How is CDA-AMC getting involved with AI in HTA in Canada and beyond?

Recognizing that the use of AI in HTA is more than simply a local issue, CDA-AMC has been proactive in understanding, adopting, and creating leading guidance on its implementation to help support the international HTA community (Reference Ting, Badin and Gaind1;Reference Gaind, Badin and Jacob2). For HTA agencies, international collaborations in the AI space are increasingly important to address the fast-moving environment and the growing need for guidance on how AI is being used in HTA. At CDA-AMC, we are working to shape the appropriate use of AI through participating in forums like the International Network of Agencies for Health Technology Assessment (INAHTA), collaborating with other international agencies, and sharing our experiences, tools, and initiatives (e.g., CDA-AMC AI-backed HTA information retrieval search tool).

What CDA-AMC projects support the appropriate use of AI in HTA?

Currently, CDA-AMC has three major lines of work to inform appropriate AI implementation practices:

• • Governance: Since 2024, CDA-AMC has been developing formal guidance (6) on AI methods in HTA, working together with NICE and adapting their Use of AI in Evidence Generation position statement (7). This work is being extended to address how AI-generated real-world evidence (RWE) should be reported for the Canadian context (8). Additionally, we established an AI Risk and Governance Framework to classify AI use cases into risk tiers and apply governance accordingly. Pilot projects will explore workflows and functions where automation and AI can enhance efficiency while maintaining accuracy and rigor.

• • Methods: When AI methods are used to generate evidence in support of an HTA submission by a technology developer, CDA-AMC expects them to uphold the core HTA principles of transparency, appropriateness, trustworthiness, and ethical considerations. As a practical example, this is as straightforward as clearly declaring that AI was used within submission materials. CDA-AMC also follows these principles for our own internal AI methods, processes, and procedures.

• • Community collaboration: As the international HTA community is developing additional guidance, CDA-AMC continues to collaborate with other agency partners. For instance, together with IQWiQ in Germany, our information specialist team is working on new appraisal tools for AI-generated data and ensuring that the right evidence can be found by searching gray literature with AI tools (9). As part of the HTAi Global Policy Forum (GPF), and in collaboration with industry, we are ensuring that implementation guidance developed jointly across jurisdictions reflects what agencies actually need from technology submitters.

What are the future developments for AI in HTA?

The work that the CDA-AMC is doing is clearly critical to support today’s HTA processes. For us at Loon, as an early-adopting AI developer, my view looks a bit further into the horizon of future developments for AI in HTA. I believe it is less about the technology itself and more about how the technology is used. The most consequential development in AI for HTA will be the adoption of an “engineering philosophy” where the architecture itself controls the output, ensuring scientific rigor by design in the way that curved roads passively slow drivers. As generative AI capabilities accelerate, the temptation is to deploy large language models (LLMs) directly into HTA workflows through open-ended chat interfaces. But this is prone to high error rates and large variability in outputs dependent on user skill and technical fluency. The systems most likely to earn trust are those where AI is embedded, executing specific, well-defined tasks (e.g., eligibility assessment, data extraction, and study classification) without exposing users to unpredictable open-ended interaction. A system that performs identically, whether operated by a senior methodologist or a junior analyst, can be reliably measured and trusted.

On the topic of building trust in AI, we break HTA workflows into granular tasks, engineer purpose-built agents, then prospectively benchmark against held-out datasets following strict machine-learning principles (e.g., all test data outside the training corpus). Retrospective validations against studies already in the training corpus are of limited value: when models already know the answers, the validation is corrupted. We evaluate correctness of outputs and the system’s ability to recognize when it is uncertain (calibrated confidence). Once deployed, we run continuous validation, triggered automatically whenever a material change occurs (e.g., a new underlying model, an updated system prompt, and a change in data handling).

Are there new technical developments in AI for HTA?

