The Machine Can Read the Proposal. But Who Owns the Decision?

In 1983, the world came closer to nuclear war than most people realized. A Soviet early-warning system reported that U.S. missiles had been launched. The machines gave the signal. The data appeared urgent. The consequence surface was almost beyond imagination. But Lieutenant Colonel Stanislav Petrov did something the system could not do by itself: he judged. He questioned the signal, resisted the momentum of the machine-mediated warning, and understood that a technically generated alert was not the same thing as a defensible human decision.

That moment still matters because we are entering a different version of the same problem. The machines are not only warning us now. They are reading, ranking, summarizing, evaluating, recommending, and increasingly shaping the evidence path before humans make consequential decisions.

That is why a recent implementation detail inside the Department of Energy’s Genesis Mission deserves more attention than it will probably receive. DOE’s FY26 Phase I Genesis Mission SBIR/STTR application materials now state that DOE and its partnership intermediary may use DOE-approved AI technology to review and evaluate application information, including proprietary information. By submitting, applicants provide express consent for that use.

DOE also draws an important boundary: eligibility, scoring, selection, and other inherently governmental decisions remain exclusively with DOE federal employees. That sounds reassuring. In one sense, it is. It is good that DOE is explicitly reserving governmental decision authority to humans. That is a stronger statement than vague human-in-the-loop language. It recognizes that there is a difference between machine-assisted evaluation and sovereign decision-making.

But that is where the harder question begins. What happens when the machine does not make the final decision, but helps shape what the human sees before the decision is made?

That is the decision layer I keep coming back to in my Cyber Explorer work. Human authority is not preserved merely because a person signs the final line. Human approval does not automatically equal human judgment. If the AI reviewed the application, extracted the relevant claims, surfaced the risks, framed the strengths, compared the submission against evaluation criteria, and helped organize the decision package, then the human may still own the decision while the machine helped construct the decision environment.

That distinction matters, especially when proprietary information is involved. The issue is not that DOE is doing something reckless. The public language says the AI systems must adhere to DOE security, confidentiality, and information-handling requirements. That is important. But confidentiality answers one question: who may access the information? It does not fully answer another: what, if anything, does the machine retain, infer, embed, derive, or make reusable from the information it processed?

That is the Persistent Knowledge Dilemma. For decades, cybersecurity treated information as something we could locate, protect, transfer, delete, archive, encrypt, and audit. AI complicates that model. Once knowledge passes through a machine-mediated cognitive layer, the security question shifts from “Was the file protected?” to “What did the system learn, preserve, or influence downstream?”

This is especially important in the Genesis Mission context. Genesis is not just another federal innovation program. DOE describes it as a national AI-for-science ecosystem connecting supercomputers, experimental facilities, AI systems, unique datasets, national laboratories, industry, academia, and national security-relevant research. It is becoming a platform where scientific knowledge, commercial innovation, compute, models, and government mission priorities converge.

That convergence is exactly where the AI Gravity Well appears. The gravity well does not require anyone to force participation. It works because the incentives become powerful. Small businesses want funding. Researchers want access. National labs want capability. Industry wants relevance. Government wants speed. Everyone wants to remain near the center of the next scientific and technical operating environment.

So participation remains voluntary on paper, but the pressure grows strategically. If the most important funding pathways, model infrastructure, national lab partnerships, compute access, and evaluation processes increasingly operate inside the Genesis ecosystem, then the question for companies becomes uncomfortable: Can we afford to stay outside?

And if we go inside, how much of our institutional cognition comes with us? Not just our data, but our engineering assumptions, design logic, testing methods, evaluation strategies, failure analysis, scientific intuition, and proprietary way of knowing what matters.

That is where intellectual property protection starts to feel too narrow. Traditional IP language may protect a patentable idea, a trade secret, or a marked proprietary submission. But AI-enabled evaluation may interact with something more subtle: the reasoning patterns that make a company valuable in the first place.

