Traditional educational
software delivers content and records structured inputs. A reading program
presents a passage and records whether the student answered comprehension
questions correctly. A math platform tracks which problems were attempted,
which were right, and how long each took. The software follows defined rules:
if the student scores above a threshold, advance; if below, repeat.
The data these tools collect is
predictable and bounded. You know what it is, where it goes, and what it is
used for. The tool does not change its own behavior based on what it learns
from aggregated student interactions. Its logic was set at the time it was
built.
Student privacy risks still
exist with traditional edtech. In 2023, the FTC obtained a $6 million order
against Edmodo, a traditional classroom platform, for collecting children's
data without proper parental consent and using it to target students with advertising,
which is a serious violation of COPPA.1 But the risk profile of traditional edtech is one that most districts have
frameworks to manage.
AI-powered edtech collects the
same structured inputs as traditional software — and then goes further. It also
collects behavioral signals: how a student reads through content, where they
pause, what they skip, how they phrase open-ended responses, which paths
through material they take and abandon. From these signals, the system builds
inferences: this student struggles with inferential questions, this one
disengages when content is text-heavy, this one accelerates when given choice.
Those inferences then drive the
tool's behavior. The AI adapts pacing, selects content, and in some systems
surfaces recommendations to teachers about which students may need
intervention. This is the core promise of adaptive AI edtech, and it is a genuine
improvement over one-size-fits-all delivery.
But it introduces three
questions that traditional edtech evaluation frameworks were not designed to
answer.
The first is the model training
question. The Future of Privacy Forum's guidance on vetting AI tools in schools
identifies this as the critical issue that most district reviews miss: will
student data be used to train or improve the underlying AI model?2 Some vendors explicitly prohibit this. Others are ambiguous. If a student's
written responses, behavioral patterns, or reading history are fed back into
the model, that data may persist in the system indefinitely and could surface
in outputs for other users. A traditional edtech privacy review does not ask
this question because traditional tools do not have underlying models to train.
The second is the transparency
and explainability question. When an AI system recommends that a student be
assigned to an intervention pathway, or flags a student as at risk of falling
behind, who can explain how that recommendation was made? Traditional software
follows explicit rules that a teacher or administrator can trace. AI systems
often work through statistical patterns that even the vendor may not be able to
articulate fully. The FPF guidance notes that some proposed AI use cases
involve substantive decision-making affecting students' educational
trajectories, and that these warrant particular scrutiny, in some cases
requiring human oversight or parental consent before the AI recommendation
carries weight.2
The third is the equity
question. AI systems learn from data, and the data they learn from reflects the
students and contexts in which they were trained. A tool trained predominantly
on data from well-resourced suburban classrooms may perform differently, and
potentially less accurately, for students in other contexts. The American
Association of School Administrators identifies this as one of the primary
concerns districts should raise with AI vendors: what datasets were used to
train the AI, and was the system tested for performance across different
student subpopulations?3
None of this is a reason to
default to caution or avoid AI-powered tools. Many of them offer meaningful
benefits that traditional edtech cannot: genuinely personalized pacing, early
identification of students who need support, reduced administrative work for
teachers. The point is that a traditional app vetting checklist is not
sufficient to evaluate them.
Three questions belong in any
AI edtech review that are not standard in traditional edtech review. Does
student interaction data train the underlying model, and if so, how is it
de-identified and who has access? How does the system make its recommendations,
and can it explain them to teachers and parents in plain terms and how was the
AI trained and tested? More specifically, was it evaluated for accuracy and
fairness across different student populations, including those who look like
your students?
These are not questions a
vendor should struggle to answer. If they do, that itself is useful information.
