SAIN Research Hub Handbook

Version: 1.0 Date: 29/04/2026 *Contact information: research@safeainetherlands.org Authored by* Alexander Müller (Co-Director) Thomas Brcic (Co-Director) lija Lichkovski (Research Lead)

Table of Contents

*Purpose, Principles, and Scope 3*

Who This Handbook Is For 3

Mission and Research Scope 3

Core principles 4

*Structure of the SAIN Research Hub 5*

Program Modes 5

Eligibility 5

Application Requirements and Criteria 5

Roles and Responsibilities 5

*Research Lifecycle at SAIN 7*

Onboarding 7

Research Proposal Stage 7

Assignment & Setup 7

*Expectations and Standards 9*

*Deliverables and Milestones 9*

*Funding, Support, and Constraints 10*

*Communication Norms 10*

*Research Integrity, Safety, and Ethics 10*

Integrity 10

Responsible Use of AI Tools 11

Dual-Use and Infohazards 11

*Wellbeing and Support 11*

*Escalation, Conflict Resolution, and Sanctions 12*

When to Notify Someone 12

Notification Path 12

Possible Sanctions or Remedies 12

Purpose, Principles, and Scope

Who This Handbook Is For

This handbook applies to

Researchers and students working under SAIN’s Research Hub
Supervisors affiliated with SAIN
Collaborators and visiting researchers under open projects
Outsiders who are interested in SAIN’s Research Hub

It functions as a comprehensive guide for expectations, processes, and norms related to the SAIN Research Hub.

Mission and Research Scope

SAIN (Safe AI Netherlands) exists to raise awareness of the full spectrum of existing and potential harms from AI, contribute to shaping mitigation priorities through ongoing discourse, and support the realization of effective solutions.

We care about high-quality research. The SAIN Research Hub offers a platform where supervisors, collaborators, and students can connect to advance AI safety research on a global and interdisciplinary scale.

We welcome any project that meaningfully contributes to AI safety, and are open to technical, governance, or conceptual AI safety research. To make this concrete, our hub especially encourages work in (but not limited to) the following areas:

Technical alignment & interpretability

Mechanistic interpretability
Representation learning & feature sparsity for safer control
Steering and controllability of LLMs and other foundation models
Adversarial robustness and red-teaming of models
Detection and mitigation of deceptive or misaligned behaviors
Scalable oversight, debate, constitutional AI, and related methods
Robustness of RL agents and multi-agent systems
Safety benchmarks and evaluation metrics
Agent Foundations
Neuroscience-inspired alignment

Societal impacts, governance & policy

AI governance, regulation, and standards (with emphasis on EU / Dutch context)
AI and democratic processes (misinformation, polarization, information integrity)
AI and labor, job displacement, economic and social impacts
Privacy, surveillance, and data protection in AI deployment
Risk assessment for high-stakes domains (e.g. CBRN, cybersecurity, critical infrastructure)

Meta and foundational topics

Research methodology and evaluation in AI safety
Benchmarking AI systems’ persistence, persuasion, or autonomy
Forecasting and scenario analysis for transformative AI
Epistemics, information hazards, and responsible communication in AI safety

Importantly, if you’re unsure whether a topic “counts” as AI safety, the default answer is yes, if you can articulate a plausible pathway from your work to reducing AI-related risk. With AI-related risks, we mean it broadly: risk to lives, risk to equality, risk to rights, etc.

Core principles

Across all projects, SAIN aims to uphold various principles.

We appreciate intellectual honesty. Always accurately communicate what is known, unknown, and uncertain.
Proactively consider both near-term and long-term harms, as well as dual-use concerns (explained more in-depth below).
Maintain a supportive environment across disciplines and seniority levels. There is zero room for harmful authority.
Feedback, critique, and review are essential tools to ensure your work is up to standard. When receiving feedback, remember this is not a personal attack. When giving feedback, attack the ideas, not the person.

