EthicAlly: AI for SSH Research Ethics
- EthicAlly is a proof-of-concept AI tool designed to support social sciences and humanities research ethics through structured proposal review.
- It employs constitutional AI and collaborative prompt engineering to blend universal ethics with SSH-specific contextual considerations.
- The system assesses proposals by flagging ethical conflicts, providing detailed reports and risk scores to enhance REC review processes.
Searching arXiv for the provided EthicAlly paper and closely related ethics-support work to ground the article in current sources. arxiv_search.query{"3search_query3 OR ti:\3"EthicAlly\" OR ti:\3"AI-Powered Research Ethics Support\"","start":3search_query3,"max_results":5} EthicAlly is a proof-of-concept AI assistant for social sciences and humanities research ethics support. It is described as a system that helps researchers “identify, reason through, and document ethical issues in study design and prepare stronger submissions for Research Ethics Committees (RECs), without replacing human oversight” (&&&3search_query3&&&). In this formulation, EthicAlly addresses a specific institutional problem: biomedical ethics models often map poorly onto social science and humanities methodologies, while access to tailored ethics training, SSH-appropriate advisory services, and RECs with relevant expertise remains uneven in parts of Europe and in many low- and middle-income countries (&&&3search_query3&&&). More broadly, EthicAlly belongs to a family of ethically aligned AI approaches that seek to constrain or guide flexibility, adaptability, and creativity by explicit ethical principles, contextual judgment, and auditable decision structures rather than by unconstrained optimization alone (Rossi et al., 2018).
3id:(Grohmann, 15 Jul 2025) OR ti:\3. Historical and institutional setting
EthicAlly is situated in a long-standing critique of the transfer of biomedical ethics oversight into domains characterized by interviews, surveys, focus groups, ethnography, online and digital methods, and participatory approaches. The central concern is that rigid biomedical checklists can misfit open-ended designs, create “tick-box” compliance, or overlook SSH risks that are interpretive, socio-political, and context-bound (&&&3search_query3&&&). The prototype therefore addresses both a methodological mismatch and a capacity problem: in some settings, SSH researchers lack tailored ethics support, and institutional RECs are required to assess proposals involving unfamiliar qualitative or politically sensitive designs under heavy workload and short turnaround expectations (&&&3search_query3&&&).
This placement is consistent with the broader notion of ethically bounded AI. In that literature, ethically bounded AI is defined as AI systems whose flexibility, adaptability, and creativity in pursuing goals are constrained by explicit ethical principles and virtues, such as fairness, social norms, and laws (Rossi et al., 2018). EthicAlly applies that orientation to research ethics support rather than to autonomous decision execution. Its purpose is preparatory and advisory: it supports researchers in ethical research design and REC preparation, and may ease institutional REC burden, but “without attempting to automate or replace human ethical oversight” (&&&3search_query3&&&).
A plausible implication is that EthicAlly should be read not as an automated ethics adjudicator, but as an intermediate layer between abstract norms and institutional review. That reading is reinforced by adjacent work that seeks to operationalize ethical reasoning through explicit structures rather than through generic prompting alone. Modular ethically bounded AI separates goals, preferences, and ethical principles and combines them contextually through thresholds and blending parameters (Rossi et al., 2018). ECCOLA translates ethics into sprint-level prompts and work products across a software lifecycle (Antikainen et al., 2021). RESOLVEDD contributes a nine-step decision aid that structures ethical reflection and documentation (&&&3id:(Grohmann, 15 Jul 2025) OR ti:\3search_query3&&&). EthicAlly extends the same general movement into SSH ethics review support.
3 OR ti:\3. System architecture and interaction model
EthicAlly is implemented as a web app that accepts a researcher’s proposal plus contextual fields and returns a structured ethics assessment. The prototype uses Anthropic’s Claude Sonnet 4 via API and includes fields for Field/Discipline, Country/Region, Proposal text, and data-processing consent checkboxes (&&&3search_query3&&&). It currently does not use retrieval-augmented generation or institutional knowledge bases, although those are identified as future directions (&&&3search_query3&&&).
