Chinese EVs are cheaper due to scale, supply chains, and subsidies Targeted EU tariffs and price floors correct subsidy-driven undercutting Link trade defense to localization, skills, and investment to protect competitiveness
Factory tours scaled US–Japan know-how Aerospace: from licenses to composite wings Open collaboration with firm research security
In the late 1950s and early 1960s, 400 to 500 Japanese firms sent managers on organ
UK GDP per head is 6–8% below its no-Brexit path The slow-burn hit was masked by transition rules and the pandemic Without lower frictions and restored mobility, the drag endures; the U.S.
AI tools exclude people through missing data and bugs Count “no-decision” cases and use less-exclusionary methods with human review Set exclusion budgets, fix data flows, and publish exclusion rates
A quiet fact sets the stakes. In 2024, a global HR survey found that 53% of organizations already use AI in recruitment, often to filter résumés before a human ever reviews them. Yet in a recent, legally required audit of a résumé-screening tool, 11.4% of applicant criteria were labeled “uncertain” and excluded from analysis. In another audit of a central matching platform, over 60 million applications had unknown race or ethnicity fields, making fairness hard to test and easy to overlook. When systems cannot decide or measure, people fall out of view. This is algorithmic exclusion, and it is not a rare issue; it is a structural blind spot that eliminates qualified applicants, obscures harm, and weakens trust in AI across education and work. We can reduce bias only if we also address silence—those moments when the model returns nothing or hides behind missing data-so stakeholders feel reassured that no one is invisible.
The familiar story about unfair algorithms focuses on biased predictions. But exclusion starts earlier. It begins when models are designed to prioritize narrow goals like speed, cost, or click-through rates, and are trained on data that excludes entire groups. It worsens when production code is released with common software defects. These design, data, and coding errors together push people outside the model’s view. Leading institutions now recognize “no prediction” as a significant harm. A recent policy proposal suggests that regulations should recognize algorithmic exclusion alongside bias and discrimination, as sparse, fragmented, or missing data consistently yield empty outputs. Suppose the model cannot “see” a person’s history in its data. In that case, it either guesses poorly or refuses to think entirely. Both outcomes can block access to jobs, courses, credit, and services—and neither situation shows up if we audit only those who received a score.
Exclusion is measurable. In face recognition, the U.S. National Institute of Standards and Technology has long recorded demographic differences in error rates across age and sex groups; these gaps persist in real-world conditions. In speech-to-text applications, studies show higher error rates for some dialects and communities, affecting learning tools and accessibility services. In hiring, places like New York City now require bias audits for AI-assisted selection tools. However, most audits still report only pass/fail ratios for race and gender, often excluding records with “unknown” demographics. This practice can obscure exclusion behind a wall of missing data, making it crucial for stakeholders to understand the silent failures that undermine fairness and transparency.
The numbers are precise. One résumé-screening audit from January to August 2024 examined 123,610 applicant criteria and reported no formal disparate impact under the four-fifths rule. However, it also showed that 11.4% of the applicant criteria were marked “uncertain” and excluded from the analysis. Among the retained records, selection rates differed significantly: for instance, 71.5% for White applicants versus 64.4% for Asian applicants at the criterion level. Intersectional groups like Asian women had impact ratios in the mid-80s. These gaps may not reach a legal threshold, but they indicate drift. More critically, the excluded “uncertain” pool poses a risk: if the tool is more likely to be uncertain about non-linear careers, school re-entrants, or people with fragmented data, exclusion becomes a sorting mechanism that no one chose—and no one sees.
Scale amplifies the problem. A 2025 audit summary for a large matching platform listed over 9.5 million female and 11.9 million male applicants, but also recorded 60,263,080 applications with unknown race or ethnicity and more than 50 million with at least one unknown demographic field. If fairness checks depend on demographic fields, those checks become weakest where they are most needed. Meanwhile, AI in recruiting has become common: by late 2024, just over half of organizations reported using AI in recruitment, with 45% specifically using AI for résumé filtering. Exclusion at a few points in extensive talent funnels can therefore deny thousands of qualified applicants a chance to be seen by a human.
Regulators are taking action, but audits must improve transparency to reassure stakeholders that fairness is being actively monitored. The EEOC has clarified that Title VII applies to employers’ use of automated tools and highlighted the four-fifths rule for screening disparities. The OFCCP has instructed federal contractors to validate AI-assisted selection procedures and to monitor potential adverse impacts. New York City’s Local Law 144 requires annual bias audits and candidate notification. These are fundamental steps. However, if audits lack transparency about “unknown” demographics or “uncertain” outputs, exclusion remains hidden. Education systems face similar issues: silent failures in admissions, course placement, or proctoring tools can overlook learners with non-standard records. Precise, transparent measurement of who is excluded is essential for effective policy and practice.