The most promising technical direction is “neurosymbolic” AI: the combination of the “neural” layer, typically LLMs, with a “symbolic” layer of deterministic rules, logic, and algorithmic processes. The neural layer understands language, interprets context, and handles the variability of scientific reporting. The symbolic layer enforces structure, consistency, and truthfulness. Together, they address a fundamental tension: generative models are extraordinarily capable but inherently stochastic, but HTA demands rigor. Wrapping stochastic processes in deterministic guardrails allowed us to build confidence scores that correspond directly to human-AI disagreement rates. Our organization independently validated this neurosymbolic approach on the task of citation screening in systematic reviews. Loon neurosymbolic AI screened 3,796 citations from 8 publicly posted CDA-AMC systematic reviews, where each AI decision carried a confidence score that was assigned through a deterministic algorithm. This allowed human-in-the-loop validation to focus on lower-confidence “Include” AI decisions (5.8 percent of the sample). The workflow yielded 99 percent sensitivity and 90 percent precision, reducing screening workload by over 95 percent (Reference Janoudi, Uzun and Disher10).

Are we missing anything?

We do not need infallible AI. We need AI that knows when it is in doubt. This reframes the question from “is AI accurate enough?” to a value-of-information problem the community already understands: how much confidence can we place in a given source, and how should we act on it? The more precisely we measure AI confidence, and the more rigorously we manage the environment in which it operates, the more useful these systems become, not as replacements for human judgment, but as instruments whose uncertainty is quantified, auditable, and actionable.

What are the risks you see in AI development?

AI development is progressing at an unprecedented rate. AI systems can now reason and problem-solve across text, images, audio, and video within a single platform and conduct independent agentic work, presenting enormous opportunities to all of science (11), including HTA, where AI may soon be capable of automating research end-to-end (Reference Lu, Lange and Yamada12).

However, this rapid advancement in AI development has significantly outpaced the speed of AI governance. New versions of AI models are released frequently without adequate testing or safeguards. Easy access to powerful models is shifting the types of AI used from dedicated single-purpose models to general-purpose LLMs, and we are starting to see risks of harm associated with them unfolding in real time, which should concern users in all sectors, including HTA.

Given the risks to the community, what type of AI is trustworthy enough for use in HTA?

Similar to what Ghayath mentioned earlier, small, task-specific AI, also known as “narrow AI,” trained on carefully curated data, is not only efficient to develop and run but also easier to test and validate using specific benchmarks and therefore safer to use. While general-purpose LLMs, trained on massive amounts of diverse data, are ubiquitous, the HTA community should consider using narrow AI, specifically intended and evaluated for performance in HTA, to mitigate the risks and build public trust in the HTA process.

Is there still a need for humans in HTA? And what does appropriate human oversight look like?

As AI models become more capable, our use of and reliance on them will grow, and so will our automation bias. This could challenge meaningful human involvement, so the HTA community will need to define what appropriate human oversight is. For now, the use of AI should be limited to specific tasks humans can oversee within existing HTA workflows. For example, AI could be used to extract data during evidence synthesis or build economic models, with humans checking the work as second reviewers. Going forward, as new state-of-the-art AI models emerge, how to handle AI capabilities that enable end-to-end research or surpass human abilities will be an important conversation within agencies, technology developers, and others in the HTA community.

From industry’s lens, any key takeaways from the previous panelists on AI in HTA?

All of the panelists’ perspectives have a common theme of trust-building when it comes to AI in HTA. Through the shared development of guidance documents, design of validation methods, and definitions of appropriate human involvement, trust-building is the next crucial step for AI in HTA.

We come back to a key HTAi GPF recommendation: the development of CoPs to allow networking and sharing of experiences with specific AI tools and use cases. CoPs are critical to safely implementing AI and unlocking its benefits for HTA, and HTAi can lead in facilitating these communities.

As HTAi has a broad view of what a CoP may look like in different geographies and healthcare systems, HTAi can share and promote best practices across its membership. CoPs may be situation-dependent, but we believe that in all cases, their main objective is to maintain trust, transparency, and human agency in HTA. As per the GPF’s suggestion for an inclusive and purposeful CoP, these groups should have open membership, including HTAi members and others within the HTA ecosystem.

From an industry perspective, we must consider specifically where it makes sense for AI to be deployed in HTA, and avoid using AI just for the sake of using it. In other words, use of AI should be intentional, in service of the HTA, and not the other way around.

It is critical to recognize that industry has a valuable role in CoPs. Industry partners can efficiently automate administrative processes and alleviate HTA agency workflow burden, help co-create validation methods for AI-augmented real-world evidence, and show where AI was used (or not) within submissions so HTA can be conducted transparently. In all cases, CoPs can build trust and best practices so that AI can be appropriately adopted for HTA.