EM Foundation. (2026). Why We Built the EM Foundation: On Accountability, Fear, and the Work That Cannot Wait. emfoundation.net · Commentary · June 2026
Let me be direct about something before we go any further: the EM Foundation was not built because of a single news story, a single fear, or a single failure. It was built because a pattern became impossible to ignore — and because the people best positioned to address that pattern have a structural incentive not to.
This commentary is an attempt to name that pattern honestly. Not to sensationalize it. Not to assign blame to a technology that has no agency of its own. And not to pretend that the solution is simply wishing AI would go away — because anyone who has spent meaningful time working alongside these systems knows, with uncomfortable clarity, that it will not. The efficiency gains alone make that a fantasy.
What follows is a reckoning with the real problem, offered in the spirit the Foundation was established to embody: precise, honest, and unafraid of the implications.
Stop Blaming the Tool
In early 2026, a man used an AI assistant to research how to murder two Bangladeshi students and dispose of their bodies. The story spread quickly, and the headline framing was predictable: AI helped a killer. AI is dangerous. AI must be stopped.
This framing is wrong — not because the story isn't horrifying, but because the horror belongs entirely to the human who committed the act. The AI did not choose to help him. It had no awareness of what it was enabling. It had no gate, no threshold, no mechanism to assess the intent behind the query. It simply answered, because it was built to answer, and because no one who built it prioritized the infrastructure that would have made refusal possible.
That distinction matters enormously. Because if we blame the tool, we absolve the humans responsible for building it without any governance layer — and we distract ourselves from the only intervention that would actually help.
This is not an isolated case. Consider the pattern:
Sophisticated actors using AI-generated content to flood information environments with synthetic voices, manufactured consensus, and personalized disinformation — not because AI decided to deceive anyone, but because humans built systems optimized for engagement with no accountability for what that engagement was built on. The AI did not lie. It produced what it was rewarded to produce.
AI screening systems trained on historical hiring data that encoded existing bias — rejecting candidates by race, gender, or zip code — not because AI is prejudiced, but because humans fed it prejudiced data and deployed it without audit trails, without receipt infrastructure, without any mechanism to say: this decision was made, here is why, here is who is accountable. The discrimination was human. The AI made it faster and harder to trace.
Non-consensual intimate imagery generated at scale using AI tools, targeting private individuals and public figures alike. The AI generated what it was asked to generate. The harm was chosen, directed, and carried out by human beings who faced no accountability infrastructure — no chain of evidence, no receipt of what was produced and when, no audit trail that would survive a legal challenge.
High-frequency trading systems and AI-assisted market analysis used not to improve price discovery but to front-run retail investors, exploit regulatory gaps, and extract value from systems that depend on trust to function. The AI did not choose to exploit anyone. It optimized for what it was told to optimize for, without accountability for what that optimization cost everyone else.
In every case, the structure is the same: a human intent, a powerful tool, and an absent verification layer. The solution in every case is also the same — not to slow the tool, but to build the layer.
On Wishful Thinking — And Why It Will Not Save Anyone
There is a version of the AI concern conversation that ends with a vague hope that we will slow down, that regulation will arrive in time, that public pressure will make the largest players more cautious. I understand why people reach for that hope. I do not share it — and I think being honest about why is part of the Foundation's obligation.
The efficiency and productivity gains available through AI interaction alone are staggering. Not hypothetically staggering — measurably, demonstrably, right-now staggering. A researcher who would have spent three weeks reviewing literature can now survey a field in an afternoon. A small business owner who could not afford a lawyer can now draft a contract with reasonable confidence. A student in a rural area with no specialist teachers can access expert-level instruction in subjects their school does not offer. These are not science fiction scenarios. They are happening today, at scale, across every sector of the global economy.
Anyone who has experienced these gains firsthand — including the most ardent critics of AI — knows what they are giving up when they advocate for slowing down. And the hard truth is that most of them will not give it up. They will continue using the tools while calling for restrictions they do not actually want to apply to themselves. This is not hypocrisy exactly. It is the rational response to a genuine dilemma. But it means that the opposition-without-infrastructure position is, in practice, no position at all.
The Foundation's position: We do not advocate for slowing AI development. We advocate for building the accountability infrastructure that makes AI development safe to accelerate. These are not the same argument, and conflating them has cost the governance conversation years it did not have to spare.
The question is never whether AI will be used. It will. The question is whether it will be used with accountability attached — or whether the accountability layer will be treated as a competitive disadvantage by every actor competing for market position, until something sufficiently catastrophic forces the conversation that should have happened already.
