Overview

When individual researchers use AI tools without coordination, research teams accumulate inconsistency: different team members prompt differently, apply AI to different workflow stages, validate outputs with different rigor, and document their AI use in incompatible ways. The aggregate result is a research environment where AI assistance is present but invisible, variable but untracked, and increasingly central to research outputs while remaining outside the quality systems that govern everything else. Team-based AI research protocols address this by bringing AI use inside the quality management infrastructure of the research group, defining standards, building shared resources, and creating accountability mechanisms that scale AI's benefits without amplifying its risks.

Title

Lesson 2.1: Team-Based AI Research Protocols

Purpose

This lesson teaches you how to develop and implement standardized AI protocols for your research team or lab. You'll learn to create explicit guidelines for when and how team members use AI tools, establish shared prompt libraries and templates, design verification procedures that maintain quality across collaborators, and build team capacity for consistent AI integration in research.

Designing Team AI Protocols

A team AI protocol is a formal document specifying which AI tools are approved for which uses, under what conditions, with what verification requirements, and with what documentation obligations. Well-designed protocols enable junior team members to use AI assistance confidently within clearly defined boundaries, enable senior team members to trust that AI assistance has been applied with appropriate rigor, and create the documentation infrastructure needed for research accountability and compliance.

The protocol architecture: An effective team AI protocol has five components. First, a scope statement: which research activities AI assistance is permitted for and which are excluded. Second, an approved tools list: which specific AI systems are approved for research use, with what versions, and under what data handling requirements. Third, application guidelines: for each permitted use case, how AI should be prompted, what context must be provided, and what output review standard applies. Fourth, documentation requirements: what must be recorded for each AI interaction, in what format, and where. Fifth, escalation procedures: what to do when a team member is uncertain whether a specific AI application is within protocol, or when AI output quality concerns arise.

Scope decisions in AI protocols: The scope decision, which activities AI assistance is permitted for, is the most consequential design choice. A useful framework distinguishes between four zones. The Automate zone contains activities where AI can execute the entire task and human review is a quality check: formatting, citation standardization, transcript transcription, code formatting. The Augment zone contains activities where AI accelerates human work and human judgment is substantively applied: literature summarization, draft generation, data query construction, visualization creation. The Support zone contains activities where AI provides background information but human judgment drives the work: research design consultation, hypothesis exploration, interpretation support. The Exclude zone contains activities that require human judgment unmediated by AI: participant interaction, ethical decision-making, final scientific interpretation, data analysis for primary outcomes. Protocols that specify which zone each activity belongs to give team members actionable guidance rather than vague principles.

Tool approval processes: Not all AI tools are suitable for research use. Approval criteria for tools used in team AI protocols should include: data handling terms (does the tool train on submitted content? what is the data retention policy?), access controls (does the tool support role-based access? can usage be audited?), output documentation (does the tool provide session logs? can outputs be exported with metadata?), and institutional licensing (does the institution have an approved contract with the provider?). For research involving sensitive data, these criteria become mandatory rather than preferable. A two-tier tool list, approved for all research uses, approved for non-sensitive uses only, accommodates differential data sensitivity within the same protocol.

Prompt standardization: Shared prompt templates are the mechanism by which research teams achieve consistency in AI outputs. When every team member prompts for a literature summary using the same template structure, providing the same domain context, the same output format specification, and the same scope constraints, the variance in outputs across team members is reduced substantially. Prompt templates should be developed collaboratively, versioned, stored in the team's shared documentation system, and reviewed when teams identify recurring output quality issues. A prompt library with 20-30 templates covering the team's most common AI use cases is a realistic and high-value initial target.

Building and Maintaining Shared Prompt Libraries

A shared prompt library is a curated collection of prompt templates that team members use to standardize their AI interactions. The library reduces training overhead for new team members, makes best practices explicit and transferable, and creates a knowledge base that improves as the team's AI experience accumulates.

Prompt library structure: Organize prompt templates by use case category: literature tasks (summarization, gap identification, synthesis prompts), analysis tasks (descriptive statistics prompts, qualitative coding prompts, data cleaning prompts), writing tasks (section drafts, abstract generation, peer review response drafts), and administrative tasks (meeting agenda generation, status report drafts, grant narrative section templates). Within each category, templates should specify the prompt text, the required context inputs (what the user must provide), the expected output format, the verification standard (how to check the output), and the documentation requirement.

Template development process: The best prompt templates emerge from real usage rather than theoretical design. The most productive development process is: identify a high-frequency AI use case, have 3-4 team members each attempt the task using their own prompts, compare outputs, identify what prompt features produced better outputs, and synthesize a template that incorporates those features. This empirical development process produces templates that actually work for the team's specific research context rather than templates that seemed reasonable in the abstract.