First, count silence. Any audit of AI-assisted hiring or educational tools should report the percentages of “no decision,” “uncertain,” or “not scored” outcomes by demographic segment and by critical non-demographic factors such as career breaks, school changes, or ZIP codes with limited broadband. This specific measurement approach helps stakeholders identify hidden biases. It also discourages overconfident automation: if your tool records a 10% “uncertain” rate concentrated in a few groups, having a human in the loop is not just a courtesy; it is a safety net. The Brookings proposal to formalize algorithmic exclusion as a category of harm provides regulators a tool: require “exclusion rates” in public scorecards and classify high rates as non-compliance without a remediation plan.
Second, make less-discriminatory alternatives (LDAs) the standard practice. In lending, CFPB supervision now scrutinizes AI/ML credit models, and consumer groups advocating for LDA searches argue these should be part of routine compliance. This same logic applies to hiring and education. If a ranking or filtering algorithm shows negative impacts or high exclusion rates, stakeholders should test and document an equally effective method that excludes fewer candidates. This could involve deferring to human review for edge cases, replacing rigid résumé parsing with skills-based prompts, or using structured data from standardized assignments instead of unclear proxies. Prioritizing the reduction of exclusion while ensuring job-relatedness guides organizations toward fairer AI practices.
Third, fix the code. Many real-world harms stem from ordinary defects, not complex AI math. The best estimate of the U.S. cost of poor software quality in 2022 is $2.41 trillion, with defects and technical debt as significant factors. This impacts education and HR. Flaws in data pipelines, parsing, or threshold logic can quietly eliminate qualified records. Organizations should conduct pre-deployment defect checks across both the data and decision layers, not just the model itself. Logging must ensure unscored cases are traceable. And when a coding change raises exclusion rates, it should be rolled back as we would treat security regressions. Quality assurance directly relates to fairness assurance.
Educators and administrators should identify where their systems fall short in delivering results. Begin with placement and support technologies: reporting dashboards must indicate how many students the system cannot score and for whom. If remote proctoring, writing assessments, or early-alert tools have trouble with low-bandwidth connections or non-standard language varieties, prioritize routing those cases to human review. Instructors should have simple override controls and a straightforward escalation process. At the same time, capture voluntary, privacy-respecting demographic and contextual data and use it only for fairness monitoring. Suppose the “unknown” category is large. In that case, it signals a need to improve data flows, not a reason to disregard it in analysis.
Employers can take a similar approach. Require vendors to disclose exclusion rates, not just impact ratios. Refuse audits that omit unknown demographics without thorough sensitivity analysis. For high-stakes roles, set an exclusion budget—the maximum percentage of applicants who can be auto-disqualified or labeled “uncertain” without human review. Use skills-based methods and structured work samples to enlarge the data footprint for candidates with limited histories and log instances when the system cannot parse a résumé, allowing candidates to correct their records. Lastly, follow legal regulations and the direction of change: comply with New York City’s bias-audit rules, the EEOC's Title VII disparate-impact guidelines, and the OFCCP's validation expectations. This is not mere compliance; it represents a shift from one-time audits to ongoing monitoring of the areas where models often falter.
The key statistic that should dominate every AI governance discussion is not only who passed or failed. It is those who were not scored—the “uncertain,” the “unknown,” the missing individuals. In 2024–2025, we learned that AI has become integral in the hiring process for most firms, even as audits reveal significant gaps in unmeasured applicants and unresolved ambiguities. These gaps are not mere technicalities. They represent lost talent, overlooked students, and eroded policy legitimacy. By treating algorithmic exclusion as a reportable harm, establishing exclusion budgets, and mandating less-exclusionary alternatives, we can keep automation fast while ensuring it is fair. This should be the guiding principle for schools and employers: ongoing measurement of silent failures, human review for edge cases, and transparent, public reporting. We do not have to wait for the perfect model. We can take action now to recognize everyone that the algorithms currently overlook.
The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.
Ashby. (2024, August). Bias Audit for Ashby’s Criteria Evaluation model. FairNow.
Consumer Financial Protection Bureau. (2024, June). Fair Lending Report of the Consumer Financial Protection Bureau, FY 2023.
Consumer Reports & Consumer Federation of America. (2024, June). Statement on Less Discriminatory Algorithms.
Eightfold AI. (2025, March). Summary of Bias Audit Results (NYC LL 144). BABL AI.
Equal Employment Opportunity Commission. (2023). Assessing adverse impact in software, algorithms, and AI used in employment selection procedures under Title VII.
HR.com. (2024, December). Future of AI and Recruitment Technologies 2024–25.
New York City Department of Consumer and Worker Protection. (2023). Automated Employment Decision Tools (AEDT).
NIST. (2024). Face Recognition Technology Evaluation (FRTE).
OFCCP (U.S. Department of Labor). (2024). Guidance on federal contractors’ use of AI and automated systems.
Tucker, C. (2025, December). Artificial intelligence and algorithmic exclusion. The Hamilton Project, Brookings Institution.