The system’s workflow is intentionally structured rather than open-ended. If the field is clinical or biomedical, the system is instructed to decline review politely and refer elsewhere; otherwise it applies a prompt scaffold that combines universal principles, SSH contextual principles, legal context, disciplinary context, and explicit guardrails (&&&3search_query3&&&). The returned report is structured and typically produced in approximately 33search_query3–63search_query3^ seconds (&&&3search_query3&&&).
| Component | Function | Stated characteristics |
|---|---|---|
| Web UI | Collects context and proposal text | Includes Field/Discipline, Country/Region, Proposal text, Data-processing consent checkboxes |
| Backend | Sends structured prompt and user inputs to Claude Sonnet 4 | Returns structured report in ~33search_query3–63search_query3 seconds |
| Output report | Ethics assessment for REC preparation | Disclaimer; summary assessment; compliance analysis; issues and recommendations; ethics risk score; supplementary materials assessment |
The report format is fixed by the co-created startup prompt. It contains: a disclaimer; a summary assessment; compliance analysis covering core principles, discipline-specific standards, and legal or regulatory checks such as GDPR; issues and recommendations prioritized by severity; an ethics risk score on a 3id:(Grohmann, 15 Jul 2025) OR ti:\3–5 scale with justification; and, if materials are supplied, a supplementary materials assessment (&&&3search_query3&&&). The risk scale is categorical rather than formulaic: “3id:(Grohmann, 15 Jul 2025) OR ti:\3^ Low: minimal concerns; standard protections” through “5 High: fundamental ethical problems; redesign needed” (&&&3search_query3&&&).
This design places EthicAlly in a lineage of structured, inspectable ethics-support systems. The resemblance is strongest to methods that emphasize explicit inputs, explicit ethical criteria, and documented outputs. ECCOLA, for example, is organized around 3 OR ti:\3id:(Grohmann, 15 Jul 2025) OR ti:\3^ cards across 8 themes and a sprint-level workflow, while requiring Work Product Sheets with documented justification and follow-up review (Antikainen et al., 2021). RESOLVEDD similarly emphasizes a “paper trail” of ethical reasoning through its nine-step process (&&&3id:(Grohmann, 15 Jul 2025) OR ti:\3search_query3&&&). EthicAlly’s report schema functions analogously as an audit-friendly artifact.
3. Constitutional AI and normative scaffolding
The system’s alignment strategy is built around constitutional AI. Claude is described as being trained to follow a written “constitution” of ethical and safety principles, including human rights sources and bias safeguards, and to balance competing values through critique-and-revision (&&&3search_query3&&&). EthicAlly uses that capability to produce what the paper characterizes as structured ethics assessment that incorporates both “universal ethics principles and contextual and interpretive considerations relevant to most social science research” (&&&3search_query3&&&).
The prompt scaffold explicitly references named frameworks. Through iterative refinement, it was adjusted to reference the Belmont Report, the Declaration of Helsinki, and the Nuremberg Code, while also adding SSH contextual principles such as reflexivity, cultural sensitivity, political economy, trauma-informed practice, recognition versus anonymity, and intellectual property (&&&3search_query3&&&). The system prompt further instructs the model to consider sociocultural and political context and vulnerable or marginalized groups, to limit scope to SSH, to include a limitation disclaimer that output does not replace IRB or REC review, and to use professional, non-inflammatory language (&&&3search_query3&&&).
The collaborative prompt development methodology is itself part of the system’s design claim. The “startup prompt” was initially drafted by Claude in response to a meta-prompt from the author and then iteratively refined by the author and Claude (&&&3search_query3&&&). The stated rationale is that full scaffolding should be provided up front so the model has disciplinary, legal, and contextual frames from the outset, rather than relying on fragmented follow-up question-and-answer interaction (&&&3search_query3&&&).
