and no senses, and only a few dozen electrodes to read it. In the fly brain every one of 139,255 neurons is wired into a known circuit ^[49]; in a dish the cells connect at random. For the same reason, we do not give any dish an IQ: see the last section.

06 How many neurons are in a Cortical Labs CL1?

Cortical Labs does not say: its CL1 page prints no neuron count, price or power figure (live on 3 November 2026; same in copies archived on 4 March and 11 July 2025) ^[39]. Press coverage of the 2025 launch gives about 800,000 neurons in some places and about 200,000 in others. Each number matches a Cortical Labs paper about a different chip: about 800,000 mouse cells plated for DishBrain in 2022 ^[1], and 200,000 NGN2 neurons (plus 50,000 astrocytes) in a 2025 drug study ^[13]. The two 2026 preprints that actually used a CL1 system print 500,000 cells per 64-electrode chip ^{[24] [27]}.

Figure 6

Where do 800,000 and 200,000 come from? Every number in a primary source

Cells plated per chip, as printed in Cortical Labs papers. The company's own CL1 page gives no count. The press figures of roughly 200,000 and 800,000 each match a paper about a different chip; the two preprints that used a CL1 system print 500,000 cells per chip (64 electrodes). Cells plated are not living neurons counted.

Lower figureHigher figure

DishBrain 2022 (other chip): mouse, then human cells800,0001,000,000

Drug study 2025 (other chip): NGN2 neurons, then with astrocytes200,000250,000

CL1 chips, Loeffler 2026 preprint500,000

CL1 chips, Abu-Bonsrah 2026 preprint500,000

Press reports of the CL1: the two figures200,000800,000

0200,000400,000600,000800,0001,000,000

Cells or neurons per chip

Which figure does this page use? The 500,000 cells per chip from the 2026 CL1 preprints, labelled as cells seeded and not as a company specification, because they are the only numbers tied to the product’s chips. A seeded cell is not a living neuron: the 2022 paper plated 800,000 cells and does not count the neurons that survived ^[1]. The CL API paper describes the system’s software and 64-channel examples, with no neuron count ^[38].

Figure 7

How "the CL1 has about 800,000 neurons" travelled

Each step in time order. Steps are evenly spaced. None of the steps is a Cortical Labs statement about the CL1 itself.

Where the number beganMass mediaFact-checkMeasured by a study

About 800,000 neurons in the CL1Verdict: no company source

1. 2022DishBrain paper: about 800,000 mouse cells plated on a 26,400-electrode chipKagan et al. 2022, Methods: cells plated, not neurons counted
2. 2025CL1 launch coverage repeats 800,000 (and, elsewhere, 200,000)news coverage we saw; not used as a source
3. 2025Drug study: 200,000 NGN2 neurons plus 50,000 astrocytes per chipWatmuff et al.; a DishBrain chip, not the CL1
4. 2026Two preprints on CL1 chips: 500,000 cells per 64-electrode chipAbu-Bonsrah et al.; Loeffler et al.
5. 2026CL1 page, 3 November 2026: no neuron count, no priceSame on the copies archived on 4 March and 11 July 2025

07 Did brain cells really learn to play Pong?

Partly: in the 2022 DishBrain study, rally length rose within 20 minutes in 9 mouse and 11 human cultures, but not in controls ^[1]. The controls were media-only arrays (6), rest sessions (20) and an in-silico run (3 seeds), over 399 test sessions. Human cultures started worse than the controls and ended better. The game is a simplified Pong: ball position is sent as stimulation on 8 electrodes, two motor regions move the paddle, and a hit gives predictable stimulation, a miss unpredictable ^[1].

Follow-up work widens the picture. A 2023 analysis of 14 cultures found critical dynamics during gameplay but not rest, and that criticality alone did not give learning ^[9]. A 2025 study reported that, in a 20-minute real-time window, mouse and human cultures reached more hits per rally than three deep reinforcement-learning algorithms ^[10]. A 2025 drug study found that carbamazepine at 200 uM improved gameplay ^[13]. All three come from the same company’s group: we found no other laboratory that has repeated the Pong result.

Critics have two main points. First, the word. The 2022 paper uses “sentience” in a formal sense, as being “responsive to sensory impressions” through adaptive internal processes ^[1]. A response by 30 authors disputed claims of emergent intelligence and sentience in a dish, and the authors replied in the same issue of Neuron ^{[33] [34]}; we read the titles and venue, not the letters. Second, controls. A 2024 paper reported that an electro-active polymer hydrogel, which has no cells, exhibits emergent memory when embodied in a simulated game, and the cartpole authors cite it as a reason for careful controls ^{[37] [18]}. A 2026 Cortical Labs preprint found that an empty saline dish classified digits above chance when stimulation artifacts were not blanked, so some apparent accuracy came from the stimulus itself ^[24].