Template versioning and maintenance: Prompt templates should be versioned because AI model behavior changes over time and templates that performed well with one model version may perform worse after model updates. Version each template with the date of last validation, the model version it was validated against, and the output quality rating assigned during validation. When a team member notices that a template is producing lower-quality outputs than expected, they should flag it for re-validation rather than modifying it locally, a local modification creates template drift and defeats the standardization the library exists to provide.

Integration with research documentation: Prompt library use should be integrated with research documentation. When a team member uses a prompt from the library, they record the template ID, the specific inputs provided, and the AI output in their AI usage log. This integration enables retrospective reconstruction of AI assistance in any part of a research project and enables the team to analyze which templates were used most frequently, informing library maintenance priorities.

Training new team members: The prompt library serves as the primary training resource for new team members' AI use. An onboarding process that walks new members through the library by category, explaining not just what each template does but why it is structured the way it is, builds both practical skill and conceptual understanding of what makes AI assistance effective. New team members who understand the reasoning behind prompt design are better positioned to adapt templates appropriately than those who treat them as opaque recipes.

Team-Level Verification and Quality Assurance

Team-based AI protocols require team-level verification procedures, not just individual researcher judgment about whether their own AI outputs are acceptable, but systematic processes that provide consistent quality assurance across all team members and all research activities.

Tiered verification framework: Verification requirements should be scaled to the AI use's impact on research integrity. Tier 1 applies to AI assistance in low-impact administrative tasks (meeting agendas, email drafts, formatting): self-certification by the user that outputs were reviewed and meet quality standards. Tier 2 applies to AI assistance in research-adjacent tasks (literature summaries, data descriptions, report drafts): second-team-member review of AI outputs before use. Tier 3 applies to AI assistance in research-critical tasks (primary analysis assistance, results interpretation, methods section generation): expert verification that AI outputs are factually accurate, methodologically sound, and within the scientific scope the data support. The tiered framework prevents both under-verification (treating all AI outputs the same regardless of impact) and over-verification (requiring expert review for meeting agenda generation).

Calibration exercises: Teams that use AI consistently benefit from periodic calibration exercises, structured comparisons in which multiple team members apply the same prompt template to the same input material and compare their outputs. Calibration reveals: differences in how team members interpret template instructions, differences in what team members accept as adequate output quality, and differences in what team members document. Calibration exercises are the team-based equivalent of inter-rater reliability checks in research methodology. A team with high calibration produces AI-assisted research with consistent quality. A team with low calibration produces variable quality that may be difficult to detect until peer review or post-publication scrutiny.

Output auditing: Beyond routine verification, teams benefit from periodic audits of AI-assisted outputs. An audit selects a random sample of AI-assisted work products from a specified period, applies the relevant verification standard to each, and assesses the proportion that meet the standard. Audit findings identify whether verification is working as intended or whether there are systematic gaps, cases where team members pass AI outputs through verification without catching errors that the verification standard should detect. Audits can also identify AI use patterns that were not anticipated when the protocol was designed, suggesting scope or guideline updates.

Feedback mechanisms: Team members who identify AI output quality problems should have a designated feedback mechanism, not an informal complaint, but a structured process that captures the problem, the template or use case involved, the quality issue observed, and the team member's recommendation for addressing it. Structured feedback enables the team lead to distinguish isolated incidents from systematic patterns and to update templates, verification procedures, or scope guidelines in response to accumulating evidence.

Documentation Standards and Disclosure Protocols

Research accountability requires that AI assistance in research be documented with sufficient specificity to reconstruct what AI did, when, and to what effect. This documentation obligation extends beyond publication disclosure to the complete research record maintained for the team's own accountability.

Team AI use logs: The foundation of team AI documentation is a structured AI use log maintained for each research project. The log records: date, team member, AI tool and version, use case category (from the protocol scope framework), prompt template used and any deviations from the template, summary of input material provided, summary of AI output received, verification tier applied and by whom, and documentation of any output quality issues identified. This level of detail enables retrospective reconstruction of AI assistance in any research activity and enables team leads to review AI use patterns across the project.

Project-level disclosure documents: In addition to per-interaction logs, teams should maintain a project-level AI disclosure document, a summary of how AI was used in the project overall, by phase, with the verification standards applied and the proportion of outputs that were modified from AI draft to final use. This document serves as the basis for any required disclosures in publications, grant reports, or research integrity reviews. Producing a project-level disclosure from complete per-interaction logs is straightforward; producing one retrospectively without logs requires imprecise reconstruction.

Funder and institutional reporting: Many funders and institutions now require disclosure of AI use in funded research. The scope and format of these requirements vary: some require only high-level disclosure (AI was used in data processing), others require detailed accounting of AI contributions by research phase. Teams that maintain complete AI use logs can meet any reporting requirement; teams that rely on informal recollection cannot. Proactive documentation is less labor-intensive than retrospective reconstruction, particularly for multi-year projects with multiple team members.

