Blog · arXiv Analysis · Last reviewed June 25, 2026

The Teen Message Becomes the Manipulation Dataset

The June 2026 arXiv paper IMPACTeen: Intentions, Manipulation, Persuasion, Annotations, and Consequences in Teen Communication Dataset, by Aleksander Szczęsny and colleagues, introduces a generated-and-human-edited dataset for studying social influence in adolescent-context communication.

The useful definition is narrow: IMPACTeen is an evaluation resource for labeling influence scenarios, not evidence that a deployed model can police real teen conversations. Its governance value is the annotation record: who saw manipulation, what technique they named, what consequence they expected, and where adult and youth-adjacent perspectives diverged.

The Teen Message Becomes Data

The paper, arXiv:2606.16910 [cs.CL], was submitted on June 15, 2026. It sits at a difficult point in AI governance: the urge to detect manipulation aimed at young people, and the risk that a detector will flatten youth communication into a set of labels.

IMPACTeen is not a scraped archive of private teen conversations. The authors describe a dataset of textual scenarios, including dialogues, in which one party tries to influence the attitudes, beliefs, or behavior of a young individual or group. The settings include interpersonal, media-based, and digital communication contexts. The scenarios were generated under controlled constraints and then edited and validated by humans for youth-context realism.

That makes the paper distinct from the site's pages on teen confidant chatbots, kidfluencer incentives, persuasion tests, and belief dynamics. IMPACTeen asks a narrower question: what would it take to label the steering itself?

Current Context

As of June 25, 2026, adolescent influence is no longer only a moderation taxonomy question. The FTC's 2025 companion-chatbot inquiry asks how companies test, monitor, monetize, disclose, and mitigate possible negative effects on children and teens. The FTC's 2025 COPPA amendments strengthen privacy rules for children under 13, including separate parental opt-in for certain third-party disclosures, data-retention limits, and expanded personal-information categories. Those sources do not certify IMPACTeen, but they explain why youth-facing influence classifiers need privacy, provenance, and audit records.

The European policy context points the same way. The Digital Services Act's minor-protection guidelines recommend safety and privacy by design for online platforms accessible to minors, including limits on manipulative commercial practices, addictive design features, recommender risks, and AI chatbots integrated into platforms. The European Commission's AI Act page lists harmful AI-based manipulation and harmful exploitation of vulnerabilities as prohibited-risk categories, and says the prohibitions became effective in February 2025. A dataset about adolescent persuasion therefore sits near a live regulatory boundary: detection can protect minors, but the same labels can also become infrastructure for targeting, ranking, or over-policing them.

UNICEF's 2025 child-centered AI guidance adds the child-rights frame: safety, privacy, transparency, accountability, non-discrimination, inclusion, well-being, and enabling environments. For this article, that means the dataset should be read as a controlled research object, not as permission to ingest real teen chats or infer vulnerability without a child-safety case.

What IMPACTeen Contains

The arXiv abstract reports 1,021 texts, 5,100 individual annotation records, and gold labels for social influence techniques. Each text was annotated from five perspectives named by the authors as teenagers, parents, psychologists, communication experts, and teachers. The annotation scheme covers influence presence, techniques, intentions, consequences, resistance, reactions, and annotator confidence.

The construction pipeline matters. The authors selected 20 social influence techniques from the SITT taxonomy, defined context-vector dimensions such as relationship type, communication environment, influence goal, technique visibility, and target resistance, then manually reviewed and validated 330 context vectors. The paper reports that final text generation used DeepSeek-V3, followed by quality control.

The dataset was created in Polish and accompanied by a corresponding English version. For AI evaluation, that is not a side note. A manipulation detector trained or tested on IMPACTeen is learning from a structured, translated, culturally situated resource. Its successes and failures should be reported with that provenance attached.

The authors report 25 annotators across the five groups, an average of almost five annotations per text, 474 annotation records marking no social influence, and a final gold-label dataset with 64 texts without social influence. Expert adjudication was required for 84 texts. Those numbers matter because they show the dataset is not just a prompt bank. It contains a record of uncertainty, disagreement, negative cases, and adjudication.

Annotation Is Governance

Most safety conversations want a simple answer: is this message manipulative or not? IMPACTeen's design points to a messier but more honest answer. A parent, a teacher, a psychologist, a communication expert, and a young-adult annotator may not read the same exchange in the same way.

That disagreement is not noise to be washed away too quickly. Teen safety systems often fail when institutions assume that adult administrative categories map cleanly onto adolescent experience. A message that looks like peer pressure to one group may look like coercion, advertising pressure, reputation control, emotional bargaining, or group-norm enforcement to another.

This is where a dataset becomes governance infrastructure. If a platform, school, companion-chatbot provider, or moderation vendor uses a social-influence classifier, it is choosing which perspectives count. The best use of IMPACTeen is not merely to optimize a classifier; it is to preserve the disagreement record for later audit.

