AI at Speed, Humans at Risk

The ubiquity of AI is perhaps matched only by its’ banality[i]. For most of society, there is no longer a meaningful opt-out to its reach. AI’s utility as a cost-cutting accelerant has – for the moment – dwarfed discussions of more pedestrian things such as accuracy or effectiveness. In national security, it is increasingly treated as both the illness and the cure; adversaries’ use of it mandates our development in a self-reinforcing cycle that circles around the question of ‘why’[ii]. The Defense Department has incentivized[iii] its’ workforce to use AI through GenAI.mil. This is bigger than it appears, as it has broader industry and education implications; it is ‘if you build it, they will come’ on the grandest scale. “Future generations will look back upon this time not as a statement of technology but rather an extensive examination of what it means to be human.[iv]”

While the stakes are different in the Public and Private Sectors, the fundamental issues remain the same. The question for career professionals in analytics such as us, especially for high-stakes decisions, is not whether AI will be used, but what standards we will demand before we let speed substitute for judgment. Furthermore, we argue that current policies give AI an unfair advantage, and that taxation schemes should move from taxing labor to taxing results. In short, the answer to the question ‘should robots pay taxes’ is an unequivocal “Yes”.

When we don’t use AI to solve decision problems, statisticians have a battery of ‘goodness’ measures which are all different spins on the same underlying questions, being:

AI has inexplicably slipped past these questions and nowhere more than in the realm of Large Language Models (LLMs). There is a notion among decision makers at all levels that AI is somehow ‘better’ by virtue of its’ existence. Thus, AI receives an unfair advantage when compared to other scientific methods. We have had numerous engagements with clients who express joy at the fact that their model has generated ‘an answer’ without a discussion around how ‘good’ that answer may or may not be[viii]. This is because usability is either implicitly or explicitly the most important metric when developing an LLM. How much people like interacting with it is the main thing; correctness is only relevant to the quality of interaction. No LLM in current use exemplifies this more than DeepSeek[ix].

Against this backdrop, professionals in industry and government are asking the same fundamental questions. First, how much ‘nug work’ (some people call this ‘feeding pigeons’) can we offload to a computer to make our work more focused, and secondly, how can we be sure that the answers we get back from AI are trustworthy?

AI is already being used to process and analyze volumes of data at speed and scale previously inaccessible to decision-makers with a dozen novice to journeyman analysts. Thus “AI will not replace you, but the professional who uses it to work faster and more effectively may.” Cold comfort for those of us not completely on the AI train. AI for managers is like aeronautical engineering for pilots; building the machine and using the machine are fundamentally different tasks; however, the incentives of those who build may (or may not) be well matched to those who use it. In physical systems, disconnects between development and use are generally obvious, e.g. the plane doesn’t fly like we expected it to. We have (yet) to develop a good technique for determining when your machine does not think like we expected it to.

National security frequently requires a balance between urgency and accuracy that does not always tilt towards accuracy. The ability to act swiftly can offer a significant strategic advantage, making speed a vital asset when addressing critical threats or opportunities. AI, when well trained, excels at identifying known patterns, generating insights, and synthesizing data at speed and scale. The key ingredients in the preceding sentence were ‘when well trained’ and ‘identifying known patterns.’ The ability to make an inference with little to no data, or when the pattern is unknown, remains the domain of Natural Intelligence, or in a word, humans. What if ‘thinking’, means a little bit more.

It is hard for responsible performers to build an AI to responsibly perform an inherently human task, such as firing employees. Unfortunately, it is almost trivial for irresponsible performers to build an AI to do a crap job at it. Given our comments above about customer service and ‘enshittification’, we are not optimistic about the future of

AI deployments in national security, healthcare, criminal justice, and other high-stakes domains requiring oversight. Unfortunately, in the current environment, these AI misses tend to be obfuscated. The critical question is ‘when do we stop chalking failures up to development and actually call them failures.’ This is another way in which AI has a perverse advantage over humans.

The line between humans and AI is a source of friction. Too many professionals are developing an (incorrect!) preference for AI over humans. We soundly reject this primase and adopt a ‘Carbon Supremacist[xi]’ point of view, which in summary states that Natural Intelligence has primacy and is superior without proof, and relegates the role of silicon-based intelligences to be subservient and supporting. The confluence of forces pushing us in this direction are partly societal, partly economic, and partly workforce development[xii]. We have partly considered the workforce piece previously and leave the societal piece to experts in that field[xiii].  We briefly consider the economic incentives for AI.

On the surface, AI is ‘better’ than humans because it does not get sick, demand a promotion, take holidays or retire. When you are finished with it, you can simply throw it away. Furthermore, businesses do not pay taxes on work that is produced by machines, while they do have a tax related to the work done by humans (e.g. payroll taxes). Companies get tax incentives to invest in capital and for depreciation on their equipment, both of which can directly contribute to AI systems. These are outdated, industrial-era policies that unfortunately do nothing to adjust for increased reliance on AI over humans. Taxation as a policy tool is frequently used to shape certain activities by a government’s citizens, which are humans. Examples are home ownership and retirement savings. The Government should explore the implications of taxing work performed by AI at a rate equal to if it were done by a human. This could also be extended into taxing usage of AI in a similar fashion to road taxes. AI requires infrastructure that is creating burdens on traditional utilities and customers they serve.

AI has moved from a tool used by programmers to an intrusive guest in our everyday lives. The danger is not that machines are becoming too powerful, but that speed, convenience, and misplaced deference allow humans to become too weak, bypassing judgment, accountability, and ethics. Future humans – which we remain optimistic towards – will see this moment as a test of human resilience rather than a surrender. If we insist on evidence, demand explainability, train the next generation to think before they automate, and align economic incentives with responsibility rather than replacement, AI can remain what it should be: an amplifier of human intelligence, not a substitute for it. The future is not determined by how clever our machines become, but by whether we have the discipline to decide when (and when not) to trust them.

Nick Ulmer is a Principal Analyst at CANA, LLC and a retired Navy Supply Corps Officer. He is President of the Military Operations Research Society.

Harrison Schramm is Chief Technologist at Attritable Machines, LLC and a retired Naval Aviator. He is a Fellow of the Military Operations Research Society.

[i] https://medium.com/@kambizhoma/the-banality-of-ai-415a3e8d6a0a

[ii] For discussion of National Security and AI, we recommend the NSCAI report, which interested professionals should read in its’ entirety.

[iii] https://www.war.gov/Spotlights/Artificial-Intelligence/

[iv] H. Schramm, First Word, Phalanx, Spring 2026

[v] https://hbr.org/2025/10/why-ai-at-work-makes-us-so-anxious

[vi] https://www.wsj.com/business/has-the-era-of-the-mega-layoff-arrived-928f061d?utm_source=substack&utm_medium=email

[vii] https://locusmag.com/feature/commentary-cory-doctorow-social-quitting/

[viii] Other techniques that produce answers without measures of quality include monkeys with typewriters and my cat with a Ouija board.

[ix] https://www.realcleardefense.com/articles/2025/08/19/deepseek_taiwan_and_the_knights_who_say_ni_1129577.html

[x] With apologies to The Grinch, who ‘puzzled and puzzled till his puzzler was sore’

[xi] https://thehill.com/opinion/technology/4874396-carbon-supremacist-ai/

[xii] https://thehill.com/opinion/technology/3504097-how-to-restore-americas-dominance-in-technology/

[xiii] https://warontherocks.com/making-a-u-s-digital-service-academy-work/