08 How good is the evidence? Our A to D grades

Of 20 graded experiments, 10 report all five of our checks (grade A) and we found none repeated by another laboratory. The rest: 1 grade B, 6 grade C and 3 grade D; three old papers we could read only as abstracts are not graded. The rule uses five yes or no checks observed in the text: peer review, a control condition, 3 or more independent cultures or organoids, data or code in a public repository, and a number with a test or interval. A = all five; B = peer review, control and replicates plus at least one more; C = at least three checks; D = fewer. A grade measures how completely a result is reported, not whether the claim is true.

Figure 8

Which of the five checks does each experiment pass? Only 10 of 20 pass all

1 = yes, 0 = no, from the text we read. Rows are sorted by grade (letter before the name). The three abstract-only papers (Shahaf 2001, Ruaro 2005, Bakkum 2008) are not graded. A grade says how completely a result is reported; it does not say the claim is true.

NoYes

JournalPrControlCt3+ unitsnDataDaStatStA: Isomura 201511111A: DishBrain 202211111A: Habibollahi 202311111A: Isomura 202311111A: Sumi 202311111A: Alam El Din 202511111A: Khajehnejad 202511111A: Watmuff 202511111A: Cartpole 202611111A: Sono 202611111B: Brainoware 202311101C: Deng 202511001C: Kuchler 202510110C: Shao 202510111C: Braille 202610111C: CL1 digits 202601101C: Brainobot 202601101D: DeMarse 200110000D: Paired organoids 202501000D: Brainopheno 202601000

Figure 9

All 23 experiments by evidence grade

One square per experiment, in order of year within each grade. Rule: A = all five checks; B = peer-reviewed with control and 3 or more units; C = three or four checks; D = fewer than three.

Grade A: all five checks 10Grade B 1Grade C 6Grade D 3Not graded (abstract only) 3

23 experimentsone square = one experiment

Of the 10 experiments that give cells a learning or closed-loop task, 7 are peer-reviewed with a control condition ^{[7] [1] [17]}. Most grade A results are rodent cultures or the DishBrain family; the organoid results are younger and less often graded A: Johns Hopkins’ plasticity study ^[12] and the UC Santa Cruz cartpole study, which also found that gains did not last through a 45-minute rest ^[17]. Brainoware gets grade B because the preprint we read states no data location (the published paper may) ^[22]. Organoid studies test fewer units (median 5, against 20 for cultures) because each organoid is grown one at a time.

Figure 10

How many independent cultures or organoids did each experiment test? Cultures: median 20; organoids: median 5

Each dot is one experiment that prints a single count of independent cultures, organoids, chips or networks (13 of 23 do). Boxes: the middle half; whiskers: lowest and highest. Organoids are made one at a time and are harder to replicate in number.

Middle halfMedianWhiskersexperiment

Cultured neurons9 experiments20

Organoids4 experiments5

051015202530

Figure 11

From tissue to task to grade: where the well-reported results are

23 experiments. Left: kind of living tissue. Middle: what it was asked to do. Right: evidence grade. Organoids are 7 of 23 experiments and 2 of them reach grade A.

Living tissueTissueTaskTaskEvidence gradeGrade

Cultured neuronsCultured neurons 15OrganoidOrganoid 7Mixed cultures and organoidsMixed 1

Learning by stimulationStimulation 3Closed-loop game or controlClosed loop 9Reservoir computing and classificationReservoir 10Plasticity assayPlasticity 1

Grade AGrade A 10Grade BGrade B 1Grade CGrade C 6Grade DGrade D 3Not gradedNot graded 3

Independent replication is the missing column. The two Tokyo papers are by one laboratory ^{[7] [8]}; Brainoware’s and Brainobot’s authors overlap ^{[22] [25]}; every DishBrain paper shares authors with Cortical Labs. A framework published in 2026 proposes matching words such as learning, memory and adaptation to specific controls, because the same output can come from passive material physics or an interface ^[36]. In our funnel of 10 headline claims, 6 trace to a peer-reviewed paper, 4 to one with a control, and none to a second laboratory.

Figure 12

Of 10 headline claims, 6 trace to a peer-reviewed paper and none to a second laboratory

Ten claims that circulate in press and marketing (listed in the CSV). Stage 2: the claim traces to a peer-reviewed paper reporting that result. Stage 3: that paper has a control condition. Stage 4: another laboratory has repeated the result. "A million times less energy" and "800,000 neurons" trace to no measurement.

