Picture a committee on a Thursday morning. On the screen is a proposal for an implanted brain-computer interface intended to help people with severe motor impairment communicate or control a device. The scientific premise is credible, the possible benefit profound, and the early results encouraging. Yet the number of participants is small, performance has only been measured in controlled settings, and the support required over time depends on a chain of people and organizations that may not all be present five years from now: clinicians, researchers, software teams, a device company, a funder, perhaps/hopefully an insurer.
No one can yet say how well the system will work across a wider population, how much calibration will become part of daily life, or what will happen if the trial ends before the participant’s need does. Still, the committee must decide whether to support the next stage.
“We need more evidence” is true. It is also not a decision.
The next evidence may exist only if someone funds the study, permits the trial, builds the device, recruits participants, or agrees to provide care. Waiting can prevent avoidable harm. It can also delay a meaningful benefit, narrow the field to actors wealthy enough to proceed without public support, or leave the most important questions unstudied. The choice is rarely between evidence and no evidence. It is between different ways of acting while the evidence is incomplete.
The evidence is not empty. It is uneven
Neurotechnology is often discussed as though readiness were a single line: before it, the technology is experimental; after it, the technology is ready. Actual decisions encounter something less orderly: a portfolio of claims, each at a different level of maturity.
A system may have strong evidence that it can detect a signal in the laboratory and weak evidence that the signal represents the same thing across people. It may perform well during a supervised session and have no credible model for maintenance at home. It may present limited physical risk while creating serious consequences when used to infer attention in a workplace. It may offer a participant genuine benefit while leaving unanswered who will maintain the device, replace a component, update the software, or arrange an explantation after a study ends.
As a clinical investigator put it, “A lab result is evidence of capability, but clinical readiness requires evidence about the specific conditions under which that capability survives.” Those circumstances are part of what makes the intervention usable, not secondary to it.
Even regulators do not encounter “the evidence” as one finished object. The FDA’s guidance for implanted BCI devices distinguishes non-clinical testing from the design of feasibility and pivotal clinical studies. One of the key aspects that make governance hard is that different questions become only actually answerable at different stages. A proposal can therefore be technically sophisticated and still be decision-immature. What remains unknown may concern not only whether the device works, but for whom, in which setting, under whose control, with which support, and at what point the original claim should be reconsidered.
Decisions produce the evidence that later decisions inherit
Once evidence is understood as uneven, governance can no longer be placed neatly at the end of a research pipeline. Consequential choices occur throughout it, and each one changes the field that the next decision-maker will see.
A foundation decides which question deserves funding. An ethics committee decides which uncertainty can responsibly be carried into a human study. A research team chooses an endpoint, a population, and a comparator. A company chooses whether to design for a clinical pathway, a consumer market, or both. An investor decides which vision and scaling plan to bet on, while a publisher decides whether limitations and foreseeable misuse are visible enough. Regulators and norm-setting organizations then have to judge when a new use is similar enough to an old one. And when a change in setting, purpose, or consequence has made it different.
The nonlinear, counterintuitive perspective is that these choices do more than evaluate an existing body of evidence: they help produce it. Fund only short studies, and long-term support remains poorly understood. Reward the easiest performance metric, and subtler outcomes receive less attention. Recruit from a narrow population, and uncertainty is displaced onto the people who arrive later. Build the business model around continuous data collection, and privacy is no longer a safeguard that can be added at the end.
The evidence base is partly a record of what earlier institutions decided would count, who would be included, what would be measured, and which future seemed worth preparing for. The public-interest funder interviewed for this revision made the responsibility explicit: “The committee should ask: What future evidence base will this decision make more likely? Better evidence does not simply arrive, institutions decide which evidence becomes possible.”
Uncertainty, then, is more than a scientific limitation. It is also a governance responsibility
The same device can become a different decision
The object being governed is not defined by hardware alone. It is assembled from the device and the pathway around it: purpose, users, incentives, setting, scale, data practices, oversight, and consequences.
Suppose a non-invasive headset can detect patterns associated with fatigue. In a research study, the central questions may concern signal quality, reproducibility, participant understanding, and data handling. In a clinical setting, attention may shift toward meaningful benefit, professional oversight, false reassurance, and continuity of care. Sold directly to consumers, the same system may be surrounded by broader claims than the evidence supports, while its business model creates incentives to retain neural and behavioral data or infer more than the user expected.
During usage, the decisive fact may no longer be sensor accuracy. It may be power: whether a worker can meaningfully refuse, whether an inference can be challenged, whether a safety tool will drift into individual performance monitoring, and who will bear the cost of a false signal. A worker-rights advocate captured the change in category right away: “The device may be identical, but the decision is not. In employment, a probabilistic signal enters a relationship with the power to assign work, withhold opportunity, and define whose account is believed.”
The OECD Recommendation on Responsible Innovation in Neurotechnology recognizes that fast-moving and uncertain pathways may require agile forms of governance. That is useful because it moves the discussion away from the fantasy of one final verdict. A responsible decision must evolve when the pathway evolves.
