The Capability Gap | AI Wars of 2026
Access, fluency, judgment, and transfer. Source-backed analysis from AI Wars of 2026.
Part IV — Preparing the AI-Enabled Workforce
The Capability Gap
In research
This chapter distinguishes access, fluency, judgment, and transfer. Evidence about one does not establish the others. The chapter remains in research because transfer into work is still missing from its examples. The one judgment study found is outside education-technology deployment: it tests only how a disclosed label affects grading, not general error detection.
Every part of this book so far has asked what the technology can do and who controls it. This part asks a question the rest of the book quietly assumes: can people direct delegated work well enough for any of it to produce value?
Announcements about AI in education almost always report one thing and imply another. A license count is reported; capability is implied. The two are separated here into four layers, and the distance between them is the subject of this part:
- Access. The tools are provisioned and reachable. Measured in licenses, seats, or enrolment.
- Fluency. A person can operate the tool competently. Measured in assessment and task completion.
- Judgment. A person can tell a good output from a merely plausible one, and knows when not to delegate at all. Measured, when it is measured, in error detection and calibration.
- Transfer. The capability survives into unsupervised real work. Measured longitudinally, or by employers, or not at all.
Delegated work depends on human judgment. Recognizing a confident wrong answer can prevent a mistake from becoming an action. Review still takes time, and detecting an error does not necessarily reverse harm already done.
The public record is concentrated almost entirely in the first layer, with some in the second. That is the gap this chapter has to describe precisely before the rest of the part can say anything useful.
What the available measures show
Bitkom's 2026 telephone survey of 603 German companies with at least 20 employees reports employers' assessments of staff AI skills separately from whether they provide training. These are reports about perceived skills and training provision, not direct tests of employees' ability to detect errors or decide when to delegate. Company-level answers also do not show which trained individuals became more capable. Source 1
Microsoft's Philippine Department of Education account distinguishes a teacher-access commitment from a reading pilot with reported student outcomes. The pilot is vendor-published and has no control group. Student reading progress is not a test of teachers' judgment about AI, and promised teacher access is not observed use. The populations, activities, and outcomes cannot be substituted for one another. Source 2
Google's U.S. Future Report describes a national survey of 1,003 US teenagers aged 13-17, conducted from April 23 to May 26, 2026, alongside separate state-level studies and focus groups. Its national findings describe what teenagers report about technology use, information evaluation and preparation. They are not observations of how well those respondents detect an incorrect AI answer. Source 3
The report also distinguishes students who say their school offers digital-literacy or technology classes from those who say the classes are unavailable or are unaware of them. The latter group combines reported absence with lack of awareness; it is not a direct count of schools without provision. This vendor-commissioned study adds evidence about reported preparation, not proof that classes produce better judgment or that capability transfers into work. Source 3
Outside these institutional accounts, one peer-reviewed study measures judgment. A preregistered randomized experiment involved more than 1,300 teachers in Greece. Each teacher reviewed student work with an intentionally inaccurate grade that was labeled as either AI-generated or human-generated. Under a harsh recommendation, the grading fairness gap — the distance between the teacher's revised grade and the objective rating — was 22% larger when the grade was labeled AI-generated. Under a lenient recommendation, correction rates did not differ by label. Source 4
The study directly tests how experts correct an error when an AI label is disclosed. It measures one labeling effect in one grading task, not a general ability to detect unlabeled AI errors. It also does not establish whether training or experience changes the effect.
These examples make the gap more specific. Training can be offered without a capability assessment. An educational outcome can be measured without testing the capability needed to direct agents. Reported preparation can differ from demonstrated judgment. A disclosed AI label can also increase the error that remains after an expert reviews a harsh recommendation. None establishes that education produces reliable delegated work later in a career.
What remains unanswered
Evidence of judgment must test whether people detect errors and choose when to delegate. Evidence of transfer must follow capability into unsupervised work. This chapter includes one judgment study, but it measures how a disclosed label affects grading-error size in one task. It does not test unlabeled error detection in general. Evidence of transfer remains absent, and the four layers should not be combined into a score.
This chapter builds on Chapter 00, which frames voluntary judgment, perceived value, trust, and agency as the conditions connecting availability to repeated use. It asks where those conditions are formed.
Sources
Source references
Source 1: Erstmals nutzt die Mehrheit der Unternehmen KI
Bitkom e. V.
Published: 2026-09-14
Bitkom reports that 57% of surveyed German companies now use AI, against 36% a year earlier and 20% two years earlier, and that 4% say AI is not a topic for them. Its methodology note states that Bitkom Research telephoned 603 companies with 20 or more employees in Germany, in calendar weeks 28 to 33 of 2026, and describes the survey as representative of the economy as a whole. The same instrument reports that no company using AI says it fully exploits the potential, 3% say they exploit it strongly, 28% little, and 59% 'ueberhaupt nicht' - not at all. The measure is self-assessed at company level and 'AI use' is not defined on the page, so it establishes breadth of use rather than depth, outcome, retention, or value. The fieldwork period precedes publication by roughly six to eight weeks and must not be read as a measurement of the publication date.
Source 2: DepEd and Microsoft partner to advance Philippine education in the AI era, equip 1M teachers with Microsoft Copilot
Microsoft
Published: 2026-08-27
Microsoft states it will provide 1 million Filipino public school teachers with access to Microsoft Copilot under a renewed Department of Education partnership. It reports that a nationwide AI skilling program drew nearly 75,000 live attendees against more than 175,000 registrations, and that a 21-week Reading Progress and Reading Coach pilot reached 3,431 students across eight regions, analyzed 8,381 reading submissions, trained over 2,100 teachers, and recorded 91.5 percent teacher satisfaction, with 780 re-assessed students showing reading progress, 319 moving between Phil-IRI bands and 461 improving within a band. This is a vendor-published account of its own partnership. Access provisioned is not a paid seat, an activated seat, or active use; the pilot is small relative to the 1 million figure, reports no control group or comparison cohort, and its satisfaction measure is self-reported.
Source 3: U.S. Future Report 2026: Growing Up with Technology & Learning How to Shape Its Role
Published: 2026-09-16
Google's U.S. Future Report 2026 describes a national survey of 1,003 US teenagers aged 13-17, fielded April 23-May 26, 2026, with separate state-level studies and focus groups. The preparation finding combines students reporting unavailable classes with those unaware of them. The national findings are self-reported and must not use the combined state samples as their denominator. This vendor-commissioned study does not establish demonstrated error detection, causal learning gains or transfer into work.
Source 4: Why do experts miss AI's errors? Evidence from a randomized labeling experiment
PNAS Nexus
Published: 2026-06-09
A preregistered randomized labeling experiment with more than 1,300 teachers in Greece. Education experts reviewed identical student work paired with an intentionally inaccurate score, randomly labeled human- or AI-generated, with the error direction also randomized. Under a harsh recommendation, the grading fairness gap was 22% larger when the score was labeled AI-generated; under a lenient recommendation, correction rates did not differ by label. This tests error-correction deference to a disclosed AI label in one grading task, not unlabeled error-detection capacity in general, and does not establish whether deference changes with experience, training, or over time.