The organoid intelligence award with no neurons in it

What the work claims

This is a grant abstract, and it should be weighted as an intention rather than a result. Matthew Bennett at Rice, with co-investigators Caroline Ajo-Franklin, Anastasios Kyrillidis and Kirstin Matthews at Rice and Kresimir Josic at the University of Houston, holds 1,998,147 dollars obligated in fiscal 2025, running from August 2025 to July 2029 under the Emerging Frontiers in Research and Innovation office.¹

The proposal is to build an electrogenetically networked Cyber-Bacterial Organoid: engineered microbial consortia that take chemical and electronic inputs, integrate them, and emit electronic outputs which are relayed to a network of parallel consortia, performing what the abstract calls high-dimensional chemical pattern recognition. The team intends to add cellular memory and adaptive learning through long-term electrogenetic interfacing and transcriptional feedback that sensitises cells to inputs, which requires continuous culture systems able to keep the consortia alive and electrically addressable for extended periods. Target applications are diagnostics and environmental monitoring.

The bold claim is learning. Sensing and classification in engineered bacteria are established; a consortium that refines its own responses over time through transcriptional feedback is not.

How it works

Electrogenetics is the load-bearing idea, and the team is well chosen for it. It means using electrical signals to control gene expression, and reading cellular state back out electrically. The return path is extracellular electron transfer, the ability of certain bacteria to move electrons across their membrane to an external acceptor such as an electrode. Ajo-Franklin's published work is squarely on this: extracellular respiration in Escherichia coli, bioelectronic sensors built from electroactive bacteria coupled to organic electrochemical transistors, and multichannel bioelectronic sensing with engineered E. coli. Bennett's is on synthetic gene circuits and the dynamics of microbial consortia. Josic supplies the mathematics of coupled dynamical systems, Kyrillidis the machine learning.

Functionally this is a read and write channel into living cells, the same role a microelectrode array plays for neural tissue. It is not, however, an equivalent channel, and the difference is worth stating precisely because it bounds everything the platform can do. Writing by transcription takes minutes to hours, against milliseconds for an electrode depolarising a neuron. The write is effectively broadcast to the culture rather than addressed to individual cells, and the readout is a bulk measurement rather than hundreds of spatially resolved channels. What the architecture buys instead is parallelism between consortia rather than within one, which is why the abstract describes a network of parallel consortia rather than a single larger device.

That is a real trade and not obviously a bad one. Spatial addressability and bandwidth are surrendered; in exchange the substrate becomes clonable, cheap, genetically tractable and reproducible in a way no organoid is.

Where a skeptic should push

Nothing here has been demonstrated. Checked four independent ways, against the NSF award record, OpenAlex filtered on the award identifier, Europe PMC funding acknowledgements, and searches for the award's own coinages, this project has produced no publications. That is the expected state of affairs eleven months into a four-year award and is not a criticism of the team; it is a statement about what a reader is entitled to conclude, which is nothing about whether any of this works.

The word doing the most work is learning. Transcriptional feedback that sensitises cells to inputs describes, on its face, adaptation or hysteresis: the system's response changes as a function of its history. That is a necessary condition for learning and nowhere near a sufficient one, and the distinction is exactly where biocomputing claims tend to inflate. The solicitation makes this harder rather than easier to police, because it explicitly asks investigators to define the bounds of intelligence and learning for themselves. A programme that lets each project set its own bar for the central claim will produce a portfolio whose successes are not comparable with one another.

There is also a mechanism in the proposal working against itself. Adaptive learning here requires long-term continuous culture, because the feedback has to accumulate. Continuous culture is also the condition under which engineered genetic circuits are selected against: they impose a metabolic burden, and any mutant that loses or silences the construct outgrows its neighbours. The requirement that makes the learning possible is the one that erodes the copy fidelity that made bacteria attractive as a substrate in the first place. Whether the team can hold a functioning circuit stable long enough to demonstrate learning is, on this evidence, the project's central technical risk, and the abstract acknowledges the need for continuous culture systems without saying how stability will be maintained.

Timescales bound the application space too. A classifier whose internal state updates on transcriptional timescales cannot serve any task requiring fast closed-loop control. Chemical pattern recognition in patient samples or environmental monitoring, the applications actually named, tolerate that. Little else would.

What the funding line actually defines

My first reading of this award was that it was an anomaly, an organoid intelligence programme spending two million dollars on a substrate with no neurons and no tissue. Checking the solicitation destroyed that reading, and the correction is more interesting than the claim.

