Natural language is the foundation of PyAuto’s approach to artificial intelligence. Our goal is to allow users and developers to begin by clearly describing what they want to achieve, while AI translates that intent into the necessary software APIs, syntax and implementation.
Natural language does not remove the need for expertise. Instead, it helps distinguish between knowledge that is essential to the science and technical detail that can reasonably be delegated to AI.
Use natural language to remove unnecessary barriers, but not to bypass the knowledge required to direct, understand and validate the work.
This policy describes how that principle applies to the use, teaching and development of PyAuto software. It is informed by How to Craft the Right Language AI Policy For Your Research Group (Some Assembly Required), which argues that AI policies should reflect the values of the communities adopting them.
PyAuto combines three priorities:
- Natural-language accessibility: reducing the technical distance between a scientific idea and its implementation.
- Learning and expertise: helping users develop the knowledge required to describe, direct and evaluate an analysis.
- Scientific trustworthiness: validating AI-assisted work through testing, reproducible examples, attribution and transparent development records.
This is a living policy and will evolve alongside the capabilities and risks of AI systems.
PyAuto aims to make natural language the primary starting point for scientific analysis and software development.
As a rule of thumb, before delegating a consequential task to AI, you should be able to describe:
- what you are trying to achieve;
- the relevant data and assumptions;
- the intended scientific or technical method;
- the expected output; and
- how you will determine whether the result is correct.
If you cannot yet describe these clearly, you may not have the knowledge required to direct or validate the task. In this situation, AI should first help you learn, explore the available approaches and refine your description—not immediately produce a final scientific result.
The PyAutoLens AI Assistant demonstrates this approach. A user can describe a lens-modelling problem in natural language, including the dataset, lens and source models, computational requirements and desired figures. The assistant translates this scientific description into the appropriate PyAutoLens workflow without requiring the user to know every Python class, configuration option or API call.
A new user may not initially be able to provide such a detailed description. They can instead begin with questions, discuss the available modelling approaches or use Teacher Mode to build the knowledge needed to formulate the analysis. Natural language is therefore both an interface for performing tasks and a means of learning how to define them.
All PyAuto AI support must include routes for developing the knowledge and expertise needed to use the software responsibly. AI should be able to:
- explain the scientific and statistical concepts underlying an analysis;
- help users express their objectives and assumptions more precisely;
- describe the available modelling choices and their consequences;
- explain generated code and results;
- identify assumptions, limitations and possible failure modes; and
- direct users to human-readable learning material and runnable examples.
The principal learning resources are:
- HowToLens, which teaches gravitational lensing, Bayesian inference and lens modelling from first principles;
- the autolens_workspace, which provides runnable examples and guides covering PyAutoLens use cases; and
- the guided and Teacher modes provided by the PyAutoLens AI Assistant.
AI should cultivate the understanding required to hold a meaningful scientific conversation about the task. It should not turn researchers into passive reviewers of generated work. The person using AI remains responsible for the scientific decisions, interpretation of the results and recognition of when additional expertise is required.
Natural language can appropriately replace knowledge of syntax that is incidental to the scientific objective.
For example, a researcher may understand how they want a figure to look without knowing the exact Matplotlib API needed to adjust its axes, labels, colour map or layout. Similarly, they may understand the intended structure of a paper or equation without remembering the required LaTeX commands. In these cases, natural language allows the researcher to communicate the desired result while AI handles the implementation details.
Understanding the inner workings or complete syntax of Matplotlib or LaTeX is not normally integral to the lensing science being performed. Requiring researchers to master every incidental API before making a small change would create an unnecessary barrier.
This does not mean that knowledge is unnecessary. Some general understanding of the tool or domain is usually needed to describe the intended result, steer the AI and recognize incorrect output. The required expertise should depend on the consequences of the task: detailed knowledge of Matplotlib syntax is rarely essential, whereas understanding the assumptions of a lens mass model is.
The distinction is between:
- syntax and implementation details, which can often be safely delegated through natural language; and
- scientific concepts, assumptions and judgments, which the researcher must understand sufficiently to describe and defend.