Headline claims we traced10all

60.0% of the stage above

Trace to a peer-reviewed paper660.0% of all

66.7% of the stage above

…with a control condition in that paper440.0% of all

0.0% of the stage above

…repeated by another laboratory00.0% of all

09 Do biological computers beat AI chips on energy?

Nobody has shown it: we found 0 measured per-task energy comparisons between living neurons and a chip in 23 experiments, and “a million times less” is a ratio of two power ratings, about 1,050,000 to 1 (21 megawatts for the Frontier supercomputer against 20 watts for a brain) ^[28]. The roadmap paper itself adds that the two are “performing quite different tasks”. Other versions of the ratio: 10 megawatts against 20 watts on FinalSpark’s site (500,000 to 1) ^[43] and 8 megawatts in the Brainoware preprint (400,000 to 1) ^[21]. FinalSpark’s own 2023 article says “one billion times” ^[42]. Its peer-reviewed platform paper makes no such claim; it states 20 watts for a brain and about 10 gigawatt-hours for training GPT-3 ^[2].

Figure 13

How "a million times less energy" travelled

Each step in time order. Steps are evenly spaced. The only calculation is a power ratio of a whole supercomputer to a whole brain; nothing compares energy per task.

Where the number beganRepeated, no new dataMass mediaFact-check

Biological computing uses a million times less energyVerdict: a power ratio, not a measurement

1. 2023-02Smirnova et al.: 21 megawatts (Frontier) against 10 to 20 watts (a brain)Table 1: a million-fold power efficiency, "albeit while performing quite different tasks"
2. 2023-07FinalSpark article: "one billion times more energy efficient"No calculation shown
3. 2024Headlines: "a million times less power" (January and May 2024)Listed on the FinalSpark press page
4. 2024FinalSpark's peer-reviewed platform paper makes no such claimIt gives 20 W for a brain and the energy of training GPT-3
5. 2026FinalSpark site: silicon would need 10 megawatts, "according to" BoahenRatio 500,000 to 1; we did not read his paper
6. 2026Brainobot preprint: low energy claimed, none measured; equipment "bulky and energy-intensive"Our search found 0 measured per-task comparisons in 23 experiments

The numbers behind the comparison differ by source. Google engineers estimated GPT-3 training at 1,287 megawatt-hours ^[45]; FinalSpark’s paper says about 10 gigawatt-hours, 7.8 times as much ^[2]. On the inference side Google measured a median Gemini text prompt at 0.24 watt-hours ^[46], which is what a 20-watt brain uses in 43 seconds (IQ Metrics calculation). A fair test would measure the energy of both systems for the same task at the same accuracy, including the incubator, life support and computers that keep the cells alive. The Brainobot authors say that equipment is “bulky and energy-intensive” ^[25], and the CL1 and Neuroplatform pages give no wattage ^{[39] [41]}.

For what AI systems can do, and how benchmarks are scored, see whether AI models share a g factor, the 2026 benchmark review and why benchmark scores disagree.

10 What does a biological computer cost, and who sells one?

Neither of the two companies that sell access prints a price on its own page (checked 3 November 2026): Cortical Labs (the CL1 and the Cortical Cloud) and FinalSpark (the Neuroplatform) ^{[39] [40] [41]}. The CL1 page says neurons stay alive for up to 6 months and that the device is self-contained; the Neuroplatform page lists a Shared plan (4 shared organoids, one user) and a Dedicated plan (4 dedicated organoids, several users) and says “contact for pricing”. Press reports of a launch price and a weekly cloud rental exist; they are not on the company pages we saved, so we do not use them. Price per neuron is therefore not computed: no company page prints both a price and a neuron count.

Table 1

What the platform makers say on their own pages, 3 November 2026

Prices and neuron counts are shown only where the company prints them. "Not printed" means we saved the live page and found nothing; copies archived on 4 March and 11 July 2025 say the same for the CL1.