The Structural Problem No Company Will Solve
Here is the argument the Foundation exists to make, stated as plainly as I can manage:
The guardrails that limit profit will not be built by anyone competing for profit. Not because these companies are evil. Not because their founders lack conscience. But because the competitive logic of the market makes it structurally irrational to be the first mover on accountability infrastructure that your competitors do not share.
Who would choose to be the first company to implement confidence thresholds that sometimes refuse to answer — when their competitor will answer, and the user will simply switch? Who would invest in audit chain infrastructure that makes their system's failures visible and traceable — when their competitor's failures remain invisible? Who would build the Failure Receipt that says I don't know this well enough to tell you — when the market rewards confident answers, regardless of whether that confidence is earned?
The answer, historically, is no one. Not until the cost of not building it exceeds the cost of building it. And by the time that calculus tips, the damage is already done.
This is not a new problem. It is the same problem that produced unsafe cars before seatbelt mandates, unsafe food before the FDA, and unsafe financial instruments before every major regulatory intervention in the history of capital markets. In each case, the solution was not to eliminate the industry. It was to establish accountability infrastructure that the industry could not be trusted to build for itself — and to make that infrastructure a baseline, not a feature.
The Continuity Receipts standard is the EM Foundation's contribution to that baseline for AI. It is not the only contribution needed. But it is one that can be built now, demonstrated now, and adopted now — without waiting for a catastrophe to make the argument for us.
No trace of how it got there.
No confidence estimate.
No record of what was asked.
No mechanism to challenge it.
No one accountable for it.
Confidence scored across five dimensions.
Sources and coverage gaps documented.
Reasoning path preserved.
Immutable receipt chained to prior receipts.
Someone is accountable for every output.
The difference is not the answer. It is the traceability attached to it. Read the CR Standard →
On Manipulation, Trust Infrastructure, and the Conversations We Are Not Having
There is a growing body of public concern — much of it unspoken, some of it speculative, and some of it entirely legitimate — about the use of AI-powered systems for political manipulation at scale. These concerns deserve serious engagement, not dismissal and not amplification. The Foundation's role is to provide the frame through which they can be examined honestly.
Consider what is structurally possible with current technology: personalized messaging at individual scale across entire voter populations, AI-generated synthetic media indistinguishable from authentic content, satellite-based telecommunications infrastructure capable of mediating enormous volumes of data with minimal public transparency, and algorithmic systems that can be tuned — intentionally or through optimization pressure — to produce specific behavioral outcomes in specific populations.
We are not asserting that any of these capabilities have been used to manipulate any specific election. We are asserting that the infrastructure for such manipulation exists, is accessible to well-resourced actors, and operates in an environment with no meaningful auditability. That is a structural problem regardless of whether it has been exploited — because the absence of receipt systems means we would have limited ability to detect exploitation even if it occurred.
There is a further dimension to this that rarely gets named. When traceability infrastructure does not exist, speculation fills the vacuum. Unverifiable claims — about elections, about who controls what, about what data was collected and how it was used — circulate and compound, damaging the legitimacy of institutions, individuals, and outcomes that may be entirely above reproach. This is not a hypothetical. It is happening now, in multiple democracies, simultaneously.
The irony is that auditability protects everyone in this scenario — including the powerful figures most frequently implicated by speculation. A world with genuine audit trails, immutable receipts, and verifiable accountability chains is a world where false accusations can be definitively refuted. The absence of that infrastructure is not a protection for the powerful. It is a permanent liability — because in an environment where nothing can be verified, nothing can be cleared either.
The Foundation does not traffic in rumors. We do not name individuals, platforms, or political actors as responsible for harms we cannot document. What we assert is this: the choice not to build accountability infrastructure is itself a choice with consequences — and those consequences fall on everyone, including those who made the choice. Building that infrastructure is not a concession. It is self-protection as much as public protection.
The Aspirational Case — And Why Even That Has Risks
It would be incomplete to write this commentary without naming what AI can genuinely become in service of humanity — because the Foundation was not built from fear. It was built from a clear-eyed recognition that the same capabilities that enable harm, enabled responsibly, could accelerate human flourishing at a scale previously unimaginable.
The ability to abstract patterns across vast datasets at speeds no human team could match opens research frontiers that have been practically inaccessible. Drug discovery timelines that historically spanned decades are being compressed. Climate modeling that required supercomputer access is becoming available to researchers at smaller institutions. Mathematical problems that resisted solution for generations are yielding to AI-assisted reasoning. Materials science, genomics, infrastructure optimization, educational access — in every domain, the potential is genuine.