NSF 24-508 defines its terms, and the definition pre-authorises exactly this project. The solicitation states that the term organoid has broad meaning, capturing a range of designer three-dimensional cellular constructs, microphysiological systems and engineered tissues; that projects may use organoids representative of any organ system; and, decisively, that cells used to construct organoids need not be mammalian, giving three-dimensional plant cell or biofilm based constructs for biocomputing as an explicit worked example of what is allowable.² A biofilm is a bacterial consortium. This award is not an outlier or a stretch; it is the definition operating as written.

The programme's criterion is functional throughout, never anatomical. Its objective is three-dimensional in vitro biological systems capable of information processing and actuation, organ-like systems that problem-solve, learn and adapt. Across the entire solicitation text the words neural, neuron, bacteria and microbe do not appear at all. The name carries a neural connotation that the document itself never asserts.

This matters because the public and professional ethics conversation about computing on living tissue has organised itself around the phrase organoid intelligence, and treats it as though it denotes neurons. In the founding funding instrument, it does not. Anyone drafting governance for organoid intelligence, and anyone reasoning about whether that governance is adequate, is working with a term whose institutional definition is substrate agnostic. Rules written for the phrase would sweep in biofilm and plant-cell projects that raise no question of sentience whatever, while offering the neural subset no treatment specific to the thing that actually makes it contentious. The tension is not between this programme and this award; it is between a name and its own definition, and the field inherited the name.

The programme does draw one hard boundary, and it is a revealing one. Proposals using human embryonic stem cells, and proposals involving human or non-human chimeras, are returned without review; induced pluripotent stem cells are permitted. That is governance by funding eligibility rather than by ethics review, and it is a far blunter and more consequential instrument than any guideline. It does not weigh a project's merits against its risks. It removes a category of substrate from the programme entirely, in advance, at the point of submission. Much of what determines which biocomputing substrates get built is settled this way, in eligibility clauses, before any ethics committee is convened.

Two cautions on the substitution argument, both of which I had initially overstated. First, bacteria do not escape governance; they change which governance applies. Engineered consortia sit under recombinant DNA biosafety and institutional biosafety committee review, and a continuously cultured organism engineered to recognise chemical signatures sits closer to dual-use and environmental release concerns than a dish of neurons ever will. Second, bacterial and neural biocomputing are not competing for the same applications. Transcriptional-timescale chemical classification and millisecond adaptive control are different problems, and one will not displace the other on task merit. Where they genuinely compete is for programme dollars and for definitional authority, and this award demonstrates that competition resolved in the bacteria's favour on both counts.

One further correction. The award funds work on the ethical, legal and social implications of biocomputing, covering regulatory frameworks, public perception and responsible development, and Matthews is a genuine science and technology policy scholar at Rice's Baker Institute whose published work covers embryo and embryoid research policy and synthetic biology regulation. It would be wrong to present this as the team's distinctive initiative, because the solicitation makes such work Research Thread 3 and requires every proposal to address it. Portfolio-wide mandated anticipatory governance is a notable design choice by NSF and better than the alternative. Its limitation is that each project scopes its own ethics work to its own substrate, so a portfolio defined functionally across biofilms, plant cells and neural tissue produces substrate-specific findings and nobody positioned to do the comparative moral-status work that the functional definition makes necessary.

The bottom line

Established: NSF has committed just under two million dollars to bacterial biocomputing under its organoid intelligence line, and the solicitation explicitly authorises non-mammalian and biofilm-based constructs. Both facts are documentary and verifiable, and together they show the programme's category is functional rather than anatomical.

Hypothesis, and nothing more at this stage: that engineered microbial consortia can be made to learn, that the electrogenetic interface will carry enough information to support high-dimensional classification, and that continuous culture can hold the circuits stable long enough to demonstrate either. No results exist yet.

What would confirm the technical claim: a demonstration that classification accuracy improves with exposure history in a way not explained by simple adaptation, holding across a culture period long enough for selection against the construct to have operated. What would break it: circuit loss in continuous culture on a timescale shorter than the learning protocol requires, which would make the two central requirements mutually exclusive rather than merely difficult. On the governance reading, what would overturn my argument is a subsequent solicitation that narrows the definition of organoid back toward neural tissue, which would make the mismatch between the name and the portfolio a transitional artefact rather than a structural feature.