Source Ledger

Data collected about children today can be repurposed later, but a particular future use requires evidence. The song correctly asks the question rather than claiming a specific undisclosed training pipeline.

Keep this track source-light. Its evidentiary role is to establish the problem statement: consent across time, unforeseeable secondary use, and the asymmetry between data subjects and future system builders.

The government alleged that Amazon retained children’s recordings indefinitely unless a parent requested deletion and failed in some cases to delete transcripts from all databases.

The FTC stated that retained children’s voice recordings were used to improve Alexa’s speech recognition and processing capabilities.

TightenLines naming specific internal storage formats, “training corpora,” gradient descent, embeddings, checkpoints, and model residue go beyond what the FTC order itself establishes. They are technically plausible descriptions of ML systems, but should not be presented as proven facts about Amazon’s exact pipeline without engineering documentation.

Human Rights Watch analyzed 164 government-endorsed EdTech products and later published technical profiles for 163 of them, finding widespread tracking risks across the sector.

ClassDojo’s own current policy says retention and deletion of Student Data and education records are directed by the school. Its historical policy also states that school-awarded Feedback Points are automatically deleted after 12 months.

Important“Google Analytics calls,” “Facebook SDKs,” and “derived statistics trained downstream” require a ClassDojo-specific technical source. HRW’s sector-wide evidence does not, by itself, prove those exact flows for this product.

ClassDojo publicly states that it does not sell or rent student personal information and describes COPPA/FERPA and GDPR-aligned safeguards. A fair ledger should show those claims alongside criticism.

Google Workspace for Education identifies Classroom and Chrome Sync as Core Services. Google states that no ads are shown in Core Services and personal information collected there is not used for advertising.

InferenceThe song’s vertical-integration question is appropriately framed as scrutiny of governance boundaries, not proof that education data is being fed into Google Ads.

Historical claims about scanning student email for advertising or profiling should be tied to the policy and litigation period in which they occurred. They should not be used to describe current Workspace for Education practices without date labels.

The FTC settlement with Musical.ly, TikTok’s predecessor, alleged COPPA violations and imposed a then-record $5.7 million civil penalty.

AllegationThe DOJ complaint alleges that TikTok knowingly allowed children under 13 to create regular accounts, collected data without parental consent, and frequently failed to honor deletion requests.

The European Commission issued preliminary findings that TikTok’s infinite scroll, autoplay, push notifications, and highly personalized recommender system raise DSA concerns related to addictive design. Those findings remain preliminary.

Specific claims about “reinforcement learning cycles,” “facial geometry,” or an internal “vulnerable” cluster require product-specific technical evidence before being labeled fact.

Teams supports voice and face enrollment, speaker attribution, voice isolation, and recognition features. Admins can manage whether enrollment capabilities are available.

Microsoft states that users must opt in to create voice or face profiles and may delete those profiles. Current documentation says the profiles are not used to train models or for other purposes beyond the Teams recognition feature.

Source neededThe lyrics referring to an NSW department being unaware of biometric capture, a “silent push,” and unclear default state require the specific reporting or government record that established those facts at the time of release.

AllegationA bipartisan coalition of 33 attorneys general filed the original federal complaint in 2023, alleging Meta designed features that cultivated compulsive use among youth and violated COPPA.

Public reporting and the unredacted complaint cite Meta’s own research on teen well-being and body image. When used in an article, quote the underlying exhibit or complaint rather than collapsing every internal finding into a universal causal claim.

CurrentBy August 2026, the federal multidistrict case had moved to jury selection for an initial four-state trial. Allegations remain allegations unless and until adjudicated.

Meta points to Teen Accounts, restrictive content settings, nighttime nudges, parental controls, and age-assurance systems as evidence of ongoing safety work. The ledger includes those measures because structural criticism is stronger when the company’s response is represented accurately.

Facebook/Meta content moderators brought claims concerning psychological injuries from repeated exposure to graphic content. A major U.S. settlement provided compensation and workplace measures.

TIME reported that OpenAI contracted with Sama in Kenya for workers to label toxic text used in safety-related model work, with workers describing low pay and traumatic exposure.

Recent work involving African moderators and researchers, including Adio-Adet Dinika, documents structural labor conditions and mental-health burdens beyond mere exposure to harmful content.

The lyric’s vendor roll-call should not imply that every named contractor performed every described task for the same client or model. Keep vendor, geography, task, and client relationships separately sourced.

OpenAI states that ChatGPT Edu data is not used to train its models by default.