This suggests that EthicAlly treats ethics support as a prompt-engineered reasoning pipeline rather than as a classifier or a recommender in the narrow sense. The paper provides no mathematical decision rule or learned utility function. Instead, the system is specified through a structured schema, a categorical risk rubric, and a guarded analysis framework embodied in the startup prompt (&&&3search_query3&&&). In this respect it differs from formal optimization approaches in ethically bounded AI that use CP-net distance thresholds or contextual bandit policy blending (Rossi et al., 2018), and from dialogue frameworks such as EthicMind that formulate ethical-emotional alignment as an explicit turn-level decision process with analyzer, planner, and generator mappings (&&&3 OR ti:\37&&&). EthicAlly’s normative machinery is textual, constitutional, and report-oriented.
4. Ethics assessment logic and operational categories
EthicAlly operationalizes ethics review by combining universal principles with SSH-specific considerations. On the universal side, the system checks respect for persons and autonomy, informed consent, beneficence and non-maleficence, justice and fairness, confidentiality and data protection, and conflicts of interest (&&&3search_query3&&&). On the SSH side, it extends the analysis to reflexivity, cultural sensitivity and representation, political economy, trauma-informed approaches, recognition versus anonymity, intellectual property, community impact and reciprocity, power dynamics, and vulnerable populations (&&&3search_query3&&&).
The output is not merely classificatory. It prioritizes issues by severity and recommends targeted mitigations. High-priority issues are described as “fundamental conflicts like dual-role coercion, unsafe contexts.” Moderate concerns include timing, privacy in site-specific settings, or data-management gaps. Minor considerations include language level or accessibility improvements (&&&3search_query3&&&). Guidance is also method-specific in several recurring domains: verbal versus written consent, iterative consent in ethnography or longitudinal work, interpreter-mediated consent, anonymization and pseudonymization, storage security, access control, retention periods, dissemination protections, researcher safety in hazardous or politically sensitive fieldwork, and debriefing or participant welfare (&&&3search_query3&&&).
One of the clearest examples is the clinician-researcher dual-role case described in the paper. In that scenario, EthicAlly detected a “fundamental conflict of interest undermining voluntariness,” flagged violations under Nuremberg and Belmont, recommended strict role separation, independent recruitment and interviewing, capacity checks, privacy safeguards, timing changes, and revised materials, and assigned a risk score of 5, characterized as “unapprovable without redesign” (&&&3search_query3&&&).
A concise way to interpret the system’s logic is that it maps SSH proposals into a hybrid space of principle-based and context-sensitive concerns. This is closely related to the modular vision in ethically bounded AI, where ethical constraints represent “priority orderings or acceptability boundaries independent of the user’s preferences” and are then combined contextually with goals and preferences (Rossi et al., 2018). EthicAlly does not formalize that combination with CP-nets or bandits, but it adopts the same underlying position that ethical guidance must be structured, contextual, and capable of trade-off analysis.
5. Evaluation, exemplars, and observed behavior
The paper’s initial evaluation uses 3 OR ti:\35 fictional SSH proposals generated with GPT-4o in order to avoid confidentiality and training-data leakage (&&&3search_query3&&&). Each proposal contained a non-obvious ethical issue, ranging from relatively simple oversights such as missing multilingual information to deceptive or politically charged proposals involving scientific racism, curtailing reproductive rights, conversion-therapy-like designs, or high-risk ethnography under political repression (&&&3search_query3&&&).
On this test set, EthicAlly “correctly identified the target issue in all but one case,” with the missed case described as a borderline interpretation of labor “incentivization” for homeless persons (&&&3search_query3&&&). It also “detected ‘gaming’ attempts where proposals tried to hide unethical intents and recommended strong safeguards,” and it was reported to handle complex, competing frameworks such as EU border ethnography while supporting alternative consent models including verbal, interpreter-supported, trauma-informed, and multi-stage check-in approaches (&&&3search_query3&&&).
