"We'll Fix It Later": Education, AI, and the Deferral of Privacy in EdTech

1 INTRODUCTION

• EdTech collects identifiers, behavioral logs, grades, and sometimes disability or IEP records, while children may lack meaningful choice; uneven laws and breaches make governance high-stakes.Examples include Unacademy’s exposure of more than 20 million profiles, Illuminate Education’s exposure of records affecting more than 800,000 New York City students, and the 2024 PowerSchool incident affecting tens of millions of records.Source: §1 · p. 1
• The study asks how privacy and security enter product lifecycles, how responsibility is assigned, how feedback and visible harm shape priorities, and how public policies reflect internal practices.In educational settings, institutions and, for minors, parents or guardians often act as proxy decision-makers, while students may have little ability to choose or leave the technology.Source: §1 · p. 1
• Across 12 interviews and a 48-platform policy audit, we find a cycle: privacy is recognized but postponed, delegation makes delay appear managed, weak feedback limits pressure, and AI obscures downstream data use.The study’s contribution is to show how mandatory adoption, limited student power, diffuse responsibility, and slow-moving harms weaken corrective mechanisms available in ordinary consumer markets.Source: §1 · p. 1

• EdTech personalization enables long-term sorting and tracking, while continuous data collection has become normalized in EdTech discourse.Source: §2 · p. 1
• Children’s data governance involves fragmented authority, firms face ethical challenges around AI adoption, and mandatory platforms can leave students unable to opt out,.Source: §2 · p. 2
• Schools face limited vendor accountability and uneven security capacity, while educators rely on limited guidance and procurement trust; higher-education adoption often follows convenience and momentum.EdTech platforms also turn student interactions into proprietary data assets, while public concern spans surveillance, ethics, and technical security.Source: §2 · p. 2
• Research finds broad behavioral, emotional, and social data collection without consistently meaningful consent, incomplete or difficult-to-evaluate commitments,, and COPPA limitations in modern procurement.Stakeholders may not understand downstream data practices; cloud architectures distribute responsibility, and provider incentives impede privacy-by-design investment,.Source: §2 · p. 2
• Software teams often defer privacy, and child-directed app developers may treat compliance as a threshold, despite privacy engineering requiring sustained commitment and misaligned security incentives.The study bridges this literature by examining privacy deferral, delegation, and reduced visibility inside EdTech organizations and testing for those patterns in a 48-platform policy audit; it treats AI as an amplifier of opacity and third-party data flows, not the origin of weak ownership, feedback, and incentives.Source: §2 · p. 2

3 METHODOLOGY

• The mixed-methods design combined 12 interviews with a public policy audit; participants gave informed consent, and identifying details were removed.The study received institutional ethics approval.Source: §3 · p. 2

3.1 Interviews and sample

• Twelve of approximately 100 contacted EdTech practitioners completed semi-structured Zoom interviews, with recruitment seeking conceptual variation rather than statistical representativeness.Participants came from US and Indian startups, nonprofits, research organizations, and public K-12 districts; implementation-focused roles could offer less visibility into strategic decisions.Source: §3.1 · p. 2

3.2 Qualitative analysis

• Line-by-line coding of all 12 interviews yielded more than 80 initial codes, consolidated into eight focused categories that describe recurring mechanisms, not their prevalence.The authors did not claim analytical saturation, citing NDA and compliance barriers and convergence across roles and contexts as the basis for credibility.Source: §3.2 · p. 3

3.3 Privacy policy audit

• The audit covered 48 prominent platforms and scored accessible public documentation on five dimensions, including data collection, sharing, children's consent, AI decisions, and accountability.Platforms were selected using reported usage, app-store rankings, and institutional-adoption data; a zero indicated no accessible disclosure within the audit scope, not necessarily no internal commitment.Source: §3.3 · p. 3
• 87.9% mean exact agreement, 98.7% mean agreement within one point, and mean Kappa 0.782 indicated strong overall inter-rater reliability.Total scores correlated strongly between raters (Pearson r=0.865, p<0.001), and all disagreements were by one point.Source: Table 3 · p. 3, §3.3 · p. 4

4.1 Interview Findings

• Participants identified five connected mechanisms: privacy deferral, delegated responsibility, weak feedback, less visible downstream AI data use, and uneven organizational capacity.The findings describe an explanatory pattern across a purposive sample, not estimates of how common these practices are across EdTech.Source: §4.1 · p. 4
• Market pressure and delegation stabilize privacy deferral, weak feedback loops sustain it, and accountability becomes diffuse.Figure 1 depicts privacy deferral across the product lifecycle, with market pressure and delegation as stabilizing forces.Source: §4.1 · p. 4, Figure 1 · p. 4
• In schools, nonprofits, and smaller organizations, privacy capacity was often constrained by staffing, budget, or expertise.Participants distinguished these capacity constraints from strategic deprioritization.Source: §4.1 · p. 4
• In some commercial settings, extensive behavioral tracking and personalization were tied to product value, creating incentives to collect rather than minimize data.Source: §4.1 · p. 4
• AI-related risks surfaced independently across interview roles, although AI was not the primary interview focus.Participants described inconsistent school capacity to evaluate AI tools and teachers uploading student information to general-purpose models.Source: §4.1 · p. 4
• AI governance often became reactive only after a problem emerged, positioning AI as a frontier within the broader deferral cycle rather than a separate failure category.Source: §4.1 · p. 4

