67% against 39%, and the gap is training

TL;DR. RAND's American School District Panel found 67% of low-poverty districts provided AI training to teachers by autumn 2024, against 39% of high-poverty districts. A related survey found 61% of primary teachers in schools with mostly nonwhite students had received no AI training at all, against about 35% in schools with mostly white students. Adoption itself is not the gap. Districts providing training doubled from 23% to 48% in a year, and 86% of educational organisations use generative AI, the highest rate of any industry. The gap is in the thing that determines whether it works. The tutoring evidence shows designed and supervised systems producing large gains while unrestricted access improves assisted performance and reduces unassisted performance by 17%. Both groups of students have the tool. Only one has the version that helps.

---

Status: one strong panel source, and a surrounding literature of variable quality. RAND's American School District Panel is the load-bearing evidence and states its design. The international comparisons and market figures come from mixed sources and are attributed where used. This corpus does not take positions on contested political questions, and education funding policy is one.

---

The measured gap

RAND's American School District Panel tracks AI training and adoption across the income spectrum, and its findings from the 2024-25 school year are the clearest numbers in this subject.

By autumn 2024, 67% of low-poverty districts had provided AI training to teachers, against 39% of high-poverty districts.

A related RAND survey found approximately 61% of primary teachers in schools with mostly nonwhite students had received no AI training at all, compared with about 35% of teachers in schools with primarily white students.

And the gap is not new. RAND records it as first observed in 2023 and persisting through 2024, which means two consecutive years of the same disparity during the period when adoption roughly doubled.

Overall provision rose from 23% of districts in autumn 2023 to 48% in autumn 2024, with a further 26% planning implementation, potentially reaching 74% by autumn 2025.

So the picture is fast growth with a stable gap, which is the pattern that widens absolute differences even as both groups improve.

Why training rather than access is the right variable

The standard framing of this subject is the digital divide: devices, broadband, and who can reach the tools at all. That framing is now behind the evidence.

The tutoring literature establishes the distinction that matters. A purpose-built tutor designed on sound pedagogy produced learning gains of 0.73 to 1.3 standard deviations against active learning. A supervised system, where expert tutors revised every drafted message, matched human tutors.

Unrestricted access to a general model during practice produced assisted performance up 48% and unassisted exam performance down 17% against control.

The difference between those outcomes is not the model. It is whether somebody designed the interaction and whether somebody supervised it.

Which means the operative variable in educational AI is instructional design and teacher capability, and that is exactly what the training figures measure.

A student in a district where teachers received training has access to a structured, supervised deployment. A student in a district where 61% of teachers received none has access to the same chatbot with nobody structuring how it is used.

Both students have the tool. The evidence says those are different interventions with opposite signs on the outcome that matters.

One research paper states the reframing directly: inequity in AI education extends beyond access to tools or curricula to include differences in instructional support and integration, so the divide reflects disparities in the quality and depth of engagement rather than the mere availability of AI-related content.

The compounding with detection

This is where the two halves of this territory meet, and the interaction is worse than either finding alone.

The detection article established that AI detectors misclassify human writing in 10 to 20% of cases, and that an analysis of 10,725 assessments found flags falling disproportionately on younger students, male students, and those with lower prior educational attainment.

One secondary source puts the non-native English false positive rate above 25%.

So a student in an under-resourced school is more likely to receive the unsupervised version of the tool, which the evidence says damages unassisted performance, and more likely to be flagged by a detector when their writing looks conventional.

Two mechanisms, same population, and neither was designed with the other in mind.

Neither is a deliberate disadvantage. The training gap follows district budgets, and the detector sorts on a statistical property of text. The compounding is emergent, which is why it appears in no policy document and in no vendor comparison.

What the infrastructure picture adds

The older divide has not closed, and one policy change moved against it.

A 2026 study found rural and low-income schools facing persistent and in some cases worsening gaps in broadband access.

In May 2025 the US Senate voted to repeal FCC rules that had allowed E-Rate funds to cover off-campus Wi-Fi hotspots, a decision reported as threatening $27.5 million in hotspot funding already requested by more than 20,000 schools and libraries.

This corpus takes no position on that vote, which is a contested political question. What can be stated is the sequence: connectivity gaps persist, a funding mechanism narrowed, and the training gap sits on top of both.

And internationally the same shape appears. 24.7% of the working-age population in the Global North uses generative AI tools against 14.1% in the Global South. UNESCO's 2025 survey found around 70% of higher education institutions in Europe and North America having or developing AI guidance, against 45% in Latin America and the Caribbean.