Platform and maker	What it is	Neurons and price on the page	Energy claim on the page
CL1, Cortical Labs (Melbourne)	Human neurons on a silicon chip with life support; neurons live up to 6 months	Neurons: not printed (CL1 papers: 500,000 cells seeded). Price: not printed	“A fraction of the energy” (no figure)
Cortical Cloud, Cortical Labs	Remote code deployment onto CL1 units with a Python SDK	Neurons: not printed. Price: not printed (sign-up page)	“A fraction of the energy” (no figure)
Neuroplatform, FinalSpark (Vevey)	4 arrays of 4 organoids, 8 electrodes each; remote, around the clock; Shared and Dedicated plans	Neurons: not printed (organoids about 500 um across). Price: “contact for pricing”	2023 article: “one billion times”; site: 10 MW against 20 W

The FinalSpark platform paper describes what a buyer gets: 4 arrays with 4 organoids each and 8 electrodes per organoid (128 electrodes in all, IQ Metrics calculation), organoids about 500 micrometres across, remote access around the clock, and over 1,000 organoids used in three years with more than 18 terabytes of recordings ^[2]. Researchers at the University of Bristol used it remotely for the Braille study ^[19]. Cortical Labs lists a Python SDK and open API documentation ^[38].

11 Who is working on organoid computing?

Eight countries, led by the USA (8 experiments), Australia (5) and Japan (4); 17 of 23 experiments come from university groups alone and 6 involve Cortical Labs ^[1]. The USA has 8 (Indiana University, Johns Hopkins, UC Santa Cruz, Tulane, Georgia Tech, Caltech), Australia 5, Japan 4 (Tohoku, Tokyo, RIKEN), China 2 (Tianjin; Beijing) and Israel, Italy, Switzerland and the United Kingdom 1 each. Johns Hopkins leads a 2023 roadmap with co-authors at other institutions ^[28], and its team has since published organoid plasticity work with Cortical Labs co-authors ^[12]. A Tianjin University “brain-on-chip” robot was widely reported in 2024; we could not read its paper and left it out of the data.

12 Is organoid computing ethical, and could a dish be conscious?

Experts disagree on whether a dish could ever be conscious, and this page takes no side: we report who says what. Today’s organoids hold under 100,000 cells ^[28]. The 2023 roadmap says terms such as intelligence, sentience and consciousness “cannot be directly translated” to simple cultures and are used for basic functions, and it expects the concern to grow as organoids become more complex ^[28]. A 2024 paper argues that organoids’ shortcomings suggest they fit none of the operational definitions of consciousness, while saying questions will arise as organoids gain closed-loop learning and embodiment ^[30]. An ethics analysis of DishBrain discusses what a hybrid of human neurons and silicon could mean morally ^[31], a review covers consent, moral status and governance ^[29], and a 2026 consensus from the Asia Pacific Neuroethics Working Group aims to dispel “ethical red herrings” arising from sensational portrayals ^[32]. The US National Academies’ 2021 committee was asked how large or complex an organoid would need to be to attain awareness ^[35].

Behind the debate is a size fact. Today’s organoids are under 500 micrometres and under 100,000 cells, against about 3 billion cells in the neocortex of a 12-week human fetus ^[28]. Cell cultures from mouse embryos are used as well as human cells; the 2022 and 2025 DishBrain studies used both ^{[1] [10]}. Related reading on how brains are measured, in people, is in our study of AI and your brain and our brain statistics fact check.

13 What does this have to do with intelligence tests?

Nothing directly: no dish or organoid has an IQ, because IQ tests are scored against people of the same age and these cells have no norm group, no body and no way to read a question. What the experiments measure is narrower: a rally length, a classification accuracy, a regression score. In people, the nearest ideas are fluid reasoning (solving new problems) and working memory, which fluid and crystallised intelligence and working memory and reasoning separate, and processing speed (processing speed and IQ).

For machines, the same care applies to the word “intelligence”: see how AI scores 0.51% on a test that people solve in full, what an AI capability index of 169 does not mean, and what AGI would require.

14 Every experiment in our dataset

Table 2

Every experiment and platform in our dataset (26 rows)

Cells or neurons are as printed in the paper (see the CSV for what each number counts: cells plated, neurons read out or a calculation). Grades follow the rule above. The full CSV adds electrodes, task, claim, what was shown, controls, replicates, data access, funding and conflicts.