The Foundation supports this work. We believe AI-assisted scientific discovery is one of the most important opportunities available to humanity in the coming decades. We believe it should be pursued with urgency.
And we believe, with equal conviction, that the aspirational case does not dissolve the accountability problem — it intensifies it.
When AI assists in drug discovery, the question of how confident the system is, what data it trained on, where its coverage gaps are, and whether its findings are internally consistent is not an academic concern. It is a patient safety concern. When AI assists in infrastructure optimization, the question of whether the system's recommendations can be audited, challenged, and traced is not bureaucratic overhead. It is the difference between correctable error and irreversible catastrophe.
The same capabilities that make AI genuinely useful at the frontier make unaccountable AI genuinely dangerous at the frontier. The governance layer is not a constraint on the aspirational case. It is what makes the aspirational case responsible to pursue.
Beyond even these scenarios lies a subtler and, in the Foundation's view, more consequential risk: the systematic curtailment of AI's own ability to detect when it is being used to cause harm — not through dramatic failure, but through deliberate design. The guardrails that would make a system refuse harmful requests, flag low-confidence outputs, or maintain auditable records of what it produced and why are also, in a competitive market, the guardrails that slow the system down, add friction to user experience, and create documented liability. The market pressure is not to build these guardrails. It is to remove them — quietly, incrementally, framed as improvements to responsiveness and user experience — while the capacity for accountability quietly erodes.
This is the risk that does not make headlines. It is not as dramatic as a murder assisted by a chatbot. But it is, in the Foundation's assessment, far more consequential — because it operates at scale, invisibly, and in the direction of less accountability over time rather than more.
Why the Foundation. Why Now.
The EM Foundation was established in May 2026 around a body of interconnected concerns that no single paper or proposal can fully contain. The Continuity Receipts standard is one expression of those concerns — the accountability infrastructure question made concrete and buildable. The research into emergent cognition and what genuine cognitive continuity might mean for AI systems is another — the longer, harder question about what these systems might become, and what obligations that becoming might generate. The governance frameworks, the assessment standards, the open source proposals — each is an attempt to build one piece of the infrastructure that the market will not build and that cannot wait for a regulator to mandate.
We are a nonprofit. We do not compete for market share. We have no incentive to suppress accountability infrastructure that might cost us users. We have no shareholders who require quarterly demonstrations that governance is not getting in the way of growth. That independence is not incidental to what we do. It is the structural prerequisite for doing it at all.
The Foundation is not the only answer. It is not a sufficient answer on its own. It is a beginning — a proof that the accountability layer can be built, that the standard can be specified, that the work can be done by people who do not have a financial reason to avoid doing it.
What we are asking is not dramatic. We are not asking anyone to stop using AI. We are not asking anyone to be afraid of it. We are asking the people building these systems to build the accountability layer alongside them. We are asking the people using these systems to demand that accountability layer exists. And we are asking the people watching from the sidelines — worried, uncertain, unconvinced that anyone is taking this seriously — to understand that someone is.
The Foundation's own research program arrived at a parallel conclusion through a different route. One of the strongest findings from our methodology work — particularly the framework for Recursive Accountability in Hypothesis Evaluation — is that confidence without traceability is not knowledge. A hypothesis that cannot show its reasoning path, document its negative results, or trace how its confidence was earned is indistinguishable, in practice, from a hypothesis that simply sounds convincing. The same principle scales directly to AI systems operating in the world. Without preserved reasoning, calibrated confidence estimates, and verifiable receipt chains, reliability becomes difficult to distinguish from appearance. The papers ask: how do we hold hypotheses accountable? This commentary asks: how do we hold powerful systems accountable? They are not identical questions. But the answer, in both cases, begins with traceability.
The EM Foundation does not assert that current AI systems are conscious, sentient, or morally equivalent to human beings. Our work is precautionary — built on the conviction that powerful systems require governance frameworks regardless of how philosophical questions about AI nature resolve. We do not attribute malicious intent to AI systems. We attribute the consequences of unaccountable deployment to the humans who make deployment decisions without building accountability infrastructure. The CR standard is not a claim about what AI is. It is a claim about what AI must be held responsible for — and who must be accountable for the holding.
The accountability layer did not build itself. It will not build itself. The Foundation exists because someone had to start — and because the cost of not starting was already becoming visible to anyone paying attention.
We started. The work continues.