A governance gap of that size is the institutional version of the training gap: not whether the tools are present, but whether anyone has decided how they should be used.

The governance figure that frames all of it

86% of educational organisations use generative AI, reported as the highest adoption rate of any industry, while most US public schools lack formal AI policies for students.

Near-universal use with near-absent policy is the condition the academic integrity article documented from the student side, where 67% use AI weekly and 8% believe it constitutes cheating.

Here it appears from the institutional side, and the two are the same fact seen from different positions.

Which suggests the policy gap is upstream of the equity gap rather than parallel to it. A district with a policy can train against it. A district without one has nothing to train toward, and writing a policy costs less than any training programme.

The two interventions, side by side

Setting the conditions against each other shows why a training figure is a proxy for an outcome.

The first row is what the digital divide framing measures. The rows below it are where the outcome lives.

And the fourth row is the finding. Those two figures come from randomised trials of what are nominally the same technology, and they have opposite signs on the measure that matters after the tool is taken away.

The word "plausibly" in the middle rows is doing real work and should not be smoothed over. RAND measured whether training was provided, not whether it changed classroom practice, and its own note that early sessions addressed fear and confusion rather than instructional application is a reason to doubt the link.

So the honest version of the table is narrower than it looks. A training figure is a proxy for a proxy: provision stands in for capability, and capability stands in for the designed-and-supervised condition the evidence supports. Two inferential steps, neither measured.

Which is worth stating because the table is the article's central claim and its weakest link sits in the middle of it.

What would settle it

Three measurements, in ascending order of difficulty, and the first is nearly free.

Report training provision alongside training content. RAND asks whether districts provided AI training. Asking what the training covered would separate fear-and-confusion sessions from instructional design work, which the outcome evidence says are different interventions. That is one survey question.

Measure classroom practice, not district provision. Whether teachers in trained districts actually structure AI use differently is checkable by observation or by asking teachers what they changed. Nobody has, and provision figures will continue standing in for practice until somebody does.

And run the comparison the tutoring evidence implies. Two cohorts in comparable schools, one with structured supervised AI use and one with unrestricted access, measured on unassisted post-tests. That is the study that would establish whether the training gap causes an outcome gap, and it is the expensive one.

The first two are the ones worth pressing for. They cost a survey redesign and a set of classroom observations, and without them this entire subject rests on the assumption that recorded training changes what happens in a room.

Three things this establishes

The measured gap is training, not access. 67% against 39% of districts, and 61% of teachers in majority-nonwhite primary schools with no training at all, while adoption doubled across the board. Both groups have the tools.

Training is the variable the outcome evidence turns on. Designed and supervised deployments produce large gains; unrestricted access produces assisted gains and a 17% unassisted loss. The training figures are therefore a proxy for which of those two interventions a student receives.

And it compounds with detection. The same population more likely to get the unsupervised version is more likely to be flagged by a detector, through two independent mechanisms that no policy document connects.

What it does not establish

That under-resourced students are worse off with AI than without it. No study compares those conditions, and the counterfactual for many is a class size and a teacher workload that AI may relieve.

That training closes the gap. The training figures measure provision, not quality, and no study links district training programmes to student outcomes.

That the international comparisons are precise. The Global North and South figures and the UNESCO governance percentages come from surveys with varying methods.

And nothing about education funding policy. That is a contested political question and this corpus does not take positions on those.

What is unresolved

Whether trained districts produce better student outcomes. The chain from teacher training to structured deployment to learning gain is plausible at every link and measured at none.

What the training consists of. RAND notes initial trainings focused primarily on addressing fear and confusion about AI rather than instructional application, which is a different intervention from the one the tutoring evidence supports.

Whether the gap is closing or widening. Provision rose in both groups, so the ratio may improve while the absolute difference grows, and no source reports it either way.

And what happens at 74% adoption. If provision reaches that level by autumn 2025 as projected, the remaining quarter is likely to be the districts already furthest behind.

The counter-argument

Training provision is a weak proxy for what students actually experience. A district can record a professional development session and change nothing in a classroom, and RAND's own note that early trainings addressed fear rather than instruction suggests exactly that. This article treats a provision figure as though it measured deployment quality.

The tutoring evidence may not transfer. The gains came from a purpose-built tutor in a university physics course and a supervised system in five UK schools, neither of which resembles a trained teacher in an under-resourced US district, so using them to interpret the training gap imports findings from conditions that do not apply.

The compounding argument is constructed. The training gap and the detector bias come from separate literatures, no study follows a student through both, and this article's central interaction is an inference rather than an observation, which is the same weakness it flagged in the grading article.