Be extra wary of the above principles when deadlines are approaching. Every human is fallible, and in situations where compromising on the above principles can lead to publishing or not, it is extra important to uphold the above principles rather than succumbing to the pressure of research.

Structure of the SAIN Research Hub

Program Modes

The Research Hub currently operates in the two following modes:

Supervised Research Matching

The supervised research matching mode allows students and early-career researchers to apply for guided research projects with mentorship from one of SAIN’s supervisors. Through this process, SAIN provides logistical and financial support (e.g., compute). Projects aim to evolve into publishable research. Experienced researchers (often PhD+, but with exceptions) with an interest or track record in AI safety research can join SAIN’s Supervisor Team. We welcome supervisors from any discipline.

Open Collaboration Opportunities

Not all research requires formal supervision. SAIN connects researchers directly and provides support (e.g., compute). We streamline collaboration so you can focus on advancing AI safety. If you have an interesting (research) proposal, then kindly submit it via the form on our website. If you want to join a project, please reach out to the contact person of the particular project you’re interested in. In these projects, no one will play the “supervisor” role, although there will be one person responsible for leading the project. This will likely be the person who originally designed the project.

Eligibility

For open collaboration and supervised projects, we allow anyone with sufficient background to participate in the project. This includes students (Bachelor, Master, PhD), researchers from industry, and even people with no formal affiliation.
Remote participation works as well.
To become a supervisor, we mostly require you to be PhD+, although in exceptional cases this can differ.

Application Requirements and Criteria

Our Research Hub works on a rolling basis. Thus, you can apply at any time, and if there is an opportunity and you are a good fit, you will be accepted\!
For Supervised Research Matching, we will get back to you in a working week.
For starting a project for Open Collaboration, we appreciate a fully fleshed-out proposal before contacting us, but if necessary, we can help shape it.

Roles and Responsibilities

Although later sections spell roles and responsibilities out in more detail, we briefly cover them here as well.

Researchers and students drive day-to-day research
Supervisors provide intellectual guidance, supervision, and high-level project direction. If desired or need be, they step in where necessary.
Research Lead and Research Operations focus on project management, unblockers, and research support. Primary contact should go to them: research@safeainetherlands.org.

Research Lifecycle at SAIN

Onboarding

Every new researcher in the Research Hub should have a rough overview of current projects, groups, and supervisors, and get an explanation of this handbook and where it lives. If relevant, the researcher should get access to compute or other support. They should ideally join our discord, where they will receive the above relevant information and more.

Research Proposal Stage

Before a project is approved as an SAIN Research Hub project for Open Collaboration, the researcher completes a Research Proposal. Importantly, the below is only needed if the goal is a full paper and collaboration. For other outputs, exclude what you think is unnecessary (e.g., for a mechanistic interpretability tool you do not need a title and abstract).

Title, problem statement, and motivation
Background & related work
Research questions and hypotheses
Proposed methodology
Safety & risk considerations (think of dual-use, misuse, infohazards)
Needed resources (potential datasets, compute estimations, and how many collaborators are needed and for how much time)
Timeline & milestones
Expected outputs

The proposal is reviewed by SAIN’s Research Team who decides to approve it as a Research Hub project (with or without revisions) or to decline.

For Supervised Research Matching or joining an open collaboration project, please refer to the forms (Supervised Research - Expression of Interest or by reaching out to research@safeainetherlands.org).

Assignment & Setup

Once approved, either as a new project for open collaboration, to join an existing open project, or to do supervised research matching, the following happens:

You are brought into contact with the relevant people.
For supervised research matching, a supervisor is formally assigned and a new research project is created in the discord.
For joining an existing open research project, you are officially added to the project and are added to the discord.
For a new research project, we add the project to our website, add you to the discord, and wait and promote to allow people to join.
Expectations around time commitment (e.g., 5-10h/week vs. near full-time) and meeting cadence are set. We highly encourage communication through our Discord channel, although if preferred, other communications channels can be set.
The project is registered internally and updated dynamically as the project progresses.
The SAIN Research Team sets up a GitHub repository (if applicable) under the GitHub organization and gives proper access to the project members.