Publication disclosure standards: Journal policies on AI disclosure require corresponding authors to report AI use in manuscript preparation. For research teams with documented AI use protocols, this disclosure is a matter of extracting from the project log. For teams without documentation, the corresponding author must solicit disclosure information from each co-author individually, a process that typically reveals both legitimate AI use that was not disclosed and AI use that co-authors are uncertain how to characterize. Documentation protocols that run throughout the research process prevent this uncertainty.

Building Team AI Capacity

The difference between a research team with a protocol on paper and a research team with an effective protocol in practice is capacity: team members' ability to apply the protocol correctly, identify when they need guidance, and contribute to the protocol's improvement over time.

Tiered training model: AI capacity building in a research team is not a single training event. It is an ongoing process that should be matched to team members' roles and current proficiency. An entry-level training covers: the protocol's scope framework and why each zone was classified as it was, how to use the prompt library, what the documentation requirements are and how to fulfill them, and who to ask when uncertain. An intermediate training covers: prompt design principles, verification best practices, calibration exercises, and how to contribute feedback on template quality. An advanced training covers: prompt library maintenance, audit procedures, protocol revision processes, and how to evaluate new AI tools for potential addition to the approved list.

Role-specific guidance: Different team roles require different AI guidance. Principal investigators need to understand AI governance (protocol design, funder compliance, institutional policy) and AI strategy (which workflow stages most benefit from AI integration). Postdoctoral researchers and senior graduate students need to understand AI for their own research workflows and the mentoring obligation to ensure junior researchers apply AI appropriately. Junior graduate students and research assistants need to understand the operational protocol: what tools to use, what templates to apply, what to document, and what to escalate.

Mentoring AI practices: In research training environments, PIs and senior researchers have a mentoring obligation regarding AI use, parallel to their existing obligations regarding research methodology and research ethics. A graduate student who uses AI without guidance may learn practices that harm their long-term development: excessive reliance on AI-generated text that atrophies writing skill, inadequate verification practices that produce research errors, or inappropriate uses that create research integrity problems. Mentoring that explicitly addresses AI practice, through regular review of students' AI usage logs, discussion of specific AI use decisions, and modeling of effective AI-human collaboration, is a component of contemporary research training.

Evolving protocols: AI tool capabilities change rapidly, and team protocols that were appropriate for 2024 tool generations may be suboptimal or insufficient for 2026 tool generations. Protocols should be reviewed formally at least annually, and informally whenever a significant new AI tool capability is introduced or a significant AI-assisted research quality issue is identified. The review process should include all team members, not just PIs, because junior researchers who use AI most intensively often have the most direct insight into what is working and what is not in the current protocol.

Governance, Ethics, and Institutional Integration

Team-based AI protocols do not exist in isolation. They operate within institutional, funder, and professional community frameworks that set minimum standards and create compliance obligations. Effective protocol design integrates these external requirements.

Institutional AI governance: Most research institutions have developed or are developing AI use policies for research. These policies address data governance (which AI tools are approved for which data categories), research integrity (what AI disclosure is required), and compliance (what institutional oversight AI-assisted research requires). Team protocols must be consistent with institutional policy, not more permissive than what the institution permits and documented in a way that institutional compliance review can assess. Research teams that develop protocols in consultation with their institution's research integrity office are better positioned to navigate funder and publication compliance requirements.

Data governance integration: AI tool use creates data governance obligations when research data is provided to AI systems as context. For research subject to IRB protocols, HIPAA, FERPA, export control regulations, or commercial confidentiality agreements, the question 'can this data be provided to this AI tool?' must be answered based on the data governance framework that applies, not based on whether the AI output would be useful. Data governance integration in team AI protocols specifies, for each approved tool, the data categories it may and may not receive as input, and the approval process for any exceptions.

Research integrity implications: AI assistance in research creates three categories of integrity implications that team protocols must address. First, attribution integrity: ensuring that AI assistance is disclosed as required and that authorship credit reflects genuine human intellectual contribution. Second, methodological integrity: ensuring that AI-assisted analyses, literature reviews, and data processing are conducted with the verification rigor required to support reliable research findings. Third, replication integrity: ensuring that AI-assisted research is documented with sufficient specificity that the same research could be replicated, meaning that AI prompts, tools, and outputs used in critical research steps are recorded in the research archive.

Cross-team and consortium protocols: When multiple research teams collaborate, within a multi-site grant or a research consortium, divergent AI protocols create compliance and quality risks. Teams that use AI differently may produce research contributions of different reliability and with different documentation completeness. Multi-team collaborations benefit from a consortium-level AI protocol that establishes minimum standards across all participating teams, with higher-standard team protocols permitted but not lower. The lead team on a collaborative grant typically takes responsibility for establishing and monitoring the consortium AI protocol, with designated AI protocol liaisons at each participating site.