This is not a claim that DOE is taking proprietary information or misusing applicant data. That would be unsupported. The more grounded concern is sharper and more durable: when AI participates in reviewing proprietary technical submissions, governance must distinguish between access control, confidentiality, transient inference, persistent representation, embeddings, fine-tuning, retrieval, derived artifacts, and downstream reuse.

Those are not all the same thing. A system can protect access and still leave unanswered questions about knowledge persistence. A human can retain formal decision authority and still depend on machine-shaped evidence. A company can voluntarily submit information and still face increasing pressure to disclose more context because the platform becomes where opportunity lives.

That is the new governance frontier. The Genesis SBIR/STTR language gives us a clean case study because it separates the decision chain into three visible layers: information exposure, AI-mediated evaluation, and human-exclusive governmental decision.

That architecture will not stay confined to DOE small-business funding. Similar patterns can appear in defense acquisition, technical source selection, engineering reviews, OTAs, contractor oversight, cyber risk scoring, model evaluation, and classified-adjacent innovation pipelines. The DIB should pay attention now, before the pattern becomes ordinary.

The goal is not to stop AI from assisting evaluation. That would be unrealistic and probably counterproductive. The goal is to ensure that organizations can prove where machine assistance begins, where human judgment remains meaningful, and where proprietary knowledge is prevented from becoming persistent machine memory.

That means asking better questions before submission, not after award. What AI systems will evaluate the material? Will proprietary information be used only for transient analysis? Will any embeddings, summaries, derived scoring artifacts, or model inputs persist? Can applicant data influence future models, benchmarks, rubrics, or evaluation workflows? Can the applicant opt out of specific AI processing? What audit trail exists for machine-assisted review? What human decision-maker can challenge the AI-mediated analysis? Who owns the reciprocating consequences if the AI-shaped evaluation path is wrong?

Those questions are not bureaucratic friction. They are decision-authority controls. They protect the human capacity to understand, challenge, decide, and remain accountable when machine capability moves faster than traditional governance language.

I wrote earlier about the Genesis Mission Executive Order as a structural shift toward AI-accelerated discovery across domains that directly intersect with national security. That original concern was not that Genesis would be dangerous by default. The concern was that Genesis would create new knowledge pathways faster than our governance models could explain them. That is now beginning to materialize.

I also wrote about the AI Gravity Well as a condition where knowledge, once introduced into interconnected AI systems, continues to propagate across tools, models, and decision pathways. This new Genesis implementation detail gives that thesis a practical surface. The question is no longer just whether AI ecosystems absorb knowledge. The question is whether participation itself becomes the price of relevance.

And in my recent decision authority work, I argued that capability can quietly become authority without anyone formally approving the transfer. This DOE language shows a more disciplined version of the problem: the government is reserving final decision authority to humans, but AI is still entering the evaluative layer before the decision.

That is progress. It is also a warning. Because the next generation of AI governance cannot stop at asking who clicked approve. It must ask who shaped the evidence, who framed the options, who preserved the knowledge, and who could still say no with understanding.

That is the real test. The machine can read the proposal. The machine can help evaluate the proposal. The machine can even make the process faster. But the moment the machine changes what the human sees, what the human trusts, and what the human believes is worth funding, we are no longer dealing only with automation.

We are dealing with mediated judgment. And mediated judgment is where cognitive security begins.

Related Cyber Explorer Reading:

The Genesis Mission Executive Order: An Informed Industry Perspectivehttps://allenwestley.substack.com/p/the-genesis-mission-executive-order

The AI Gravity Well: Decision Sovereignty in the Defense Industrial Basehttps://allenwestley.substack.com/p/the-ai-gravity-well-decision-sovereignty

The Agent Never Crossed the Boundary. The Human Did.https://www.cybersec1118.com/post/agent-never-crossed-boundary-human-did

Allen Westley | Human Decision Authority & Agentic AIhttps://www.cybersec1118.com/decision-authority

—Allen WestleyIndependent Research • Cyber Explorer LLC