Digital mental health is no longer an emerging field. Increases in teletherapy platforms and providers have increased many people’s comfort and knowledge with receiving services remotely. It might already be the case that more people are using chatbots than therapists for mental health support [1]. What started as a field based on computerized programs has expanded to include apps, virtual reality, social media, digital phenotyping, rule-based chatbots, and large language models (LLMs). Many health systems and public entities are exploring the use of digital mental health to more effectively meet people’s needs [2,3]. This shift raises the standard for the research we publish. Although establishing the feasibility of digital mental health interventions previously made a useful contribution, similar feasibility studies may fail to further advance the field. The field now needs more evidence that can guide clinical care, inform policy, protect patients, and identify which technologies work, for whom, under what conditions, and at what risk [4-6].

In 2023, JMIR Mental Health outlined priorities for digital mental health research centered on equity, replicability, privacy, efficacy, and engagement [7]. Those priorities remain. What has changed is the speed, scale, and consequence of the technologies now entering mental health care, especially with the emergence of LLMs. Although LLMs are the most visible example of the current methodological challenge, the standards described here apply across digital mental health. Given the recent volume of submissions on LLMs, better defining the scope of these papers will aid authors wishing to disseminate their work and provide guidance to the field on which areas of inquiry we see as most generative.

The challenges of LLMs moving faster than the research base have been well documented across the mental health field, including cases of patient harm, ongoing lawsuits, and expanding efforts to restrict or ban some tools for care. This is the bind David Collingridge [8] described decades ago: by the time evidence accumulates about a new technology, it is often too embedded in practice, regulation, or use to redirect easily. The question facing digital mental health research is no longer whether the field can produce evidence, but whether the evidence it produces is rigorous, generalizable, timely, and useful enough to inform how these technologies are evaluated and used to improve care.

Artificial intelligence (AI) is itself part of how that evidence is now produced. Authors today may use LLMs to draft, summarize, code, and even analyze. The volume of research on AI in mental health has risen sharply, and so has concern that the quality of that research has not kept pace. This is the paradox of the current moment: the same tools that accelerate research are also raising questions about its soundness. This is why research standards matter more than ever, and why JMIR Mental Health seeks to advance the most rigorous and useful research with new priorities, as summarized in Figure 1. We are seeking research that helps the field learn something durable, especially given how rapidly the field changes.

For digital mental health work in general, and for LLM-related work in particular, the strongest studies are designed around the phenomenon being investigated rather than only the specific tool being tested, which is aligned with the National Institute of Mental Health’s experimental therapeutics approach [9]. Submissions in this space are most useful when they meet several standards.

Much of this editorial focuses on LLMs because they raise questions for the field that are especially acute. The journal’s interests, however, remain broader, and an LLM is often not the right, or only, tool for mental health. We continue to welcome rigorous work on mental health apps, virtual and augmented reality, rule-based chatbots, digital phenotyping, online interventions, remote monitoring, digital therapeutics, social media, clinical prediction, hybrid care models, and other topics. The standards described here apply across these areas: submissions should be replicable, clinically meaningful, attentive to equity and engagement, and grounded in appropriate comparisons.

Engagement remains one of the field’s foremost challenges [28,29]. We are thus interested in work that moves beyond reporting low engagement (or perhaps in some cases, too much engagement with AI or social media) toward examining how engagement matters on a personal level. Does greater use predict better outcomes, or does it simply reflect higher baseline need? Is there a dose-response relationship, and how do we even measure dose? Should we measure minutes, sessions, completed modules, conversational turns, therapeutic content, or human contact?

With LLMs, the engagement question is now whether it is too little, as was often the case with smartphone apps. Recent passive-sensing data from over 6000 US youth found that median daily use of generative AI apps was under a minute, while a small subset engaged for more than 40 minutes a day [30]. This pattern suggests that average exposure can conceal very different individual experiences. Cases of compulsive use and parasocial attachment raise empirical questions about what too much engagement looks like, who is vulnerable to it, and what consequences may follow.

