
A Call for Responsible Training Data Production
Spencer M.The way we produce training data is changing. Instead of authoring examples end-to-end, experts increasingly contribute to systems in which models perform some of the work.
As models grow in capability, they can take on more repeatable work with greater independence, efficiency, and consistency. This allows experts to spend less time producing routine artifacts and more time identifying where models still fail, defining what good looks like, and determining what must be built to bridge that gap.
At this frontier, the opportunity is not simply to choose between human-generated and synthetic data. Instead, it is to build production systems in which models extend the reach and utility of human expertise.
Data labs must partner with AI developers and frontier labs to lead this transition responsibly. To do so, data labs must remain accountable for the inputs that enter their production systems, the data those systems produce, and the boundaries under which they operate.
Building the data foundry
Training data is increasingly created through production-grade systems that combine domain expertise, task and environment creation, automated checks, model-based grading, and human review.
Manually authoring, auditing, and verifying each training artifact is no longer the most effective use of an expert’s time. Fully human data production typically scales by adding new contributors, but expertise does not scale uniformly with headcount; in fact, data quality can degrade as the number of contributors increases.
At AfterQuery, we take a different approach. With model assistance, a smaller group of highly qualified experts can amplify their knowledge and experience across larger and more complex data collections, concentrating their human effort in places where it adds the most value.
We think of these expert-model systems as data foundries. We build them around a target need: the work a model should learn to perform, the information it should recognize and use, and the industry standards its output must meet. Our research and engineering teams design the production system, while experts empowered by models develop the data within the system.
Our recent work with Hebbia and NVIDIA on complex credit workflows illustrates what this process looks like. A single task could require reasoning across a base credit agreement and a long series of amendments, combining quantitative terms such as margins and leverage with qualitative provisions such as covenants, governance rights, and events of default.
Producing useful data for this kind of work requires more than asking an expert to write an answer. The process must represent realistic credit workflows, organize the relevant source material, generate tasks, run them against models, evaluate model behavior, and feed those results back into task design and post-training.
Within that process, experts contribute the domain knowledge needed to identify economically meaningful questions, define strong answers, and recognize when a technically plausible response misses the point. Models and software can retrieve relevant clauses, trace amendments, check calculations, generate candidate variations, and test whether outputs satisfy structured criteria. The value comes from how these contributions work together, not from any single component operating alone.
The resulting data is neither purely human-generated nor purely synthetic. It comes from a collaborative process designed by researchers and engineers, with experts and models contributing where each is most effective.
The key question is no longer just:
Is this data human-generated or synthetic?
The more useful questions are:
What produced this data? What was each component responsible for? And why should the result be trusted?
Human expertise moves upstream
Models increasingly take on work that is specifiable, checkable, and repeatable, even when that work is complex or requires completing a long sequence of steps. As they do, expert contribution moves upstream to decisions that shape the rest of the production process.
This division of labor amplifies an expert’s ability. In the credit workflows above, an experienced credit investor might define how covenant terms should be interpreted across amendments and which evidence should inform risk assessments. Models can then apply that framework across many document sets—retrieving terms, tracing changes, generating candidate answers, and surfacing exceptions for review. A single expert decision about task design, a reference answer, or an acceptance criterion can influence an entire collection.
Model-assisted production creates new responsibilities
The same leverage that allows production systems to scale expert decisions can also magnify the consequences of upstream errors. Those consequences do not end when a dataset is delivered; they travel downstream to the models trained on it. Data labs therefore assume responsibility for the quality, provenance, and capability implications of what they produce.
An upstream error can propagate across an entire collection and ultimately become a training signal. Data labs must be able to explain what was checked, what those checks could miss, and what evidence justified releasing the data.
The origins of the data matter as well. Material contributions from experts, sources, and models remain part of the data’s provenance.
Model contributions can also transfer capabilities downstream. When outputs or trajectories from a capable model become training data, another model may learn from them without inheriting the safety training, monitoring, or restrictions applied to the original system.
Frontier labs, including OpenAI and Anthropic, have clarified or strengthened restrictions on using their models to train competing systems. These restrictions reflect both safety and fair partnership. Data labs should respect closed-model* contributions and the terms governing them—not act as proxies that enable others to train on closed-model outputs or distill their capabilities.