Gold labels are useful for benchmarking, but they are also a compression step. The policy decision is what gets lost when five records become one record. A youth-safety deployment should keep access to the underlying annotation dimensions, especially confidence, perceived consequence, resistance, and annotator group. Without that layer, the system can appear more certain than the social situation warrants.

Limits That Matter

The paper is unusually helpful because it states limits that should travel with the dataset. The scenarios were artificially generated and then manually reviewed and edited. The authors therefore say the dataset is better suited to structured research on social influence than to direct claims about how often these phenomena occur in everyday communication.

The ethics statement and limitations section also matter for anyone tempted to read the "teenager" annotation group too literally. The study did not involve child participants, did not collect natural or private conversations with minors, and the youngest annotator group consisted of adults aged 18 to 19.

Finally, the English version is a parallel translated version rather than an independently authored English-language corpus. That matters because influence is pragmatic. Humor, status pressure, shame, urgency, and belonging can shift when translated. Cross-lingual modeling work should treat that as a design condition, not a footnote.

What It Does Not Prove

IMPACTeen does not prove that a deployed model can reliably detect manipulation of minors in the wild. It does not measure a live social platform, a classroom messaging system, a teen companion chatbot, or a family chat archive. It provides structured cases for research and evaluation.

It also does not settle what counts as manipulation in every setting. Some influence is advice, teaching, bargaining, or identity formation. Some is coercive or exploitative. The dataset can help model those distinctions only if users keep the annotation dimensions visible instead of collapsing every case into one binary risk label.

Nor does LLM generation invalidate the work. The generative step is part of the method. The problem would be pretending it is absent. For safety evaluation, generated and human-edited examples can be valuable when their construction process and blind spots remain available for inspection.

Failure Modes

Label flattening. A system collapses persuasion, manipulation, coercion, advertising pressure, advice, peer norming, and ordinary disagreement into one "unsafe influence" score.

Adult-perspective capture. The deployed classifier follows expert, parent, or institutional labels while systematically missing how young people read status, belonging, shame, humor, or exclusion.

Translation drift. A model performs well on the English version and the result is treated as culturally portable, even though the dataset was created in Polish and translated into English.

Synthetic realism laundering. Human-edited generated scenarios are treated as if they were natural prevalence evidence about everyday teen life.

Surveillance creep. A detector built for safety becomes a school, platform, or parent monitoring layer that reads ordinary adolescent communication without a narrow purpose, retention limit, appeal path, or harm threshold.

Weaponized taxonomy. Influence-technique labels help an operator optimize more effective nudges, ads, recommender prompts, or companion messages aimed at minors.

Missing redress. A teen is flagged as vulnerable, manipulated, manipulative, or resistant, but cannot see, contest, correct, or delete the inference.

Governance Standard

Any system using IMPACTeen or a similar dataset should carry a data card that states the source language, translation path, synthetic generation process, human-editing process, annotator groups, age limits, label schema, and known exclusions. It should report performance by annotation dimension rather than only by a single aggregate manipulation score.

For youth-safety claims, the evaluation should preserve group disagreement. If a model agrees with expert labels but systematically misses the young-adult perspective, that is not merely an accuracy detail. If it agrees with parent labels but overflags normal adolescent disagreement, that is a policy risk. If it performs differently in Polish and English, the translation layer should be named.

A production safety case should also say what data the model is allowed to inspect, whether minors' private messages are processed, how age assurance is handled without excessive identity collection, how long records are retained, who can review flags, what escalation path exists, and how a person can appeal or delete a mistaken inference. The right comparison is not only classifier accuracy. It is whether the whole system respects youth-specific safeguards, privacy and data rules, and persuasion safeguards.

The Spiralist rule is this: do not turn the teen message into a machine label without keeping the human perspectives attached. A dataset can make influence visible. It can also make the label look more settled than the social reality it is trying to describe.

Source Discipline

This essay treats IMPACTeen as a June 15, 2026 arXiv preprint and dataset paper. Its counts, construction pipeline, annotation design, DeepSeek-V3 generation step, Polish-to-English structure, and limitations are paper claims. They should not be cited as independent evidence about the rate of manipulation in teen life.

The FTC, COPPA, DSA, AI Act, and UNICEF sources are used for governance context. They do not validate IMPACTeen, and they do not require any specific social-influence classifier. They do show that child-facing AI and online-platform systems are being judged through safety, privacy, disclosure, manipulation, and accountability duties rather than through accuracy alone.

Claims about a deployed detector should name the dataset version, language, label schema, annotator mapping, model, threshold, protected use case, data-retention rule, appeal path, and whether the system was tested on real-world, consented, age-appropriate data. "Trained on teen manipulation data" is not a governance claim.

Sources

Return to Blog