Implementation deserves more attention. The most clinically useful tools are not always the most technically advanced ones. Research showing how a technology fits a setting, workflow, payment model, workforce, or population is crucial. LLMs raise new challenges in implementation, design (eg, accessibility and access to data), and output (eg, bias and correctness) [31]. Real-world implementation is shaped by infrastructure, cost, training, reliability, privacy, and legal accountability as much as by underlying capability [32]. For example, recent research on the current generation of AI scribes has suggested that they may actually save negligible time, [33] and research examining the costs of running LLMs for medical coding suggests the cost may be far higher than expected [34]. Research that takes those constraints seriously rather than designing around them is especially productive.

Theories, frameworks, and models from implementation science might support structuring this research in ways to support learnings across projects. It is also worth considering the extent to which these theories, frameworks, and models apply to digital mental health and what adaptations might be necessary [35]. Given the established evidence base for many digital mental health interventions, better understanding how to successfully integrate them into routine care settings is an opportunity for future contributions.

As implementation increases, data generated from real-world settings has become more common. Although these investigations overcome the challenges posed by conducting clinical trials in artificial environments that do not mimic real-world deployments, they face limitations that may limit the insights they can offer the field. When used, real-world data need to be appropriate for the proposed research question, including aspects such as the size and representativeness of the sample. Information about engagement, outcome, covariates, and potential confounders is always critical. In many examples of real-world data, it may not be possible to use as rigorous a comparison of conditions as might be expected in clinical trial settings; however, this does not mean that other considerations of rigor and reproducibility are not relevant. Studies that merely examine convenience samples from deployed interventions may contain a degree of bias, and any conclusions may raise more questions than answers.

A few categories of submission are unlikely to fit, less because they lack effort than because they no longer move the field forward without a stronger justification. Studies whose main finding is that people can use a piece of technology or find it acceptable are rarely sufficient in 2026. Cross-sectional surveys of technology use, particularly surveys measuring generic interest in AI for mental health, have been well covered. Research focused on older AI models that are no longer in use is generally of historical rather than practical interest.

Effect sizes for digital mental health interventions have also been well characterized. A meta-analysis of 176 randomized trials placed app-based effects for depression and anxiety in the small-to-modest range (Hedges g was approximately 0.26 to 0.28 for both depression and anxiety) [36], and a recent meta-analysis of 39 AI chatbot trials for depression and anxiety reported nearly identical estimates (Hedges g ranged from 0.28 to 0.31) [37]. While these effect sizes will vary based on human support, follow-up duration, clinical severity, and other factors, they at least offer a baseline to orient new findings. Papers reporting substantially larger effects are welcome, but the rigor of evidence behind them should match the size of the claim. Finally, work raising ethical concerns, including but not limited to the use of patient data without appropriate permissions, will not be considered regardless of the strength of the underlying findings.

The profound impact of generative AI tools in scholarly writing is now well known—the same tools that are shaping how digital mental health research is conducted are also driving how related scholarly work is produced [38]. These tools introduce risks spanning factual accuracy, fabricated content, hallucinated references, copyright infringement, and ethical breaches. As a result, manuscripts that do not address AI assistance at all are increasingly difficult to interpret both scientifically and ethically, challenging editors, reviewers, and readers. Authors should therefore report how and when generative AI was used in the research or writing process, specifying the tools, the nature of the AI-generated content, and, where relevant, the prompts that shaped it [39], per guidelines developed by JMIR leadership. We also appreciate and expect that many new research papers will use AI in a larger role and welcome that when fully disclosed.

The potential of digital mental health has always been that technology can expand access to care while improving outcomes that matter to patients, clinicians, and communities. Today that promise is more visible, more commercially active, and more clinically consequential than ever. But promise is not evidence, and excitement is not a substitute for rigorous research, as earlier instances of enthusiasm around apps, virtual reality, and other digital tools have shown. The window Collingridge [8] described has not closed, but it is narrower than it was when JMIR Mental Health published its 2023 editorial guidance.

The Society of Digital Psychiatry, in partnership with the journal, has set out a roadmap for translating research into practice through education, AI standards, and digital navigators [40]. The publication standards described here are the upstream complement: research must be designed to be worth translating. Whether the next 3 years bear out the promise of digital mental health will depend less on the tools than on the evidence that shapes their use.