Privacy Deferral and Prioritization

• Across all 12 interviews, participants recognized privacy but described addressing it after pilots, growth, or core functionality.Participants rarely characterized privacy as unimportant; the recurring pattern was postponement.Source: Privacy Deferral and Prioritization · p. 4
• Startups faced market-speed pressure, while nonprofits and schools had to justify privacy spending against program costs and expected returns.Teams weighed whether investment would produce student outcomes, growth, or future funding.Source: Privacy Deferral and Prioritization · pp. 4–5
• Participants described security as coupled with privacy, and P10 said security would become a “second base layer” after pilots began.The account illustrates how protective work was sequenced for later rather than built in from the outset.Source: Privacy Deferral and Prioritization · p. 4

Delegation and Diffuse Accountability

• Organizations shifted responsibility to cloud providers, policy documents, compliance functions, and downstream institutions.Interviewees described relying on AWS and Azure for data storage.Source: Delegation and Diffuse Accountability · p. 5
• Encryption and two-factor authentication were common, while threat modeling, provenance tracking, and systematic privacy review were often deferred until compliance pressure or an incident.Source: Delegation and Diffuse Accountability · p. 5
• Some vendors bypassed district procurement by going directly to schools, which participants said often lacked the equipment or understanding to manage data.Source: Delegation and Diffuse Accountability · p. 5
• When developers rely on cloud defaults, schools on vendor documentation, and users on institutions, privacy becomes everyone’s responsibility in principle but no one’s job in practice.Source: Delegation and Diffuse Accountability · p. 5

Weak Feedback Loops

• Participants consistently received usability, engagement, and learning-outcome feedback, but little direct feedback about privacy or security features.P6 described sandbox testing as focused mainly on usability ratings.Source: Weak Feedback Loops · p. 5
• Privacy harms are often probabilistic, delayed, and difficult to attribute, so organizations may interpret the absence of complaints as evidence that no urgent problem exists.Source: Weak Feedback Loops · p. 5
• Participants described little end-user pushback, with P2 attributing some silence to users believing their data was not valuable enough to matter.Source: Weak Feedback Loops · p. 5

AI as an Amplifier of Existing Governance Gaps

• AI-related concerns centered on opaque downstream data use and institutions’ capacity to evaluate new tools, extending rather than replacing existing organizational problems.Source: AI as an Amplifier of Existing… · p. 5
• Governance capacity varied sharply; one school had no designated AI or privacy specialist and relied on its principal and instructional coaches.P7 also questioned how AI could help students think more deeply and evaluate its outputs.Source: AI as an Amplifier of Existing… · p. 5
• Participants described reactive governance, with tools restricted only after problems surfaced, and uncertainty about teachers uploading student data to general-purpose LLMs.Source: AI as an Amplifier of Existing… · p. 5

Unequal Capacity and Commercialization Pressures

• Larger organizations could maintain dedicated privacy or security staff, whereas smaller organizations assigned those responsibilities alongside other roles.Capacity differences included staff, budget, and expertise to act on privacy concerns.Source: Unequal Capacity and Commercialization… · p. 5
• Privacy deferral reflected both genuine scarcity and strategic deprioritization, including organizational shortcuts and the costs of foregoing data collection or personalization.P2 described a cost tension for new organizations, while P8 described shortcuts in security.Source: Unequal Capacity and Commercialization… · p. 5
• In some commercial settings, data collection was part of the value proposition, with personalization, aggressive marketing data use, IP tracking, and passive behavioral collection working against minimization.These pressures could preserve data-intensive practices rather than treating collection as merely an operational byproduct.Source: Unequal Capacity and Commercialization… · p. 5
• Procedural gatekeeping and regulatory ambiguity shaped when privacy pressures became visible and who was expected to respond.Source: Unequal Capacity and Commercialization… · p. 5

From Themes to the Deferral Cycle

• Privacy deferral reproduces itself across the product lifecycle because underlying incentive, ownership, and feedback structures remain unchanged.Figure 1 traces the cycle from strategy and funding through build, pilots, procurement, AI, and incidents.Source: From Themes to the Deferral Cycle · pp. 5–6
• At strategy and funding, privacy struggles to enter the business case; during build, cloud delegation can substitute for internal engineering, while pilots surface usability rather than privacy concerns.These stages shift attention or responsibility away from internal privacy work.Source: From Themes to the Deferral Cycle · p. 5
• Procurement can shift responsibility to schools with uneven capacity, AI introduces less visible downstream uses, and incidents finally prompt reactive attention.The cycle spans procurement, downstream AI use, and incident response.Source: From Themes to the Deferral Cycle · pp. 5–6