And the equity framing may understate the counterfactual. If AI relieves teacher workload, the districts with the highest workloads have the most to gain, so a training gap could coexist with a benefit that is larger where provision is lower, and nothing here measures that.

The short version

RAND's American School District Panel found 67% of low-poverty districts providing teacher AI training by autumn 2024 against 39% of high-poverty districts, and about 61% of primary teachers in schools with mostly nonwhite students having received none at all against 35% in mostly white schools. The gap was first observed in 2023 and persisted while overall provision doubled from 23% to 48%.

Adoption is not the divide. 86% of educational organisations use generative AI, the highest rate of any industry, while most US public schools lack formal AI policies.

The divide is training, and training is the variable the outcome evidence turns on. Designed and supervised systems produced gains of 0.73 to 1.3 standard deviations or matched human tutors. Unrestricted access produced assisted performance up 48% and unassisted performance down 17%. Both groups of students have the tool; only one has the version that helps.

And it compounds. Detectors misclassify human writing in 10 to 20% of cases with flags falling disproportionately on students with lower prior educational attainment, and one source puts non-native English false positives above 25%. The same population more likely to receive the unsupervised version is more likely to be accused of using it.

Neither mechanism was designed against that population. The training gap follows budgets and the detector sorts on text statistics. The compounding is emergent, which is why no policy document mentions it.

Common questions

What is the measured gap in AI education? Training rather than access. RAND's American School District Panel found that by autumn 2024, 67% of low-poverty districts had provided AI training to teachers against 39% of high-poverty districts, and a related survey found approximately 61% of primary teachers in schools with mostly nonwhite students had received no AI training at all, against about 35% in schools with primarily white students. RAND records the disparity as first observed in 2023 and persisting through 2024.

Is adoption itself unequal? Less than the training figures suggest. Districts providing AI training doubled from 23% in autumn 2023 to 48% in autumn 2024, with a further 26% planning implementation, and 86% of educational organisations report using generative AI, the highest adoption rate of any industry. The tools are widely present. What differs is whether anyone has been trained to structure their use.

Why does training matter more than access? Because it is the variable the outcome evidence turns on. A purpose-built tutor designed on sound pedagogy produced learning gains of 0.73 to 1.3 standard deviations against active learning, and a supervised system where expert tutors revised every message matched human tutors. Unrestricted access to a general model during practice produced assisted performance up 48% and unassisted exam performance down 17% against control. The difference between those outcomes is not the model but whether somebody designed and supervised the interaction.

How does this compound with AI detection? Through two independent mechanisms landing on one population. Detectors misclassify human writing in 10 to 20% of cases, and an analysis of 10,725 assessments found flags falling disproportionately on younger students, male students and those with lower prior educational attainment, with one secondary source putting non-native English false positives above 25%. A student in an under-resourced school is therefore more likely to receive the unsupervised version of the tool and more likely to be flagged when their writing looks conventional. Neither mechanism was designed with the other in mind.

What about infrastructure? The older divide has not closed. A 2026 study found rural and low-income schools facing persistent and in some cases worsening gaps in broadband access, and in May 2025 the US Senate voted to repeal FCC rules that had allowed E-Rate funds to cover off-campus Wi-Fi hotspots, a decision reported as threatening $27.5 million already requested by more than 20,000 schools and libraries. This corpus takes no position on that vote, which is a contested political question.

Does the pattern appear internationally? Yes, in the same shape. 24.7% of the working-age population in the Global North uses generative AI tools against 14.1% in the Global South, and UNESCO's 2025 survey found around 70% of higher education institutions in Europe and North America having or developing AI guidance against 45% in Latin America and the Caribbean. A governance gap of that size is the institutional version of the training gap: not whether tools are present, but whether anyone has decided how they should be used.

What would close it? Unknown, and the honest answer is that nobody has measured whether training works. The chain from teacher training to structured deployment to student learning gain is plausible at every link and measured at none. RAND also notes that early trainings focused primarily on addressing fear and confusion about AI rather than instructional application, which is a different intervention from the one the tutoring evidence supports. Writing a policy costs less than any training programme and most US public schools do not have one.

What is the strongest objection to this article? That training provision is a weak proxy for what students experience. A district can record a professional development session and change nothing in a classroom, and RAND's own observation that early trainings addressed fear rather than instruction suggests exactly that. A second objection is that the compounding argument is constructed: the training gap and the detector bias come from separate literatures, no study follows a student through both, and the central interaction is an inference rather than an observation.