AI meets virtual reality therapy: new review maps what works, what’s missing, and what’s next

A systematic review in Nature’s Translational Psychiatry charts how artificial intelligence is being woven into virtual reality exposure therapy (VRET) to treat fear- and anxiety-related disorders—and where the science needs to go next.

The open‑access paper, “Artificial intelligence (AI) for virtual reality exposure therapy (VRET): A systematic review,” by Kamilla Bergsnev and Ana Luisa Sánchez Laws (UiT The Arctic University of Norway), evaluates how AI is being used to personalize VR-based exposure, support clinicians, and engage patients. The bottom line: early signals are promising, but much of the work remains proof‑of‑concept and too often excludes the very people it aims to help.

Why this matters

• Exposure therapy is a gold‑standard treatment for phobias, PTSD, OCD, and related conditions, but it’s underused in routine care due to logistical hurdles and clinician/patient concerns.

• VRET can make exposure safer and more practical by simulating feared situations. AI could take it further—tailoring intensity, pacing, and context in real time, and scaling support to more patients.

• Getting that promise right requires tight alignment with exposure/extinction theory and strong “human‑in‑the‑loop” safeguards.

How the review was done

• Scope: Peer‑reviewed studies through November 14, 2025; 377 records screened; 23 studies included.

• Databases: PsycINFO, Web of Science, Google Scholar, EMBASE, CINAHL, MEDLINE.

• Method: PRISMA‑guided narrative synthesis clustered studies into:

  1. machine learning (ML),

  2. conversational AI, and

  3. knowledge‑based/hybrid AI.

• Quality: Risk‑of‑bias (RoB) appraisal classified 9 studies as Low RoB, 9 as Medium, 5 as High (all High RoB were ML). Primary conclusions emphasize Low/Medium RoB; High RoB results are treated as sensitivity analyses.

What the authors found

1. Machine learning: modest gains, best with multimodal data

• Use‑cases: outcome prediction (who responds, symptom severity), state classification for biofeedback (e.g., fear/arousal), and neurofeedback frameworks.

• Signal: ML can classify anxiety states from physiological and behavioral signals and sometimes predict severe symptom subgroups better than chance, especially with multimodal inputs (e.g., combining voice, heart rate, movement).

• Limits: Generalizability is weak; many models depend on small, homogeneous samples and don’t transfer well across sites. fMRI‑based approaches show promise but are resource‑intensive.

• Process gap: Over half of ML studies set goals without clinician/patient input; few anchored AI logic in contemporary exposure/extinction theory.

2. Conversational AI: engagement and coaching, not replacement

• Use‑cases: virtual agents that coach patients and scaffold sessions inside VR.

• Signal: Early prototypes show high patient acceptance as a supplement to human‑led therapy; automation can reduce therapist workload and improve session flow.

• Limits: No clear superiority over non‑VR comparators in small studies; speech naturalness, empathy, and continuity across sessions remain challenges; unsupervised use raises safety concerns.

• Best practice emerging: Keep therapists “in the loop”—AI can propose, but humans should decide, especially when safety thresholds are involved.

3. Knowledge‑based/hybrid AI: decision support and adaptive scenarios

• Use‑cases: systems that encode exposure rules and adapt VR scenarios to patient responses, sometimes blending rules with learned models.

• Signal: Case reports (e.g., PTSD simulations with haptics) suggest symptom reductions and high stress‑classification accuracy in controlled settings; therapist decision support often aligned with experienced clinicians’ choices.

• Limits: Small samples; maintenance burden for rule‑based systems; risk of over‑automation if tools don’t respect therapist styles and clinical judgment.

Cross‑cutting concerns

• Theory integration is thin: Few applied systems embed modern exposure/extinction principles (e.g., inhibitory learning, expectancy violation, variability, context shifts) into AI decision-making.

• Stakeholder involvement lags: Higher‑quality studies more often involve clinicians and patients; many ML efforts remain “out‑of‑the‑loop,” risking clinically misaligned goals.

• Ethics and equity are underaddressed: Few studies assess model explainability, calibration, subgroup error rates, or fairness; privacy and data‑governance plans are rarely detailed.

• Implementation barriers: Sensor cost and complexity, external validation needs, and clinician training for interpreting probabilistic outputs slow adoption.

What the authors recommend

• Build larger, multi‑site datasets and prioritize external validation to improve generalizability.

• Make AI theory‑driven: encode expectancy violation, variability, spacing, and context shifts; use AI to optimize exposure schedules within safety bounds.

• Compare head‑to‑head: run trials pitting AI‑enabled VRET against traditional VRET to identify who benefits, when, and why.

• Keep humans in the loop: design for therapist oversight, with clear override controls and session memory. Align with emerging guidance (EU AI Act; APA ethics).

• Demand explainability and accountability: report calibration, uncertainty, subgroup errors; publish model cards/data sheets; audit decisions informed by AI.

• Co‑design from day one: include patients, clinicians, engineers, ethicists, and equity experts; build AI literacy for therapists and VR literacy for developers.

The takeaway
AI is starting to make VRET smarter—personalizing stimuli, enhancing engagement, and supporting clinician decisions. But the field is early. The most credible path forward keeps therapists at the helm, bakes exposure theory into code, and involves patients and clinicians at every step. Done well, AI‑enabled VRET could expand access to high‑quality exposure therapy without sacrificing safety or the therapeutic alliance.

