AI in Impact Assessment: What Are Regulators Actually Doing?

Artificial intelligence is moving rapidly into professional practice, and impact assessment is no exception. Much of the discussion to date, however, has focused on how consultants, practitioners and project proponents might use AI to prepare reports, analyse information or improve productivity. A different — and arguably more consequential — question is receiving less attention: what are the regulators and statutory impact assessment authorities themselves actually doing with AI?

I have been examining this question through a comparative review of developments across Canada and a range of international jurisdictions, including the United States, Hong Kong, Chile, England, Denmark and Australia. The emerging picture is more nuanced than the broader discussion about AI might suggest. AI is entering impact assessment, but it is not transforming every part of the process equally. Governments are experimenting with very different applications, governance arrangements and levels of automation. And, importantly, the evidence demonstrating that these technologies actually improve assessment outcomes remains limited.

The professional context is worth noting. The International Association for Impact Assessment devoted its 2025 annual conference in Bologna to impact assessment in the age of artificial intelligence, drawing more than a thousand delegates from over eighty countries, and adopted a set of Principles for Use of AI in IA. Its journal, Impact Assessment and Project Appraisal, published a dedicated issue on the subject later that year. The profession is engaging seriously with the question. What has been examined less closely is what assessment authorities have actually built.

AI is entering the documentary layer first

Perhaps the clearest pattern is where AI is actually being deployed. Most confirmed applications are concentrated in what might be called the documentary layer of impact assessment: searching large collections of previous assessments, retrieving relevant information from lengthy documents, extracting tables, maps and figures, comparing previous decisions and mitigation measures, summarising comments and submissions, organising environmental information, and supporting document review and drafting.

These are important functions. Impact assessment has become extraordinarily information-intensive, and regulators routinely deal with thousands of pages of technical material, public submissions, baseline studies and historical records. AI may therefore offer considerable value simply by making existing information easier to find and use.

But this is different from automating the substantive analytical work of impact assessment. There is much less evidence of operational AI systems being used to predict environmental or social effects, compare project alternatives, determine the significance of effects, evaluate cumulative effects, select mitigation measures, or make statutory assessment decisions.

That distinction matters. A system that reduces the time needed to locate information may improve administrative efficiency, but that does not necessarily mean the quality of the assessment itself has improved.

Different jurisdictions are taking very different approaches

The international experience shows there is no single model of AI adoption.

Hong Kong: building the information environment upstream

Hong Kong has developed one of the most advanced examples identified in the review. Its Environmental Protection Department has created a smart environmental assessment platform designed to provide environmental data and analytical tools earlier in the assessment process. The conceptual shift is important: rather than waiting for proponents to submit an assessment and then reviewing the information, the regulator increasingly provides data and tools that proponents can use while preparing the assessment.

Hong Kong has reported substantial reductions in assessment timelines. However, these improvements occurred alongside broader regulatory and procedural reforms, making it difficult to determine how much of the improvement can be attributed specifically to AI and digital tools. This is a recurring challenge when evaluating AI performance.

United States: searching decades of assessment records

In the United States, the PermitAI initiative developed by Pacific Northwest National Laboratory is applying AI to the enormous body of documentation generated under the National Environmental Policy Act. Tools are being tested to search, interrogate and analyse historical NEPA documentation. This illustrates one of AI's most obvious opportunities in impact assessment: making decades of previous assessments accessible as a usable knowledge base rather than simply an archive of PDFs.

Denmark: curating the data before building the tool

Denmark offers the clearest illustration of a principle I return to below. Rather than applying AI to whatever documentation happened to be available, researchers associated with the Danish EIA Centre have worked to build a deliberately curated dataset of environmental assessment texts, addressing questions of quality assurance, structure, copyright and ownership, ethics and continuous updating. The premise is that the usefulness of AI in assessment depends on the quality and governance of the underlying information — and that this cannot be assumed.

Canada: governance ahead of deployment

Canada presents a particularly interesting case, because the sequence runs opposite to most jurisdictions. The federal government already has a relatively developed governance framework for automated decision-making through the Treasury Board Directive on Automated Decision-Making and its associated Algorithmic Impact Assessment. A system supporting statutory assessment decisions would likely fall within its higher tiers, attracting requirements for human decision-making, senior approval and independent peer review before deployment.

Meanwhile, other Canadian regulators have developed important information-management tools. The Canada Energy Regulator's BERDI initiative makes environmental and socio-economic information contained in historical pipeline assessments more accessible and searchable, using natural language processing and machine learning to extract tables, figures and maps from assessment documents. Notably, the same approach identifies and protects sensitive Indigenous knowledge so that it is excluded from public search results — an aspect of that work which deserves considerably more international attention than it has received.

Canada therefore illustrates an unusual configuration: comparatively mature governance of automated decision-making, meaningful data infrastructure at some resource regulators, and limited deployment within the federal assessment authority itself.

Chile: combining deployment and governance

Chile is another noteworthy example, because technological modernisation has been accompanied by the development of an internal AI governance protocol addressing privacy, bias, explainability, procurement and human oversight. This is significant because many institutions are experimenting with AI first and developing governance later. Chile's approach suggests a different model, in which those questions are addressed alongside deployment rather than after it.