Consortium for Information & Software Quality (CISQ). (2022). The Cost of Poor Software Quality in the U.S.
Koenecke, A., Nam, A., Lake, E., et al. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14), 7684–7689.
Ageing Europe needs an optimal immigration level Set ~0.6–1.0% yearly, adjusted by jobs, housing, and language Link flows to capacity and invest to keep growth and trust
Across the EU, the old-age dependency
Closing gender and 60+ gaps offsets ageing Frontiers need more hours; laggards need jobs and childcare Use 5-year targets, neutral taxes, late-career training
In 2024, the European Union still had a gender employment gap of 1
Support is broad, underestimated Local clean jobs drive votes Show local data; build pipelines
The most crucial fact in climate politics today is not about a molecule or a megawatt. It’s a number: 89%.
Export controls slow foes, not secure leadership Invest in compute, clean power, talent Make NAIRR-style Compute Commons permanent
U.S. data centers used about 183 terawatt-hours of electricity in 2024, roughly 4% of total U.S. power consumption. This figure is likely to more than double by 2030. The United States already accounts for 45% of global data-center electricity use. These numbers reveal a clear truth: the future of American AI relies less on what leaves our ports and more on whether we can provide the computing power, energy, and talent needed to build and use systems domestically. While AI chip export controls may seem practical, they are not the deciding factor in who leads. They can slow rivals marginally, but do not build labs, wire campuses, or train students. We must focus on transforming energy, infrastructure, and education into a self-reinforcing system. With every semester we delay, the cost of missed opportunities increases, and our advantage shrinks.
Recent proposals suggest a 30-month ban on licensing top accelerators to China and other adversaries, formalizing and extending the Commerce Department’s rules. The goal is straightforward: restrict access to advanced chips and make competitors fall behind. However, the policy landscape is already complicated. In 2023 and 2024, Washington tightened regulations; in 2025, Congress discussed new “SAFE CHIPS” and “Secure and Feasible Exports” bills; and the House considered a GAIN AI Act adding certification requirements for export licenses. These measures mainly solidify what regulators are already doing. They increase compliance costs but may complicate enforcement. They could also provoke reciprocal actions abroad and push trade into unclear channels, making it harder to monitor.
These unclear channels are real. According to a Reuters report, between April and July 2025, more than $1 billion worth of Nvidia AI chips entered China via black market channels despite strict U.S. export restrictions. Export-compliant “China-only” chips came and went as rules changed; companies wrote down their inventory; and one supplier reported no sales of its redesigned parts to China in a recent quarter after new licensing requirements were enforced. Controls produce effects, but those effects are messy, leaky, and costly domestically. In summary, AI chip export controls disappoint security advocates while imposing significant collateral damage at home.
There’s a second issue: innovation follows scale. U.S. semiconductor firms allocate a significant portion of their revenue to research and development, averaging around 18% in recent years and 17.7% in 2024. One leading AI chip company spent about $12.9 billion on R&D in 2025, despite rising sales, which lowered that percentage. That funding supports new architectures, training stacks, and tools that benefit universities and startups. A shrinking global market, as mandated by law, threatens the reinvestment cycle, particularly for suppliers and EDA tool companies that rely on the growth of system integrators.
However, supporters point out that the U.S. still commands just over 50% of global chip revenues, which they believe is strong enough to sustain industry leadership, according to the Semiconductor Industry Association. They cite record data-center revenue, long waitlists, and robust order books. All of this is true—and it underscores the argument. When companies are forced to withdraw from entire regions, they incur losses on stranded “China-only” products and experience margin pressure. Over time, these challenges affect hiring strategies, supplier decisions, and university partnerships. One research estimate shows China’s share of revenue for a key supplier dropping to single digits by late 2025; another quarter included a multi-billion-dollar write-off related to changing export rules. Innovation relies on steady cash flow and clear planning goals. AI chip export controls that fluctuate year to year do the opposite. The real question is not “ban or sell.” It’s about minimizing leakage while maintaining domestic growth.
China is making rapid progress. Domestic accelerators from Huawei and others are being shipped in volume; SMIC is increasing its advanced-node capacity; and GPU designers are eager to go public on mainland exchanges. One firm saw its stock rise by more than 400% on debut this week, thanks to supportive policies and local demand. Nevertheless, impressive performance on paper does not necessarily translate into equal capability in practice. Reports highlight issues with inter-chip connectivity, memory bandwidth, and yields; Ascend 910B production faces yield challenges around 50%, with interconnect bottlenecks being just as significant as raw performance for training large models. China can and will produce more, but its path remains uneven and costly, especially if it lacks access to cutting-edge tools. A CSIS report notes that when export controls are imposed, as seen with China, the targeted country often intensifies its own development efforts, which could lead to significant technological breakthroughs. This suggests that while export controls can increase challenges and production costs, they do not necessarily prevent a country from closing the competitive gap.