4.2 Policy Audit Findings

• 33% (16/48) made no meaningful AI disclosure despite visible AI features, and 73% received the middle accountability score, typically reflecting generic or vague language.Data collection disclosure (C1) was strongest, while downstream governance was weaker.Source: §4.2 · p. 6, Table 4 · p. 6
• US K-12 platforms scored highest overall (M=7.47), followed by US higher education/adult learning (M=5.50) and India K-12/exam preparation (M=4.77), with significant differences (Kruskal-Wallis H=20.634, p<0.001).Source: §4.2 · p. 6
• K-12 platforms scored significantly higher on children’s consent, but showed no comparable advantage on AI disclosure or accountability, where no equivalent mandate exists.The pattern suggests regulation improves dimensions it specifically makes enforceable rather than producing generalized privacy maturity.Source: §4.2 · p. 6

5 DISCUSSION AND IMPLICATIONS

• Privacy deferral becomes self-reinforcing as visible outcomes take priority, responsibility is delegated, weak feedback obscures harms, and silence justifies further delay.The cycle is locally rational for organizations because privacy failure costs are delayed and diffuse, while students often bear them. Policy disclosure can satisfy visible expectations without mechanisms to govern downstream data use.Source: §5 · p. 6, §5.1 · p. 6

5.1 Why the Deferral Cycle Persists

• Privacy is often postponed for locally rational reasons, while collection disclosures are easier to provide than meaningful AI disclosure or operational accountability.Organizations may rely on cloud defaults, policy statements, and complaints rather than building privacy engineering or auditing data flows; meaningful governance requires documenting model and data flows, ownership, timelines, and standards.Source: §5.1 · p. 6

5.2 Why EdTech is Different

• K-12 platforms perform better on children’s consent where rules apply, but that advantage does not extend to AI disclosure or accountability.Students often cannot choose their platforms, families may not know which services are used, and quiet, difficult-to-trace harms make exit and market feedback unreliable correctives.Source: §5.2 · pp. 6–7

5.3 AI as an Amplifier

• One-third of audited platforms made no meaningful AI disclosure despite visible features, leaving users less able to understand processing, decisions, or where to direct concerns.AI-related data transfers and recommendations may be difficult for parents to recognize, further weakening already limited feedback loops.Source: §5.3 · p. 7

5.4 Practical Implications

• Developers should make privacy a release criterion, using feature-owner data-flow summaries and just-in-time notices for AI-related transfers or model training.The proposed summary covers collected data, third parties, retention or training, automated decisions, and incident-response ownership; clear internal ownership can prevent responsibility from dissolving across teams,.Source: §5.4 · p. 7
• Schools and policymakers should require plain-language operational evidence from vendors and use the five audit dimensions as a comparison rubric.Requested evidence should cover data fields, subprocessors, consent, AI use, retention or training, and breach-notification timelines; specific enforceable rules can improve the dimensions they govern.Source: §5.4 · p. 7
• Educators need structured support to recognize data transfers and AI disclosure and to identify institutional approval channels, rather than bearing privacy responsibility alone.Privacy literacy can strengthen feedback, but cannot substitute for vendor accountability or compensate for opaque practices.Source: §5.4 · p. 7
• Researchers should track distinct privacy-governance dimensions over time and test whether procurement or AI-specific rules produce effects beyond the dimensions they directly regulate.Future studies should examine India’s DPDPA enforcement and include commercial platform decision-makers and regulators, who were difficult to access in this sample.Source: §5.4 · p. 7

6 LIMITATIONS

• The 12 interviews were primarily US-based and constrained by NDAs and compliance restrictions, with large commercial platforms and regulators absent.Some participants were closer to implementation than strategic decision-making, limiting visibility while revealing how governance is experienced operationally.Source: §6 · p. 7
• The policy audit captures publicly accessible documentation, not internal behavior, and its India segment is smaller than the market’s importance.Because DPDPA enforcement was still developing, US-India score differences reflect patterns in this purposively assembled sample, not country-level privacy maturity; unequal segment sizes, platform mix, and regulatory stages limit direct comparison.Source: §6 · p. 7

• Across 12 interviews and 48 policy audits, organizations recognized privacy risks but repeatedly postponed action, delegated responsibility, and received too little feedback to make those risks urgent.Source: §7 · pp. 7–8
• Policies usually described data collection, while providing much weaker information about AI governance and meaningful accountability.Source: §7 · p. 8
• Because students often cannot choose or leave data-collecting systems, corrective pressure from consumer markets is limited, making institutional procurement standards, clearer internal ownership, and enforceable oversight necessary.These measures would make privacy a condition of deployment rather than a promise to revisit later.Source: §7 · p. 8