Execution Phase

During the project, the standard pattern is:

Weekly / biweekly meetings
Discuss recent progress and blockers
Refine research direction and scope
Review experiments, results, and drafts
Weekly / biweekly progress forms
A form for researchers and supervisors to note what was tried, progress, next steps, etc.
This ensures accountability and allows us to track how the projects are progressing.
Contact with Research Team
Whenever necessary, anyone operating inside SAIN’s Research Hub can communicate with our Research Team. Main points of contact are research@safeainetherlands.org or the Discord channel.

Completion Phase

A project is considered “complete” when

The main research question has been adequately addressed (including negative or null results) and has been published somewhere in a presentable form of output.
If relevant, infohazard and dual-use considerations have been treated appropriately.

Post-Project

Output and relevant other resources (e.g., data, code, notes, and drafts) are stored in SAIN’s project archive.
Short note on what went well and what could have gone better, ideally shared with the SAIN community.
The project may be extended as a follow-up.

For Supervisors

The following procedure applies to the supervisors:

The research proposal is prepared as mentioned above,
Researchers are assigned to supervisors / projects,
The supervisor sends the research proposal alongside the list of researchers in the project (with contact details; discord username (and github username, if applicable) as a minimum) to research@safeainetherlands.org.
If applicable, a repository is created in the SAIN github organisation, and the researchers and supervisors are given collaborator access to the repo.
Every week, on Sundays, the researchers fill in the weekly check-in form.
Every two weeks, on Sundays, the supervisor fills in the bi-weekly check-in form.

Expectations and Standards

Researchers are expected to:

Take primary responsibility for the day-to-day progress of their project.
Show up to supervisor meetings prepared (e.g., having a clear agenda, results, or specific questions).
Fill in the weekly progress log.
Communicate early about blockers or personal constraints.
Follow SAIN’s Core Principles

Supervisors are expected to:

Provide regular guidance and mentorship.
Help ensure the project is well-scoped and realistic.
Meet with the researcher at an agreed frequency (usually weekly or biweekly) and show up prepared.
Give timely feedback on drafts and research directions.
Watch out for:
Projects that have become stuck in unproductive directions.
Researcher wellbeing issues (e.g., stress or burnout)
Mention concerns to the Research Lead or team where appropriate.

The Research Team is responsible for:

Maintaining up-to-date documentation and this handbook.
Onboarding new researchers and supervisors.
Keeping track of Research Hub projects and outputs.
Acting as a contact point via research@safeainetherlands.org or via the Discord channel.

Deliverables and Milestones

Each project should specify:

Intermediate milestones. For example:

Week 1-3: getting to know each other, the project, and doing a literature review.
Week 4-8: First experimental proof of concept.
Week 8-10: Internal discussion about the project, initial results, and what to improve.
Week 10-12: Implementation of the found weaknesses, new results.
Week 12-14: Writing up the final report and submission.

Final deliverables. One or more of, for instance:

Research paper at a conference, journal, or workshop.
Blog post at SAIN’s Substack and/or LessWrong.
Policy brief.

Authorship and Credit

Authorship should reflect substantial intellectual and implementation contributions.
Order of authorship should be handled internally within the project team.
SAIN and the Research Hub should be acknowledged in publications.
SAIN is allowed to promote your work as being enabled by the Research Hub.

Funding, Support, and Constraints

This applies mostly to technical projects.

Financial support
SAIN does not provide stipends or salaries for projects. This is purely voluntary and/or educational.
If funding for conference trips, workshop registrations, or travel are needed, we may be able to help you with this within reason.
Compute & Resources
Once the project is specified, the Research Team will strive to give your project easy access to compute and/or other resources.
The amount of compute SAIN can help with will depend on the project. This will be specified at the beginning of each project.
We encourage you to prioritize computationally inexpensive projects whenever possible.