Where controls create barriers, workarounds appear. Parallel import networks connect orders through third countries; cloud access is negotiated; “semi-compliant” parts proliferate until rules change again. This ongoing dynamic strengthens China’s motivation to develop domestic substitutes. In essence, strict bans can accelerate domestic production when they are imposed without credible, consistent enforcement and without additional U.S. investments that push boundaries in a positive direction. In 2024–2025, policymakers proposed new enforcement measures, such as tamper-resistant verification and expanded “validated end-user” programs for data centers. This direction is right: smarter enforcement, fewer loopholes, and predictability, along with significant investment in American computing and power capacity for research and education.
If the United States aims for lasting dominance, it needs a national education and computing strategy that can outpace any rival. The NAIRR pilot, initiated in 2024, demonstrated its effectiveness by providing researchers and instructors with access to shared computing and modeling resources. By 2025, it had supported hundreds of projects across nearly every state and launched a 'Classroom' track for hands-on teaching. This is more important than it may seem. Most state universities cannot afford modern AI clusters at retail prices or staff them around the clock. Shared infrastructure transforms local faculty into national contributors and provides students with practical experience with the same tools used in industry. Congress should make NAIRR permanent with multi-year funding, establish regional 'Compute Commons' driven by public universities, and link funding to inclusive training goals. According to a report from the University of Nebraska System, the National Strategic Research Institute (NSRI) has generated $35 in economic benefits for every $1 invested by the university, illustrating a strong return on investment. This impressive outcome underscores the importance of ensuring that all students in public programs can access and work with real hardware as a standard part of their education, rather than as an exception.
Computing without power is just a theory. AI demand is reshaping the power grid. U.S. computing loads accounted for about 4% of national electricity in 2024 and are projected to more than double by 2030. Commercial computing now accounts for 8% of electricity use in the commercial sector and is growing rapidly. One utility-scale deal this week included plans for multi-gigawatt campuses for cloud and social-media operators, highlighting the scale of what lies ahead. Federal and state policy should view university-adjacent power as a strategic asset: streamline connections near public campuses, create templates for long-term clean power purchase agreements that public institutions can actually sign, and prioritize transmission lines that connect "Compute Commons" to low-carbon energy sources. By aligning clean-power initiatives with campus infrastructure, we could spark the development of regional tech clusters. This would not only enhance educational capabilities but also attract industry partners eager to capitalize on a well-connected talent pool. The talent-energy feedback loop becomes essential, creating synergies that can broaden support far beyond energy committees. When AI chip export controls dominate the discussion, this vital bottleneck is often overlooked. Yet it determines who can teach, who can engage in open-science AI, and who can graduate students with real-world experience using production-grade systems.
Leadership is created, not blocked. It thrives in labs, on the grid, and within public universities, preparing the next generation of innovators. To ensure continued progress and maintain our competitive edge, we must embrace policies that amplify our strengths while fostering innovation. It is critical to envision and reach the next policy milestone that inspires collective action and drives us toward a future where American leadership in AI is unassailable.
The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.
Bloomberg. (2025, Dec. 4). Senators Seek to Block Nvidia From Selling Top AI Chips to China.
Bloomberg
Brookings Institution. (2025, Dec. 3). Why the GAIN AI Act would undermine US AI preeminence.
Brookings
CSET (Georgetown). (2024, May 8). The NAIRR Pilot: Estimating Compute.
CSIS. (2024, Dec. 11). Understanding the Biden Administration’s Updated Export Controls.
CSIS. (2025, Nov. 6). The Architecture of AI Leadership: Enforcement, Innovation, and Global Trust.
ExecutiveGov. (2025, Dec.). Bipartisan House Bill Seeks to Strengthen Enforcement of AI Chip Export Controls.
Financial Times. (2025, Dec.). Chinese challenger to Nvidia surges 425% in market debut.
Financial Times via Reuters summary. (2025, Jul. 24). Nvidia AI chips worth $1 billion entered China despite U.S. curbs.
IEA. (2024, Jan. 24). Electricity 2024.
IEA. (2025, Apr. 10). Global data centre electricity consumption, 2020–2030 and Share by region, 2024.
IEA. (2025). Energy and AI: Energy demand from AI and Executive Summary.
MERICS. (2025, Mar. 20). Despite Huawei’s progress, Nvidia continues to dominate China’s AI chips market.
NVIDIA. (2025, Feb. 26). Financial Results for Q4 and Fiscal 2025.
NVIDIA. (2025, May 28). Financial Results for Q1 Fiscal 2026. (H20 licensing charge.)
NVIDIA. (2025, Aug. 27). Financial Results for Q2 Fiscal 2026. (No H20 sales to China in quarter.)
Pew Research Center. (2025, Oct. 24). What we know about energy use at U.S. data centers amid the AI boom.
Reuters. (2025, Dec. 8). NextEra expands Google Cloud partnership, secures clean energy contracts with Meta.
