Synthetic data is moving from “nice-to-have” to infrastructure: new reviews map where LLM-generated and domain synthetic data works, where evaluation still breaks, and what governance teams should lock down before scaling.
A Scoping Review of Synthetic Data Generation by Language Models in Biomedical Research
arXiv publishes a scoping review of 59 studies (2020–2025) using large language models to generate synthetic biomedical data, spanning unstructured text, tabular, and multimodal settings. The focus is practical: using synthetic data to address data scarcity and privacy constraints in healthcare research workflows. The review also surfaces recurring gaps—especially around evaluation rigor and how accessible or reproducible these approaches are for teams outside top research labs.
- Data leads should expect more LLM-synthetic text in clinical NLP pipelines, but need standardized utility and privacy evaluation to compare methods.
- Compliance teams will push for clearer documentation of generation prompts, post-processing, and privacy controls as “synthetic” becomes a shared dataset label.
- Founders building tools here can differentiate on evaluation harnesses and auditability, not just generation quality.
Synthetic Data: The New Data Frontier
The World Economic Forum’s 2025 brief frames synthetic data as a response to scarcity, bias, and privacy barriers, with recommendations on governance, quality control, and hybrid real-synthetic strategies. The document reads like a playbook for leadership teams: define intended use, set quality thresholds, and treat synthetic data as part of a broader data governance system rather than a loophole around regulation. It also signals that “responsible synthetic” is becoming a board-level topic in regulated sectors like healthcare and finance.
- Organizations scaling synthetic data will need clear ownership: who signs off on utility, bias, and privacy risk—ML, data governance, or legal.
- Hybrid approaches (real + synthetic) will become the default, which changes how teams design train/validation splits and monitor drift.
Synthetic data generation in manufacturing: a review of methods, domains, and modalities
A DTU Orbit review surveys 18 papers (Jan 2024–May 2025) on synthetic data generation in manufacturing, categorizing methods, application domains, and data modalities. The emphasis is industrial reality: sparse labeled events, proprietary process data, and heterogeneous sensor streams. For teams deploying vision or predictive maintenance models, the review helps map which synthetic techniques are being tried and where evidence is still thin.
- Manufacturing ML teams can use the taxonomy to choose generation methods by modality (e.g., sensor time series vs. images) rather than copying healthcare playbooks.
- Security and IP concerns remain central: synthetic data may reduce exposure, but governance must still address leakage and reverse inference risks.
A Little Human Data Goes A Long Way
An ACL paper reports that mixing small amounts of human data with synthetic data materially improves performance on fact verification and evidence-based QA. The result is a pragmatic message: synthetic data is most effective as an amplifier, not a full substitute. For teams under privacy constraints, this supports workflows where a minimal, tightly governed human dataset anchors training while synthetic expands coverage.
- Budget and privacy wins: smaller “gold” datasets can still drive gains when paired with synthetic augmentation.
- Evaluation should include failure modes like hallucinated evidence and spurious correlations introduced by synthetic generation.
Synthetic data created by generative AI poses ethical challenges
NIEHS highlights that synthetic data has a long history (about 60 years) but generative AI changes the ethical risk profile—especially around privacy, bias, and downstream utility in health research. The piece underscores that “synthetic” does not automatically mean “safe,” and that governance must cover provenance, intended use, and potential harms. For public health contexts, this is a reminder that ethical review and transparency expectations are rising alongside capability.
- Ethics and IRB-style review processes may need to expand to include synthetic dataset creation and release criteria.
- Teams should document bias testing and utility limits to avoid over-claiming representativeness in sensitive populations.