OpenAI states that workspace admins can control retention. Deleted Enterprise/Edu conversations are removed from OpenAI systems within 30 days unless legal retention is required.

The Compliance Logs Platform is available for Enterprise and Edu and retains its own log data for 30 days unless customers export it for longer retention under their own policies.

Tighten“Admins with compliance panels” can overstate what every school administrator automatically sees. Visibility depends on workspace configuration, permissions, logging tools, and institutional policy. Phrase this as capability and governance rather than universal admin access.

The 2025 New York Times preservation order did not apply to ChatGPT Enterprise or Edu. OpenAI later announced the consumer/API preservation obligation had ended. That historical episode is useful as a retention-governance example, but it should be dated.

Latanya Sweeney’s demographic-uniqueness work found that 87% of the U.S. population in the studied data was likely unique on five-digit ZIP code, sex, and full date of birth.

Narayanan and Shmatikov demonstrated robust de-anonymization of the Netflix Prize dataset using sparse ratings plus outside information.

de Montjoye and colleagues found that four spatiotemporal points were enough to uniquely characterize 95% of individuals in a 1.5-million-person mobility dataset under the study’s resolution.

“Anonymized isn’t anonymous” is rhetorically effective but mathematically absolute. Some de-identification methods can provide meaningful privacy guarantees. The precise claim is that de-identification is method- and context-dependent and that pseudonymization alone is not equivalent to anonymity.

Major platforms describe data use in terms such as providing, maintaining, developing, improving, personalizing, securing, and updating services.

OpenAI’s consumer and business products have materially different default training rules. Enterprise, Business, Edu, Healthcare, Teachers, and API business data are not used for training by default; consumer ChatGPT provides model-improvement controls.

Avoid equivalenceThe lyric “same clause in different brands” works as criticism of policy breadth, but the source ledger should not imply that Amazon, Google, Meta, and OpenAI use data in the same technical way merely because privacy notices use similar verbs.

OpenAI, for example, states that deleted chats are scheduled for permanent deletion within 30 days, subject to exceptions including legal or security obligations. It also states that some internal backups may persist for a limited additional period in specific product contexts.

InferenceThe distinction between logical deletion, active-access removal, backup expiry, and derived model state is technically meaningful, but the lifecycle differs by system.

A deletion request does not universally mean “inaccessible but not zeroed.” Some systems may physically erase quickly; others may rely on cryptographic deletion, tombstones, backup expiry, or legal retention. Keep the lyric as systems criticism rather than a universal implementation claim.

The Amazon Alexa matter ended with a $25 million civil penalty and injunctive obligations. The order imposed enforceable requirements without turning every allegation into an admitted fact.

The FTC’s 2019 Facebook privacy settlement imposed a $5 billion penalty and sweeping compliance requirements. Large penalties can coexist with settlement language that avoids an admission of liability.

Editorial questionWhether a particular penalty is large enough to deter a company is an economic and policy question. Comparing penalties with revenue is relevant context, not proof that violations were economically rational.

Many ML systems are trained, tuned, evaluated, or personalized using human-generated examples, interactions, labels, feedback, or observed outcomes.

Qualification“Every new feature was learned from you” is poetic, not literal. Some features are hand-engineered, trained on synthetic data, built from licensed datasets, or derived without individual-user behavioral data.

Deleting a source record does not automatically reverse every derivative artifact already produced from it. Whether retraining, unlearning, or downstream deletion is required depends on law, contract, technical design, and the derivative system.

AI-enabled medical devices are trained and evaluated on data, and the FDA explicitly treats data quality, transparency, performance, and lifecycle management as central regulatory concerns.

Source neededClaims that hospitals “buy back” models trained on their own patients, that clinician labeling is unpaid, or that particular patient datasets become proprietary assets should be paired with named transactions, contracts, or studies. They are not safe as universal statements.

Medical-data research also operates under legal regimes such as HIPAA de-identification. A critique of commercialization should acknowledge those frameworks even when arguing they are insufficient for benefit-sharing.

GDPR Article 4 defines a “data subject” as the identified or identifiable natural person to whom personal data relates.

Legal nuance“Once [signals] enter infrastructure, they’re no longer yours” is too categorical as a legal statement. Data rights vary by jurisdiction and context. The point works as a critique of control and bargaining power, not as a universal rule of property law.

No automatic dividend, equity stake, or governance seat generally follows from being represented in data used to build a commercial system. The song argues that this allocation of value should be reconsidered. That is Horizon Accord’s normative question, not a claim that current law already requires profit participation.