No Walkovers - Humans Are Irreplaceable In AI-Assisted Human Rights Impact Assessments

Human rights impact assessments (HRIAs) are valuable tools that help identify, assess and address actual and potential adverse human rights impacts arising from activities, policies or practices, supporting compliance with human rights law and standards such as the UN Guiding Principles on Business and Human Rights . HRIAs may be used to assess for various reasons e.g., corporate supply chain operations, technology in a specific industry, or how laws might affect people .

Human rights practitioners, consultants or businesses carrying out HRIAs may use AI tools such as ChatGPT, Claude, Gemini, Perplexity or other specialised ones in the HRIA process for planning, data collection and records scanning, data review and analysis, stakeholder mapping and management, predicting model scenarios in impact evaluations, content customisation, report formatting and presentation. The benefits of using AI tools include filling gaps in assessor knowledge, enhanced coverage of impacts, identification of evidentiary gaps, time and cost efficiencies and resource optimisation.

But, while these tools facilitate HRIAs, they also present various challenges, which if not addressed will undermine the objective of HRIAs and their processes. A fundamental concern remains whether organisations should delegate and deploy AI tools that lack human judgment to carry out or support work in HRIAs and to what extent.

AI tools are fallible and limited

AI tools lack contextual awareness to navigate the complex ethical, legal, or political situations assessed in HRIAs. One study, for example, showed how values such as well-being, justice and human/animal rights are least represented in Reinforcement Learning From Human Feedback (RLHF) datasets and that this impacts how AI might approach social issues .

When used to identify impacts, AI may fail to flag previously undocumented human rights impacts. AI tools cannot yet explain all human experiences, unintended consequences, cultural, linguistic shifts, or complex social factors and actions. AI tools cannot effectively fact check all harms to vulnerable communities and groups, especially those digitally excluded. They may struggle with harms that are contextual, cumulative, subjective, relational, or experienced differently across communities and individuals.

Furthermore, the propensity of AI tools to hallucinate or fabricate false data may result in the provision of inappropriate or harmful safety-related information that affects the outcomes of the HRIA.

An HRIA requires nuanced human judgment. AI tools cannot and should not determine which rightsholder's or stakeholder's concerns matter most and should be prioritised; these judgments require human expertise and accountability. AI-generated summaries cannot be treated as a complete representation of affected rightsholder perspectives. Using automated sentiment analysis as a substitute for substantive review or excluding vulnerable group viewpoints because these are difficult or appear statistically insignificant undermines the value and the credibility of the HRIA. As outlined by the International Association for Impact Assessment, HRIAs call for a "finer disaggregation of data of those impacted (e.g., by gender, ethnicity, age, contractual relationship, etc.) to better understand how impacts affect them and to identify whether discrimination takes place."

Given these challenges and even though AI assistance and use in HRIAs is useful, meaningful human oversight is critical to ensure such HRIAs are fit for purpose and remain credible.

What is meaningful human oversight in this context?

HRIAs exist to protect human rights. Therefore, meaningful human oversight must be commensurate with the scope and scale of the HRIA. Meaningful human oversight isn't about a human assessor using the right prompts to generate information, predictions or recommendations. Nor is it just about a reviewer being part of the HRIA process or signing-off on AI-generated outputs with cursory or casual scrutiny. It requires a human-in-command approach: humans are responsible and accountable for the HRIA, define its scope and criteria, and intervene, challenge and validate AI outputs throughout the HRIA process and can ensure the assessment is an iterative process. In this case, human-in-the-loop simply does not work well, as it would undermine accountability. A human-in-command model is preferable to approaches in which human involvement is limited to reviewing AI outputs, as the final responsibility and accountability should always remain with human assessors.

What are its critical elements?

Meaningful human oversight is operational in practice. It is exercised to a good degree and quality. It does not just exist in theory or on paper as a compliance indicator or a token gesture. A human should be designated, responsible and involved in every stage and workflow of the HRIA process - be it planning and scoping, data collection, analysing impacts, stakeholder engagement, mitigation and management and reporting and evaluation.

Meaningful oversight requires the right expertise - the expert should be well-versed in human rights law, frameworks, standards and protections. They should be able to identify any potential adverse human rights impacts and additional factors that are not obvious from a technical or documented perspective. The reviewer or overseer should properly understand the scope and limitations of the AI tools used to support the HRIA, their analysis, prediction, recommendation or decision logic, and be capable of correctly interpreting and critically evaluating all AI-assisted outputs.

Reviewers must also have the ability, resources and influence to challenge, contest, intervene, disregard, override, reject or reverse an AI system's outputs, predictions or recommendations. They should test and validate AI-generated human rights analyses and scenarios against original research, stakeholder input or their own expert knowledge.

Meaningful human oversight also requires assessor self-awareness of acceptance bias - the tendency of automatically relying, over-relying, or rubber-stamping (if it looks plausible) the HRIA output produced by the AI tools without appropriate verification.

Meaningful human oversight, furthermore, is not 'one and done' - it is ongoing active review during and after the HRIA process for several reasons: older AI tools used might have used old data or provided recommendations or predictions that have become inaccurate over time. Emergent societal human rights norms and expectations might have shifted and not have been captured at the point that the AI-assisted assessment was carried out.

Assessors or overseers should ensure legal due diligence by documenting which AI tools were used for what purpose, their version, which human rights impacts were identified by the AI and who reviewed these along with their findings (including whether the reviewers/overseers agreed or disagreed with the AI, their evidence and final decisions, i.e., keep/override), the outputs of stakeholder consultations, mitigations adopted, and changes made to the use of AI tools in the HRIA process. Good documentation facilitates effective audits, and these are critical if a human rights complaint, regulatory inquiry, or lawsuit arises.

Finally, the use of AI-tools in HRIAs must undergo periodic evaluations to show if there was a review of whether the human oversight in place is working. Reviews check whether impacts were correctly identified, frequency of overrides, false-positive and false-negative rates, rightsholders' complaints, and remediation outcomes.

Future-fit for purpose

AI-assisted HRIAs will become standard practice, but they need to be fit for purpose in every context of use and must be carried out with meaningful human oversight. This oversight requires substantive, ongoing expert engagement, and cannot be an optional extra or performative exercise for compliance purposes. If human oversight is weak or missing in an AI-assisted HRIA, it dilutes an HRIA's relevance, undermines its legitimacy, accountability and trustworthiness and calls into question its value as a safeguard for human rights.

Author

Rowena Rodrigues, PhD, is Head of the Innovation and Research Services-Strategic Partnerships at Trilateral Research with expertise in the impact, governance and policy dimensions of new and emerging technologies, with a strong focus on artificial intelligence. She is co-author of Ethics of Artificial Intelligence: Case Studies and Options for Addressing Ethical Challenges and co-editor of Privacy and Data Protection Seals. She has published widely in leading journals and is well-cited in the areas of AI , impact assessment, and ethics in research & innovation.

This Insight is based on and inspired by the the Business and Human Rights Forum roundtable held on 10 June 2026.