These responsibilities will continue to evolve alongside model capabilities and production methods. Data labs are responsible for understanding those boundaries, translating them into production rules, and building systems capable of enforcing them.
AfterQuery’s policy
AfterQuery has a clear policy governing closed-model usage. We are sharing it publicly because we believe frontier AI developers and all of our partners should understand how data labs use models during training data creation.
Our policy is simple:
- We do not use closed models to generate training data.
- We do not package, sell, or otherwise provide closed-model outputs or trajectories as training data.
These commitments apply to completed answers, intermediate trajectories, and any other output generated by a closed model that could be incorporated into delivered data.
We use closed models for research, evaluation, and engineering. These uses help us understand model capabilities, identify meaningful failure modes, and build better data collection infrastructure. Our hard boundary is that closed-model outputs do not become training data.
Open models, by contrast, can contribute directly to data production through candidate generation, quality checks, synthetic rollouts, and training environments. When grounded in expert-designed tasks and engineered checks, these methods allow expert decisions to guide data production at far greater scale.
Across our production systems, we maintain provenance for the people, models, tools, and sources that materially contribute to the data we deliver. We use that provenance to determine whether each contribution is authorized for its intended downstream use and to inform quality controls and release decisions.
Operationalizing our policy
At AfterQuery, every request to a closed-model provider passes through a centralized inference gateway developed in-house. This gives us a single point of control for applying provider-specific policies and enforcing our closed-model commitments across the organization.
Before a request reaches a closed-model provider, our inference gateway classifies its intended use. Requests classified as outside policy are blocked before reaching the provider, while those requiring more context are flagged for subsequent analysis and periodic review.
Centralized logging and post-hoc analysis allow our trust and safety teams to identify patterns that may not be apparent from a single request. We also maintain an internal benchmark containing both compliant requests and adversarial examples designed to test the boundaries of our controls.
We use this benchmark to measure and tune our classifiers, reduce false positives, and identify noncompliant usage that simple rules might miss. Representative samples are periodically reviewed, and the resulting decisions are used to improve our classifiers and refine our operating guidelines.
Dataset-level provenance and release checks provide an additional control over what is ultimately delivered.
We also work directly with frontier labs and other partners to understand changes to their policies and the risks those policies are intended to address. As provider requirements and model capabilities evolve, we update our guardrails, benchmarks, and operating guidelines accordingly.
Standards all data labs should meet
As model-assisted data production scales, data labs have a choice about the standards we establish together. We can make it easier for frontier AI developers to understand and trust the data they train on, or allow production methods to become increasingly opaque.
We believe every data lab should make these three commitments:
- Respect the boundary between using closed models and training on their outputs: Closed models must not be used to generate training data. More broadly, contributions from models, experts, and source materials must be authorized for their intended downstream use.
- Preserve provenance throughout production: Providers should know which people, models, tools, and sources materially contributed to their data. That provenance should remain legible as artifacts pass through multiple stages of generation, editing, and review.
- Make commitments enforceable: Stated commitments must be reflected in actual controls, monitoring, and release decisions. When a production process violates those commitments, there must be a mechanism to stop delivery and address the failure.
These principles guide AfterQuery’s policy. We encourage other data labs to publish their own commitments and explain how they enforce them.
Organizations training on external data should ask their providers to demonstrate how that data was produced and make the evidence part of their purchasing decisions. Together, data labs and AI developers can make responsible production a condition of partnership.
Scaling expertise responsibly
The future of training data will be neither exclusively human-generated nor exclusively synthetic. It will be produced through systems that combine human expertise, models, software, and increasingly sophisticated forms of verification.
At AfterQuery, we believe data labs should compete on their ability to produce high-quality data at scale—and doing so requires production systems they can explain and govern. A lab that understands how and why its data is constructed can identify quality failures and opportunities for improvement, implement changes from source through delivery, and ensure those gains persist as production scales.
Our objective is not simply to produce more data. It is to give everyone the intelligence to achieve what was once out of reach—and to remain accountable for how that intelligence is built.
* We use closed model to describe a model accessed under provider-controlled terms and open model to describe a model whose weights are available under broadly permissive terms.