Experiment or platform	Year	Tissue	Cells or neurons	Task type	Grade	Published	Laboratory, country
DeMarse et al.: the neurally controlled animat	2001	Cultured neurons	not stated	closed-loop control	D	journal	California Institute of Technology, USA
Shahaf and Marom: learning in cortical networks	2001	Cultured neurons	not stated	stimulus-response conditioning	Not graded	journal	Technion, Department of Physiology and Biophysics, Israel
Ruaro et al.: image processing with neuronal cultures	2005	Cultured neurons	not stated	pattern recognition	Not graded	journal	Scuola Internazionale Superiore di Studi Avanzati, Italy
Bakkum et al.: embodied cortical network, goal-directed task	2008	Cultured neurons	not stated	closed-loop control	Not graded	journal	Georgia Institute of Technology, USA
Isomura et al.: blind source separation by cultured neurons	2015	Cultured neurons	500,000	stimulus-response learning	A	journal	University of Tokyo, Japan
Kagan et al.: DishBrain (Pong)	2022	Cultured neurons	800,000 to 1,000,000	closed-loop game	A	journal	Cortical Labs, Australia
Cai et al.: Brainoware	2023	Organoid	not stated	reservoir computing	B	journal	Indiana University Bloomington, USA
Habibollahi et al.: criticality in DishBrain	2023	Cultured neurons	1,000,000	closed-loop game	A	journal	Cortical Labs, Australia
Isomura et al.: causal inference in cultured networks	2023	Cultured neurons	not stated	stimulus-response learning	A	journal	RIKEN Center for Brain Science and University of Tokyo, Japan
Sumi et al.: neurons as reservoirs	2023	Cultured neurons	60	reservoir computing	A	journal	Tohoku University, Japan
Alam El Din et al.: organoid building blocks of learning	2025	Organoid	not stated	plasticity assay	A	journal	Johns Hopkins University, USA
Deng et al.: optogenetic reservoir, obstacle avoidance	2025	Cultured neurons	15 to 350	reservoir computing	C	journal	Beijing University of Posts and Telecommunications, China
Khajehnejad et al.: DishBrain against deep reinforcement learning	2025	Cultured neurons	not stated	closed-loop game	A	journal	Cortical Labs, Australia
Kuchler et al.: neurons as NAND and OR gates	2025	Cultured neurons	70,000	logic gates	C	journal	ETH Zurich, Switzerland
Rountree et al.: paired organoids, potentiation and a maze game	2025	Organoid	not stated	closed-loop game	D	preprint	Tulane University, USA
Shao et al.: repeated training of cultured networks	2025	Cultured neurons	80,000 to 135,000	pattern recognition	C	journal	Tianjin University, China
Watmuff et al.: drugs change DishBrain gameplay	2025	Cultured neurons	200,000	closed-loop game	A	journal	Cortical Labs, Australia
Liu et al.: Braille letters read by organoids	2026	Organoid	not stated	pattern recognition	C	journal	University of Bristol, United Kingdom
Robbins et al.: goal-directed learning in cortical organoids (cartpole)	2026	Organoid	not stated	closed-loop control	A	journal	University of California Santa Cruz, USA
Loeffler et al.: handwritten digits on the CL1	2026	Mixed	500,000	reservoir computing	C	preprint	Cortical Labs, Australia
Cai et al.: Brainopheno, organoid pattern classification	2026	Organoid	not stated	pattern recognition	D	preprint	Indiana University Bloomington, USA
Sono et al.: online learning in modular neuron networks	2026	Cultured neurons	1,869	reservoir computing	A	journal	Tohoku University, Japan
Cai et al.: Brainobot, an organoid-controlled humanoid robot	2026	Organoid	not stated	closed-loop control	C	preprint	Indiana University Bloomington, USA
Johns Hopkins organoid intelligence programme (roadmap)	2023	Roadmap	10,000,000	programme	–	journal	Johns Hopkins University, USA
FinalSpark Neuroplatform	2024	Platform	not stated	platform	–	journal	FinalSpark, Vevey, Switzerland
Cortical Labs CL1 and Cortical Cloud	2025	Platform	500,000	platform	–	journal	Cortical Labs Pte Ltd, Melbourne, Australia

15 How we built the dataset

1. 01

   Find every experiment

   We searched Europe PMC for organoid intelligence, synthetic biological intelligence, wetware, reservoir computing with cultured neurons, closed-loop game and robot control with neurons and the platform names, 2001 to 3 November 2026, and followed the references of the papers we read. Only experiments whose primary text we could read are rows. Three old papers could be read only as abstracts and are marked.

2. 02

   Read the primary text

   Each paper was saved as full text (open licence) and each number was read by script from the saved text: cells plated, electrodes, controls, units, data statements, funding and conflicts. News and blogs were leads only.

3. 03

   Keep conflicts

   Where sources disagree we kept both. One paper says mouse in its results and rat in its methods ^[14]; the CL1 neuron count differs between press and papers. The notes list every conflict.

4. 04

   Grade with a printed rule

   Five yes or no checks per experiment, computed by script; platforms and roadmaps are not graded. Independent replication (another laboratory, same task, matching result) is recorded separately and found none.

5. 05

   Date everything

   Company pages were saved on 3 November 2026; the CL1 page was also read in copies archived on 4 March and 11 July 2025. Prices are recorded only where a company prints them: none does.