Reuters. (2025, Dec. 4). Senators unveil bill to keep Trump from easing curbs on AI chip sales to China.
Semiconductor Industry Association (SIA). (2025, Jul.). State of the U.S. Semiconductor Industry Report 2025.
SemiAnalysis. (2025, Sep. 8). Huawei Ascend production ramp. (SMIC capacity estimates.)
The Information (archived via SOPA). (2024, Aug. 12). Nvidia AI chip smuggling to China becomes an industry.
Tom’s Hardware. (2025, Jul. 25). Underground China repair shops thrive servicing illicit Nvidia GPUs.
Tom’s Hardware. (2025, Dec.). Nvidia lobbies White House… lawmakers reportedly reject GAIN AI Act.
U.S. Bureau of Industry and Security (BIS). (2025, Jan. 15). Federal Register notice on Data Center Validated End-User program.
U.S. EIA. (2025, Jun. 25). Electricity use for commercial computing could surpass other uses by 2050.
The Wall Street Journal, LA Times, and other contemporaneous reporting on lobbying dynamics (cross-checked for consistency).
Defense AI education is the bottleneck between record defense spending and real capability Train stack-aware teams with mission credentials Scale safely with oversight and shared compute
In 2024, global military spending soared to $2.718 trillion, the most significant annual rise since the Cold War. Simultaneously, Ukraine's frontline units lost 10,000 drones each month. This marks a shift in warfare from bodies to machines. Together, these facts highlight a critical issue: we do not educate enough people to direct, manage, and improve the software that now controls the pace and reach in the field and across defense agencies. Procurement budgets are racing ahead, while practice is falling behind. “Defense AI education” represents the gap between money and real outcomes—the difference between countless systems and genuine advantage. It now determines how fast ministries learn from battlefield data, how quickly logisticians adapt policies, and whether future officers can manage drone swarms as confidently as past leaders handled platoons. If we fail to teach and credential on a large scale, the drones will operate independently, unethically, and without impact.
The past two years have driven this point home. NATO’s European members and Canada boosted defense spending by 18% in 2024. Twenty-three allies hit the 2% GDP target, more than double the number four years ago. Europe is urgently launching new industrial programs under EDIS and EDIP to push funds toward production and joint procurement. But time is running out: money alone will not deliver strategic gains. Training pathways, credentialing standards, and access to computing resources now determine whether AI tools can impact operations before the pace of threats accelerates further. That is what we must define as defense AI education: not just courses but rapid training to transform procurement, maintenance, planning, and policy toward faster, safer, and more accurate decisions. Note: the NATO figures reflect official ally submissions; EDIS/EDIP documents are legislative and program materials, not survey estimates.
We already have a proof of concept in the U.S. Department of Defense. According to Brookings, GAMECHANGER is an AI-enabled search and association platform for policy that was developed to process unclassified policy documents and improve the management of complex directives and authorities. The lesson here is not that one tool can solve all problems. Instead, enterprise AI adoption increases when users can find, test, and tailor software for specific tasks, then apply that knowledge in the future. A report from the Brookings Institution discusses the importance of addressing workforce challenges and modernizing education in national security fields, emphasizing the need for effective recruitment, retention, and up-to-date training to prepare for emerging technologies such as AI. Teach policy staff how to question models. Teach program managers how to assess task-level value. Teach leaders to maintain tools through rotations. Note: usage numbers and institutional details are from the Brookings analysis; the scope is non-classified and focused on the enterprise side.
The battlefield is already teeming with affordable autonomous systems. RUSI reports that Ukraine loses 10,000 drones per month, largely to electronic warfare—an urgent sign that success relies on learning, adapting, and replenishing at scale, not on a single advanced platform. U.S. policy has responded with haste: the Replicator initiative aims to deploy “thousands of autonomous systems” across all domains within the next 18 to 24 months. Funding, experimentation, and transition updates will come fast through late 2024. Europe’s industrial surge under EDIS and EDIP likewise calls for immediate mass production and speed. Defense AI education must not wait; it is the critical link that connects swarm tactics, resilient communications, electronic warfare-aware autonomy, and agile teamwork between humans and machines. Note: drone loss data comes from RUSI’s open-source report; Replicator details are based on official DoD announcements and analyses.
This training must reach beyond operators. The key advantage lies with “stack-aware” talent—people who understand sensors, data links, model behavior, and mission outcomes as a cohesive unit. Industry evidence supports this view. BCG estimates the aerospace and defense sector spent about $26.6 billion on AI in 2024, roughly 3% of its revenue. Yet 65% of programs remain stuck in the proof-of-concept stage. Value emerges when user-friendly solutions, domain-tailored models, and redesigned workflows come first—not yet another data pool. To translate this into curricula, it means creating capstone projects that assess mission-based return on investment rather than just model accuracy. It also means teaching students to work with electronic warfare-decoyed data, degraded GPS signals, or high-latency environments. Note: the spending and maturity estimates are from BCG’s 2025 sector report and accompanying PDF; these figures are based on surveys and modeling.