Communication Norms

Use agreed channels (ideally being the Discord channel) for project communication.
Respond within a reasonable timeframe (e.g., 2-3 working days for non-urgent matters).
For meetings (responsibility of the supervised researcher):
Have a simple agenda.
Start with briefly updating what has happened since the last time. We hope filling in our weekly progress logs helps you in keeping track of this.
End with concrete next steps.
Be explicit about availability (travelling, family time, exam periods, etc.)
Default to kind, precise, and honest communication.

Research Integrity, Safety, and Ethics

Integrity

Cite sources properly (after carefully knowing what the sources mean to say) and avoid plagiarism.
Be honest about experimental results, limitations, and negative findings.
Keep sufficient documentation to allow others to roughly reproduce your work.

Responsible Use of AI Tools

We are not against usage of AI tools such as LLMs in coding or writing, and in cases where a researcher feels it benefits them especially and they have carefully considered it, we even encourage them to use it. However:
Always check outputs carefully. Do not uncritically trust generated content.
Avoid feeding sensitive or confidential data into tools which you aren’t sure of have a responsible way of dealing with data.
Where relevant, disclose use of AI tools.

Dual-Use and Infohazards

Work on AI Safety can often, perhaps somewhat cynically, have paradoxical effects (i.e., they make AI systems, the AI ecosystem, or individuals inside the AI ecosystem, less safe). Please do not underestimate in what a complex manner these paradoxical effects may take place. Of course, it is impossible to see exactly how research findings play out, but at the very least be aware of this and spend some time thinking about whether your work could significantly increase capabilities or misuse risk if widely shared.

If in doubt, please discuss this with your supervisor and/or SAIN’s Research Lead before public dissemination.
Possible mitigations could be:
Redacting sensitive details.
Internal-only or restricted-access reports.
Delaying publication until risks are better understood.

Wellbeing and Support

We hope that people in the SAIN Research Hub is a place where personal wellbeing is something that is taken seriously, and that support is there for anyone that needs it.

AI safety research, just like any research, can be intellectually and emotionally demanding, especially because of the intense “publish or perish” that is, unfortunately, much too common.
Researchers are encouraged to:
Be honest about workload and stress levels with supervisors (and vice versa).
Take regular breaks and keep boundaries between work and rest.
Realize that what matters most is gaining a better understanding of AI safety, not whether one publishes or not.
Speak to the SAIN team if conflicts or issues arise.

Escalation, Conflict Resolution, and Sanctions

Although we hope it will not happen, it is nonetheless important to make it clear when one should consider notifying someone, how a conflict should be resolved, and potential sanctions.

When to Notify Someone

You should consider notifying someone when:

A supervisor or researcher relationship breaks down (e.g., through persistent unavailability or unresolved conflicts).
There is suspected misconduct (e.g., data falsification, (sexual) harassment, plagiarism).
There are serious concerns about whether the direction of the project is still aligned with AI safety.
Project scope, expectations, or time commitments have become misaligned.

Notification Path

A typical notification path is:

Researcher ↔ Supervisor (try to resolve directly).
SAIN’s Research Lead (or anyone in the Research Team).
SAIN Co-Directors (in case of conflict with the Research Lead or anyone in the Research Team).

Importantly, if at any stage you feel unsafe or subject to harassment that you do not feel comfortable with addressing to the party that is bothering you, please feel free to skip levels and contact SAIN leadership directly.

Possible Sanctions or Remedies

Depending on the severity and nature of the issue, possible actions include:

Clarifying expectations and adjusting milestones.
Reassigning supervisor or researcher to another project.
Re-scoping or pausing the project.
Removal from the Research Hub, or from SAIN roles, in severe cases.