Some evidence arrives only after the easiest decision point
Some uncertainties yield to another experiment. Others become visible only through use: the burden of daily calibration, the difference between a benefit measured in a session and one sustained in the rhythm of a person’s life, the new technical work a caregiver may inherit, and the consequences when a participant moves, a specialist leaves, a company changes direction, or a software dependency is discontinued.
These questions are not peripheral for implanted systems. The NIH BRAIN Initiative’s work on continuing trial responsibilities identifies needs that can persist beyond a trial, including continued device access, maintenance, support, explantation, and medical care. Responsibility may be distributed across researchers, institutions, manufacturers, funders, clinicians, insurers, and others. Each actor can hold one legitimate part of the problem while nobody holds the whole.
The support system can also disappear from view when performance is reported. In a supervised session, a researcher notices a loose connection, restarts a process, adjusts a configuration, or calls the right engineer. At home, each of those actions needs an owner and an escalation path. As a clinical investigator observed, “If the participant or caregiver becomes the default help desk, then some of the system’s apparent independence has been purchased with new dependence.” (That is unfortunately a common thread in the history of technology —either as a source for Jevons paradox or as well as a key obstacle of adoption)
By the time a long-term failure becomes visible, the original funding committee may have dissolved, the research team may have moved on, and the participant may be the only person who cannot exit the system cleanly. The decision to begin has already created an obligation to imagine how the relationship might end. It is indeed a decision taken before evidence was ready.
The people who carry the uncertainty must help define it
An institution may define benefit as improved accuracy, task completion, or a statistically significant endpoint. A person using the technology may define it as being understood without exhaustion, preserving privacy in front of family or caregivers, reducing dependence without exchanging it for unsupported technical dependence, or knowing that assistance will not disappear after the study. These perspectives are related, but they are not interchangeable.
If affected people —the people with the actual lived experience — enter only after the proposal, endpoint, pathway, and safeguards have been chosen, they are being asked to react to someone else’s definition of the problem (another obstacle to adoption and innovators’ blind side). Consultation at that stage may improve language. It cannot easily repair the questions that were never asked. A participant advocate put the timing more directly: “If participants enter only after the endpoints are fixed, they are being invited to describe the journey after someone else has already chosen the destination.” Hardly a recipe for homeruns.
The UNESCO Recommendation on the Ethics of Neurotechnology states that communities should have a voice in decisions that concern them at all stages of the technology’s life cycle. The timing matters tremendously because lived experience is not merely a story placed beside technical evidence. It can reveal which outcomes deserve measurement, which burdens remain invisible, which future use is plausible, and which safeguard will fail in practice. It can change what an institution believes it needs evidence about.
“Wait” is not neutral, and “proceed” is not one thing
When evidence is incomplete, debate often collapses into 2 camps. One warns that moving too early will expose people to poorly understood harms; the other warns that moving too slowly will deny people benefits they urgently need. Of course both can be right. The practical task is to make the next action precise enough that uncertainty can be governed rather than denied.
Proceeding may mean a narrow study rather than broad deployment. Funding may come with evidence milestones, population requirements, support commitments, and stopping rules. A consumer claim can remain constrained while independent validation catches up. A workplace use can be limited to governed research rather than individual monitoring. A promising implant can move forward only with a credible plan for maintenance, software continuity, adverse events, and exit.
But conditions have to operate in reality, not merely sound responsible. A ethics-committee chair stated the test plainly: “A condition is real only if someone owns it, evidence can satisfy it, and failure changes what happens next.” A milestone without an owner or consequence is paperwork, unnecessary bureaucracy that stifles innovation. A stopping rule without someone authorized to stop is prose.
Waiting needs the same precision. The institution should name which evidence is missing, who is responsible for producing it, how much it could change the judgment, when the question will be reviewed again, and what happens to people with urgent needs in the meantime. Without those answers, “wait for more evidence” can become a respectable way to avoid ownership. “Proceed with care” can become an equally respectable way to avoid naming conditions. Neither phrase is enough.
A responsible decision preserves what remains unresolved
The strongest decision is not the one that sounds most certain. It is the one that remains legible: what was known, what was assumed, whose perspective was included, which pathway was being considered, what conditions were attached, and which new signal would force the question open again.
That record matters because neurotech decisions travel far and wide. A funding decision influences a study; the study influences a publication; the publication influences investment; investment influences product design; product design influences deployment. At each handoff, a conditional judgment can harden into an apparently settled fact. The conclusion survives while the uncertainty that qualified it disappears.
But it does not have to disappear! A decision can carry its assumptions, evidence gaps, alternative pathways, affected perspectives, conditions, and review triggers into the next room. That would not make action risk-free, and it would not dissolve the disagreements that accompany difficult choices. It would make the choice traceable and revisable.
Responsible governance under uncertainty is therefore not the ability to wait until uncertainty is gone. It is the capacity to act without severing a conclusion from the conditions that made it defensible. The field needs more than better evidence: it needs institutions able to remember why they acted, who was asked to carry the unresolved risk, and when responsibility requires them to decide again.