Defense AI education should shift from specialized fellowships to widespread programs with quick time-to-value. First, establish mission studios that reflect real industry tasks. Pair an air base logistics dataset with a policy graph, and ask students to create prompts and agents that reduce parts approval time while adhering to ethical and legal guidelines. According to Purdue University, the Anvil supercomputer joined the National AI Research Resource (NAIRR) Pilot in May, providing expanded access to advanced computing power. Adopting successful methods like GAMECHANGER and integrating NAIRR resources could allow public universities without extensive infrastructure to conduct multi-GPU experiments and improve machine learning operations education, potentially benefiting fields such as finance, healthcare, and maintenance in realistic conditions. According to the U.S. National Science Foundation, efforts to expand access to AI resources through the NAIRR pilot are strengthened by collaboration with organizations like Voltage Park; building on this momentum, there is a growing push to establish a standardized and trusted system of credentials that ranges from foundational micro-credentials for operators to advanced leadership residencies. Link each credential to mission-based test events so learning directly supports deployment needs. Leverage European and NATO programs to coordinate and standardize, and use resources like NAIRR to ensure access is affordable and widespread. The goal: make education actionable, efficient, and directly connected to defense priorities.
To address the moral concern that defense AI education could militarize universities, take these actions: require safety cases and ethics reviews, implement export-control education, and separate research focused on lethal applications from that aimed at enterprise performance or humanitarian protection. Also, make curricula and assessment criteria public. These safeguards create transparent processes and oversight, supporting responsible AI adoption in defense.
The economic critique—why train for AI at scale if organizations cannot absorb it—demands three steps: teach practical adoption ("absorption") in courses, focus on user-friendly solutions beyond technical models, and evaluate students on their ability to drive operational change and return on investment. Capstone projects should deliver measurable outcomes—reduced policy alignment time, fewer false alarms, and less maintenance downtime. Making value visible prompts leaders to invest in talent and projects that advance organizational goals.
To address the talent supply concern, implement three key recommendations: expand the recruiting pool through apprenticeships and flexible pay, create crossover programs for educators in math and computing, and define clear roles for AI-focused jobs, such as product owners, data engineers, safety testers, and mission designers. This will help meet urgent demand and broaden access to high-priority defense roles.
A final concern is strategic: Europe’s increase in defense spending may stem more from geopolitical anxiety than from a coherent strategy. This is true, and that’s why education must be the priority. NATO’s spending increases and the EU’s industrial strategy create an opportunity to standardize training, evaluation, and data sharing, ensuring coalitions can work together seamlessly. Shared syllabi, common credentials, and NAIRR-like computing partnerships can prevent a mix of incompatible tools. If we teach to a common standard now, the next crisis won’t force us to rebuild skills while under pressure. Note: NATO statements are official records; the EU roadmap is a Commission document; computing partnerships refer to NAIRR designs.
The numbers that began this essay should capture our attention. $2.718 trillion in global military spending. Drone losses in the tens of thousands each month on one front. These figures describe a world where software dictates pace, and where disruption, jamming, and losses penalize those who cannot learn quickly. The challenge for universities, ministries, and industry is whether education itself can expand. Defense AI education is how we make this shift—from purchasing to understanding, from pilots to practical application, from hype to real value. It involves more than just courses. It requires credentials linked to missions, computing resources reaching both provinces and capitals, and leadership training that views governance as a skill, not just paperwork. If we align budgets, programs, and standards now, we can transform quantity into quality and risk into resilience. If we do not, we will spend more and comprehend less—and see the gap widen with every flight hour and every update cycle. The choice is ours, and time is not on our side.
The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.
Boston Consulting Group. (2025). Three Truths About AI in Aerospace and Defense (report and PDF). Retrieved June 2025.
Brookings Institution. (2025, Dec. 1). GAMECHANGER: A case study of AI innovation at the Department of Defense.
European Commission. (2025, Nov. 19). EU Defence Industry Transformation Roadmap.
European Commission. (2024). European Defence Industrial Strategy (EDIS) and EDIP overview.
NATO. (2024, June 17–18). Remarks and joint press conference transcripts.
National Science Foundation (NSF). (2024). NAIRR Pilot announcements and resource access.
RUSI. (2023). Russia and Ukraine are filling the sky with drones (drone loss estimates).
SIPRI. (2025, Apr. 28). Trends in World Military Expenditure, 2024 (press release and fact sheet).
U.S. Department of Defense / DIU. (2023–2024). Replicator initiative overview and updates.
U.S. DoD / CDAO. (2024). Task Force Lima executive summary (public release).
U.S. Government reporting. (2024). Cyber workforce vacancy updates.