The same natural-language principle applies to PyAuto software development. Development begins with a plain-English description of what should change, why it should change, the expected behaviour and how success will be demonstrated. AI can then translate this intent into source code, tests, examples and documentation.
PyAuto does not require every line of Python source code to have been manually written or read line by line by a human. Confidence in software does not arise from human authorship alone; it arises from evidence that the software behaves correctly.
AI-assisted changes must therefore be validated at the appropriate levels:
- Unit tests test individual components and numerical behaviour, as illustrated by the PyAutoLens test suite.
- End-to-end workspace examples demonstrate that complete scientific workflows run correctly using the public API in the autolens_workspace.
- Integration and dedicated test scripts exercise complex workflows through the autolens_workspace_test and other task-specific validation repositories.
- Scientific validation compares important results with analytic calculations, established implementations, benchmark datasets, published results or expert judgment where appropriate.
An AI-generated feature is not complete merely because the code runs or appears plausible. Its original natural-language requirements should be translated into tests and examples that demonstrate the requested behaviour. Building this test scaffolding is central to how PyAuto establishes that AI-assisted software works.
Tests do not remove the need to understand the scientific assumptions or public interfaces. The person who proposes, reviews or releases an AI-assisted change remains accountable for its behaviour and for the adequacy of its validation.
AI systems may reproduce or adapt patterns from existing software, algorithms and publications without identifying their origin. Natural-language generation can obscure this provenance because the user may receive working code without being shown where its ideas or implementation originated.
AI-generated code must therefore be treated as having uncertain provenance until it has been reviewed. Contributors must:
- investigate whether substantial or specialized generated code resembles an existing implementation;
- identify and cite the original scientific method or software where appropriate;
- check that any reused implementation is compatible with its original licence;
- avoid incorporating code whose licence or provenance is unclear; and
- add relevant papers and software references to the PyAutoLens citations documentation.
If an implementation substantially follows an existing published method or software package, attribution is required even if the AI did not reveal the connection. AI assistance does not remove normal responsibilities concerning plagiarism, copyright, licensing or scientific citation.
PyAutoScientist is the open, human-led AI development ecosystem used across PyAuto. Its purpose is to turn natural-language scientific and software requirements into planned, tested and released code while retaining a traceable connection to the original human intent.
Development begins in PyAutoMind, where a plain-English description records what should be done and why. This description is connected to a GitHub issue, development branch, pull request and dated completion record. Planning and agent responsibilities are documented through PyAutoBrain, while repository health and release readiness are independently monitored by PyAutoHeart.
Together with Git history, tests, issues and pull requests, this provides an end-to-end public record of:
- the original natural-language request;
- why the change was needed;
- how the request was interpreted and planned;
- what decisions were made during implementation;
- what tests and scientific validation were performed; and
- how the change reached the released software.
This traceability is particularly important when much of the implementation is produced through natural-language interactions with AI. It allows the community to compare the original intent with the final behaviour and to inspect, question, reproduce or improve the work.
Private, embargoed, personal or collaborator-controlled information must not be submitted to an AI service unless its use has been authorized and the service’s storage, training and privacy terms are appropriate.
Researchers must also follow the AI and disclosure policies of their institution, collaboration, journal or funding body. Where disclosure is required, it should describe how AI participated in the work and how its outputs were checked—not merely state that an AI tool was used.
Before accepting AI-assisted scientific or software work, ask:
- Can I clearly describe in natural language what was done and why?
- Do I understand the scientific assumptions and consequences?
- Can I explain and defend the important decisions?
- Can the result be independently checked or reproduced?
- Have appropriate tests, examples and validation been added?
- Have citations, provenance and licensing been checked?
- Is the development history connected to the original intent?
- Was all information shared with the AI permitted to be shared?
If the answer to any relevant question is no, the work is not yet complete.
This policy was written using natural-language instructions to guide ChatGPT 5.6 Sol. The human author defined the principles, supplied the examples, evaluated the outputs and iteratively refined the final text, and retains responsibility for its content.