Not every digital reform is AI

One of the most important lessons from the comparative review is that digitisation, automation and artificial intelligence should not be treated as interchangeable concepts.

Many assessment systems are undergoing major digital reform. Governments are creating electronic permitting portals, standardised data requirements, regional environmental databases, automated workflow systems and machine-readable environmental information. These developments may create the conditions needed for future AI systems, but they are not necessarily AI themselves.

Australia, for example, is undertaking major environmental assessment and data reforms that could provide an important foundation for future digital and AI applications. Similarly, regional databases being developed in northern Canada may significantly improve cumulative-effects assessment without currently relying on artificial intelligence. This distinction matters, because otherwise we risk overstating how widely AI has actually been deployed.

Governance is developing in two directions

Another important finding concerns who is being governed. There are at least two distinct AI governance challenges in impact assessment, and they are frequently conflated.

Governing AI used by participants. Proponents, consultants, objectors and members of the public can increasingly use generative AI to prepare material submitted to assessment authorities. England's Planning Inspectorate has issued guidance expecting disclosure where AI has been used substantively in evidence submitted to it, identifying the tool used and confirming that the submitter takes responsibility for the accuracy of the content. Improper use may be treated as unreasonable behaviour, with potential cost consequences. This raises important questions. If AI generates analysis or evidence, who is responsible for verifying it? What level of disclosure should be required? Should AI-generated material be treated differently from conventional consultant-prepared evidence?

Governing AI used by the regulator. A different set of issues arises when the assessment authority itself uses AI: procurement standards, algorithmic bias, transparency, privacy, protection of confidential or Indigenous information, explainability, human oversight, and accountability for statutory decisions.

These two directions of governance — outward toward participants and inward toward government institutions — need to be considered separately. A mature AI governance framework for impact assessment will ultimately need to address both.

Where is the evidence that AI improves assessment?

Perhaps the most important question is also the most difficult: does AI actually improve impact assessment? At present, the evidence base remains surprisingly thin.

Governments have reported substantial efficiency improvements associated with some digital and AI-enabled systems. But many of these claims are self-reported. In addition, AI deployment often occurs at the same time as regulatory reform, streamlined approval procedures, improved data systems, additional staffing, revised assessment requirements or broader digital modernisation. This makes attribution difficult. If assessment timelines fall after AI is introduced, was the improvement caused by AI, by regulatory simplification, by better data — or by all three?

Independent evaluation remains limited. One controlled academic study is particularly instructive. Cilliers and colleagues, writing in Impact Assessment and Project Appraisal in 2025, built a customised generative AI model to perform environmental screening under South Africa's list-based system and tested it against real assessment cases. Performance improved substantially across successive refinements of the model. The authors nonetheless concluded that it could not be used to screen reliably on its own, and reported reaching a point at which further development and verification would consume more time than the screening work it was intended to replace.

The lesson is not that AI cannot work. The lesson is that performance has to be demonstrated rather than assumed.

Impact assessment should apply its own standards to AI

This leads to a broader professional issue. Impact assessment exists in large part to test claims. When proponents predict that a project will have limited environmental effects, regulators normally ask for evidence: baseline data, clear methodologies, predictions, uncertainty analysis, mitigation measures, monitoring, follow-up and verification.

Yet institutional adoption of AI has not always been subjected to the same degree of scrutiny. That should change. Before governments make consequential use of AI in impact assessment, they should be able to answer some basic questions. What problem is the technology intended to solve? What is the baseline performance of the existing process? What improvement is expected? How will accuracy be measured? What kinds of errors occur, and who reviews them? How is human judgment retained? And how will the technology be evaluated after implementation?

The principles are familiar, because they are already part of good impact assessment practice.

Where AI may ultimately have the greatest value

The current concentration on document handling does not mean AI's future role will remain limited to administrative efficiency. There are potentially much more significant applications.

One is cumulative effects assessment. Cumulative effects require integrating information across projects, proponents, geographic areas and long periods of time — exactly the kind of information-integration problem where advanced computational tools may eventually offer substantial value. Other possibilities include analysing large monitoring datasets, detecting changes in environmental indicators, comparing alternatives, identifying relationships across historical assessments, improving regional and strategic assessments, supporting adaptive management, and connecting follow-up results back to future impact predictions.

But these applications will depend on something less glamorous than AI itself: good data. If previous assessments, monitoring results and baseline information are fragmented, inconsistent or inaccessible, adding AI will not solve the underlying problem. In many jurisdictions, the first step toward useful AI may therefore be building better environmental information systems.

A measured path forward

There is little doubt that AI will become increasingly important in impact assessment. The more useful question is not whether the profession should use AI, but how it should use it responsibly — and how we will know whether it is working.

The emerging international experience suggests several practical principles. Build the information architecture before the model. Distinguish AI from conventional automation and digitisation. Govern both institutional AI and AI-generated evidence submitted by others. Separate technological effects from broader process reform. And, most importantly, apply the evidentiary standards of impact assessment to AI itself: establish the baseline, define the claimed benefit, isolate the effect, test the failure modes, monitor the outcome, and publish the results.

Impact assessment has spent decades developing methods for evaluating uncertain claims about the consequences of new technologies. AI should not be an exception.