Agent AI is uneven—pilot before student use Start with internal, reversible staff workflows, add human-in-the-loop and logs Follow EU AI Act/NIST; publish metrics; scale only after proof
School leaders should pay attention to a clear contrast: nearly all teens have a smartphone, but agent AI still struggles with multi-step, real-world tasks. In 2024 surveys, 95% of U.S. teens reported having access to a smartphone at home. This infrastructure far outstrips any other learning technology. Meanwhile, cutting-edge web agents based on GPT-4 only completed about 14% of complex tasks in realistic website environments. Recent mobile benchmarks show some improvement, but even the most powerful phone agents completed fewer than half of 65 fundamental functions. This gap between universal access and uneven performance highlights a policy issue in education. Schools feel pressure to provide constant, on-device help, but agent AI can’t be counted on for unsupervised student use. The solution isn’t to ban or to buy these tools unthinkingly. Schools should require proof of reliability in controlled environments, publish the results, and expand use only when solid evidence exists.
The notable improvements in conversational models mask a stubborn reality: agent AI reliability decreases when tasks involve multiple steps, screens, and tools. For example, errors in SIS lookups or gradebook updates can lead to misrecorded student data. In WebArena, which replicates real sites for forums, e-commerce, code, and content management, the best initial GPT-4 agent achieved a success rate of only 14.41%. In contrast, humans performed above 78%. The failure points may seem mundane, but they are significant—misreading a label, selecting the wrong element, or losing context between steps. For schools, these errors matter. Tasks like gradebook updates, SIS lookups, purchase orders, and accommodation settings require long-term planning. The key measure is not eloquence; it is the reliability of agent AI.
Results on phones tell a similar story, with considerable variation. In November 2025, AI Multiple conducted a field test evaluating four mobile agents across 65 Android tasks in a standardized emulator. The top agent, “DroidRun,” completed 43% of tasks. The mid-tier agent completed 29%, while the weakest managed only 7%. They also showed notable differences in cost and response time. Method note: tasks included calendar edits, contact creation, photo management, recording, and file operations. The test used a Pixel-class emulator with a shared task list under the AndroidWorld framework. While not perfect proxies for schools, these tests reflect reality more than staged demonstrations do. The lesson is clear: reliability improves when tasks are narrow, the user interface is stable, and the required actions are limited. This is the design space that schools can control.
Consumer rollouts illustrate this caution. ByteDance’s Doubao voice agent is launching first on a single device in limited quantities, despite the widespread adoption of the underlying chatbot. Scientific American reports that the system can book tickets, open tabs, and interact at the operating system level, but it remains in beta. Reuters notes its debut on ZTE’s Nubia M153, with plans for future expansion. This approach is not a failure; it is a thoughtful way to build trust. Education should be even stricter, as errors can significantly affect students and staff.
Agent AI reliability shines in internal processes where clear guidelines are in place. This is where education should begin. Business evidence supports this idea. Harvard Business Review’s November 2025 guidance is clear: agents are not suited to unsupervised, consumer-facing work, but they can perform well in internal workflows. Schools have many such tasks. A library acquisitions agent prepares shopping lists based on usage data. A dispatcher agent suggests bus route changes after schedule updates. A facilities agent reorders supplies within a set budget. Each task is reversible, well-defined, and trackable. The agent acts as a junior assistant, not a primary advisor to a student. This distinction is why reliability can improve faster in internal settings than in more public contexts.
Device architecture also plays a significant role. Many school tasks involve sensitive information. Apple’s Private Cloud Compute showcases one effective model: perform as much as possible on the device, then offload more complex reasoning to a secure cloud with strict protections and avenues for public scrutiny. In education, this approach allows tasks such as captioning, live translation, and note organization to run locally. In contrast, more complex planning can run in a secure cloud. The goal is not to endorse one vendor; it is to minimize risk through thoughtful design. By combining limited tools, private execution when possible, and secure logs, agent AI reliability can improve without increasing vulnerability or risking student data.
For student-facing jobs, start with low-risk tasks. Reading support, language aids, captioning, and focused prompts are reasonable initial steps. These tasks provide clear off-ramps and visible indicators when the agent is uncertain. Keep grading, counseling, and financial-aid navigation under human supervision for now. The smartphone's widespread availability—95% of teens have access—is hard to overlook, but access does not equal reliability. A careful rollout that prioritizes staff workflows can still free up time for adults and promote a culture of measuring success before opening up tools for students.
Reliability should be a monitored characteristic, not just a marketing claim. Schools need clear benchmarks, such as success rates above 80%, to evaluate AI systems effectively. The policy landscape is moving in that direction. The EU AI Act came into effect on August 1, 2024, with rules being phased in through 2025 and beyond. Systems that affect access to services or evaluations are high-risk and must demonstrate risk management and monitoring. Additionally, NIST’s AI Risk Management Framework (2023) and its 2024 Generative AI Profile offer a practical four-step process—govern, map, measure, manage—that schools can apply to agents. These frameworks do not impede progress; they require transparent processes and evidence. This aligns perfectly with the reliability standards we need.