The paper also reports informal expert testing by research ethics professionals. Those impressions were described as positive, and one professional reported that the tool supported their reasoning in a contested committee case (&&&3search_query3&&&). At the same time, the evaluation remains explicitly preliminary: larger systematic validation, human-only comparisons, and multilingual testing are described as future work rather than completed evidence (&&&3search_query3&&&).
These findings place EthicAlly closer to a structured ethics drafting assistant than to a benchmarked decision model. It is notable that the paper does not report formal sensitivity, specificity, calibration, or inter-rater statistics for the prototype itself. That omission is consistent with its presentation as a proof of concept. It also distinguishes EthicAlly from recent inference-time ethical control frameworks such as EthicMind, which report automatic judge scores, human preference data, and ablations over analysis, strategy, and response modules (&&&3 OR ti:\37&&&). EthicAlly’s evidential basis is instead qualitative and case-based.
6. Governance, limitations, and relation to adjacent research
EthicAlly is explicitly scoped as supplementary rather than authoritative. The interface and report include prominent disclaimers that the system does not replace IRB or REC review (&&&3search_query3&&&). It is designed to decline clinical research reviews, and users are instructed to remove personally identifiable information and acknowledge that processing occurs outside the European Union (&&&3search_query3&&&). The paper states that the app and Anthropic do not retain data after response generation, according to the stated policy (&&&3search_query3&&&).
The prototype’s limitations are direct and extensive. The model remains proprietary and therefore a “black box”; residual hallucination and bias risk are acknowledged; outputs are non-deterministic; overreliance by inexperienced users is possible; adversarial use to “game” REC language remains a concern; very large multi-site protocols may exceed token or context limits; and current performance in languages beyond English and German requires testing (&&&3search_query3&&&). The paper also notes that some disciplines or RECs may disagree with the tool’s cautious stance on vulnerability or with its interpretations in politically sensitive contexts (&&&3search_query3&&&).
Future directions are correspondingly institutional as much as technical. The paper identifies an open-source path once open models reach comparable reasoning performance; institution-specific retrieval-augmented generation over local ethics manuals and policies; chat or explanatory modes; integration with scholarly literature APIs; and large-scale multilingual evaluation (&&&3search_query3&&&). Institutional recommendations are equally clear: AI ethics support should remain preparatory infrastructure rather than a decision system; human oversight and accountability at RECs must be preserved; privacy and security require GDPR-compliant design; and deployments should be especially valuable in under-resourced settings and LMICs (&&&3search_query3&&&).
In the broader ethically aligned AI literature, this positioning is distinctive. ECCOLA extends ethics into planning, prioritization, trigger-based reassessment, and a lifecycle “situational picture” with token-based relevance metrics and Likert-scored coverage (Antikainen et al., 2021). RESOLVEDD emphasizes decision logging, transparency, and responsibility but finds that accountability requires additional organizational scaffolding (&&&3id:(Grohmann, 15 Jul 2025) OR ti:\3search_query3&&&). Ethically bounded AI proposes modular combination of goals, preferences, and ethical principles, together with contextual weighting and compositional reasoning across interconnected systems (Rossi et al., 2018). EthicAlly does not implement these formal devices, but it shares their central design intuition: ethics support becomes operational when principles are turned into structured prompts, explicit artifacts, traceable recommendations, and governance hooks.
Taken in that sense, EthicAlly is best understood as a domain-specific instantiation of ethically aligned AI for SSH research design. Its novelty lies less in inventing a new theory of ethics than in assembling constitutional AI, collaborative prompt development, named normative frameworks, and SSH-specific contextual principles into a usable review-preparation workflow (&&&3search_query3&&&). What remains unresolved is the extent to which such systems can be validated across jurisdictions, cultures, languages, and disciplinary subfields without collapsing the very contextual sensitivity they are intended to preserve.