Guidance from the education sector supports this approach. UNESCO’s 2025 guidance on generative AI in education emphasizes targeted deployment, teacher involvement, and transparency. The OECD’s 2025 policy survey shows countries establishing procurement guidelines, training, and evaluation for digital tools. Education leaders can adopt two practical strategies from these recommendations. First, include reliability metrics in requests for proposals and pilot programs: target task success rates, average handling times, human hand-off rates, and error rates per 1,000 actions. Second, publish quarterly reports so staff and families can track whether agents meet the standards. Reliability improves when everyone knows what “good” means and where issues arise.
We also need to safeguard limited attention. Even “good” agents can overwhelm students with notifications. Set limits on alerts for student devices and establish default time windows. If financial policies permit, direct higher-risk agent functions to staff computers on managed networks rather than student smartphones. These decisions are not anti-innovation. They help make agent AI reliability a system-wide property rather than just a model characteristic. The benefits are considerable: fewer mistakes, better monitoring, and increased public trust over time.
Start with a three-month pilot focused on internal workflows where mistakes are less costly and easier to correct. Good candidates are library purchases, routine HR ticket sorting, and cafeteria stock management. Connect the agent only to the necessary APIs. Equip it with a planner, executor, and short-term memory with strict rules for data retention. Each action should create a human-readable record. From the beginning, measure four key metrics: end-to-end task success rate, average handling time, human hand-off rate, and incident rate per 1,000 actions. At the midpoint and end of the pilot, review the logs. If the agent outperforms the human baseline with equal or fewer issues, broaden its use; if not, retrain or discontinue it. This approach reflects the inside-first logic that business analysts now recommend for agent projects. It respects the strengths and limitations of the technology.
Next, prepare for the reality of smartphones without overcommitting. Devices are already in students’ hands, and integrations like Doubao show how quickly embedded agents can operate when vendors manage the entire stack. However, staged rollouts—including in consumer markets—exist for a reason: to protect users while reliability continues to improve. Schools should adopt the same cautious approach. Keep student-facing agents in “assist only” mode, with clear pathways to human intervention. Route sensitive actions—those that affect grades, attendance, or funding—to staff desktops or managed laptops first. As reliability data grows, gradually expand the toolset for students.
Finally, ensure transparency in every vendor contract. Vendors should provide benchmark performance data on publicly available platforms (e.g., WebArena for web tasks and AndroidWorld-based tests for mobile), the model types used, and any security measures intended to protect legitimate operations. Request independent red-team evaluations and a plan to quickly address performance regressions. Connect payment milestones to observed reliability in your specific environment, not to promises made in demonstrations. In 12 months, the reliability landscape will look different; the right contracts will allow schools to upgrade without starting over. The goal is straightforward: by the time agents interact with students, we need to know they can perform reliably.
We began by highlighting a stark contrast: nearly every teen has a smartphone, yet agent AI reliability in multi-step tasks still falls short. Web agents struggle with long-term tasks; mobile agents perform better but still exhibit inconsistencies; and consumer launches proceed in stages. This reality should not stifle innovation in schools; instead, it should guide it. Focus on where the tools are practical: internal workflows that are reversible and come with clear guidelines. Measure what’s essential: task success, time spent, hand-offs, and incidents. Use existing frameworks. Opt for device patterns that prioritize privacy and focus. Publish the results. Only when data prove that agents perform reliably should we allow students to use them. The future of AI in education isn’t about having a new assistant in every pocket. It’s about having reliable tools that give adults more time and provide students with efficient services. Because we have demonstrated, transparently, that agent AI reliability meets the necessary standards.
The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.
Apple Security Engineering and Architecture. (2024, June 10). Private Cloud Compute: A new frontier for AI privacy in the cloud.
Apple Security Engineering and Architecture. (2024, October 24). Security research on Private Cloud Compute.
European Commission. (2024, August 1). AI Act enters into force.
Harvard Business Review. (2025, November 25). AI Agents Aren’t Ready for Consumer-Facing Work—But They Can Excel at Internal Processes.
OECD. (2025, Mar). Policies for the digital transformation of school education: Results of the Policy Survey on School Education in the Digital Age.
Pew Research Center. (2025, Jul 10). Teens and Internet, Device Access Fact Sheet.
Reuters. (2025, December 1). ByteDance rolls out AI voice assistant for Chinese smartphones.
Scientific American. (2025, December 1). ByteDance launches Doubao real-time AI voice assistant for phones.
UNESCO. (2025, April 14). Guidance for generative AI in education and research.
Zhou, S., Xu, F. F., Zhu, H., et al. (2023). WebArena: A Realistic Web Environment for Building Autonomous Agents. arXiv.
AI Multiple (Dilmegani, C.). (2025, November 6). We tested mobile AI agents across 65 real-world tasks.