Paper details

• Title: Artificial intelligence (AI) for virtual reality exposure therapy (VRET): A systematic review

• Authors: Kamilla Bergsnev; Ana Luisa Sánchez Laws

• Journal: Translational Psychiatry (Nature), Volume 16, Article 208 (2026)

• Published: 26 March 2026

• Open access: Creative Commons Attribution 4.0

• DOI: https://doi.org/10.1038/s41398-026-03936-4

How emotions shape foreign policy: German MPs’ feelings linked to risk choices on Ukraine arms, new thesis finds

Laurence Kroese's master thesis investigates the link between emotions and decision-making under risk from an international relations perspective.

A new master’s thesis from the Center for Geopolitics, Peace, and Security argues that the feelings policymakers bring into the room can tilt how they judge risk—and thus what they advocate—for in high-stakes foreign policy debates.

In an exploratory “plausibility probe” of German federal legislators, Laurence Matthias Kroese links core affect (valence and arousal) to risk-taking tendencies on the contentious question of German weapons deliveries to Ukraine. The study suggests that more positively valenced and higher-arousal states correlate with fewer proposed restrictions on arms transfers, while negatively valenced and lower-arousal states align with tighter constraints.

Why it matters

• The study connects psychological measures of emotion to concrete policy advocacy, addressing a persistent gap in orthodox International Relations (IR) theory, which often invokes emotions (like “fear”) without specifying how they operate.
• It challenges blanket assumptions—especially in liberal IR theory—that domestic actors are uniformly risk-averse, and it offers a micro-level mechanism that could inform neorealist debates over elite consensus and (under)balancing.
• As governments across Europe weigh arms supplies, escalation risks, and alliance signaling, the thesis points to emotion as a subtle but systematic input into risk perception and policy design.

What the researcher did

• Sample: Three sitting members of the German Bundestag’s Foreign Affairs Committee (Auswärtiger Ausschuss), interviewed in September 2025. Two were from the governing Social Democratic Party (SPD), one from the Christian Social Union (CSU).
• Method: Short, semi-structured interviews featuring five brief video stimuli on Russia–NATO–Ukraine themes. After each clip, respondents marked their affect on a two-dimensional “Circumplex” grid (valence: negative to positive; arousal: low to high). They then answered concise policy questions.
• Focus: While five themes were tested, the analysis zeroes in on “Theme 3” (German weapons deliveries to Ukraine), where affect and policy positions aligned most cleanly.

What the researcher found

• All three MPs supported continued arms deliveries—but diverged on conditions.
• The CSU respondent reported positive valence and higher arousal and favored fewer restrictions.
• The two SPD respondents reported negative valence and lower arousal and favored more constraints (e.g., limiting range, emphasizing defensive use, and conditioning deliveries on wartime use).
• This pattern is consistent with an “affect heuristic” in which:
  • Positive affect nudges policymakers toward more optimistic assessments of outcomes (lower perceived downside risk) and relatively more risk-seeking stances.
  • Negative affect nudges toward more pessimistic assessments (higher perceived downside risk) and relatively more risk-averse stances.

The theory behind the link

• Circumplex Model of Affect: Emotions can be mapped along two dimensions—valence (pleasant/unpleasant) and arousal (activation level). Respondents self-reported their states after each video.
• Prospect Theory: People evaluate risky choices relative to a reference point and overweight/underweight probabilities in systematic ways. The thesis leverages findings showing that:
  • Positive affect can inflate estimated frequencies of favorable outcomes; negative affect can inflate estimated frequencies of unfavorable outcomes.
  • Higher arousal can increase reliance on affect in judgments under uncertainty.
• In the Ukraine arms case, the author argues that affect mainly shaped “decision weights”—how likely MPs judged good and bad outcomes—rather than the intrinsic value they assigned to those outcomes.

Implications for IR theory

• Liberalism (Moravcsik): The thesis qualifies the assumption that domestic actors are generally risk-averse. It shows how affect can push elites toward more or less risk acceptance depending on valence and arousal, creating variation even within governing coalitions. Practically, affective diversity may drive compromise—yielding de facto risk-averse outcomes when negatively valenced actors insist on constraints.
• Neorealism (Schweller): The findings hint at an affective micro-mechanism within “elite consensus” and “elite cohesion.” Divergent affective responses among elites can translate into disagreements over the scope and conditions of balancing measures—potentially shaping whether states underbalance, appropriately balance, or overbalance against threats.

Caveats and limitations

• Small N: Just three sitting MPs were analyzed (a fourth interview was excluded due to technical issues), limiting generalizability.
• Stimulus design: Clips were non-randomized, and targets of affect could be ambiguous (e.g., one SPD respondent focused on a figure in the video rather than the peace-talks theme itself in another section).
• No baseline mood: Without pre-interview mood measures, distinguishing incidental from integral affect is difficult.
• Measurement: A single-point self-report on the Circumplex grid can’t capture mixed or discrete emotions; facial expression notes (no video recordings) were used cautiously and not as definitive evidence.
• Context: Real-world policy is iterative and collective; the study focuses on an individual-level mechanism.

What’s next
The author calls the study a feasibility test of “importing a simple dimensional affect measure into political elite IR research,” and outlines improvements for future work:

• Larger, more diverse samples (including diplomats and bureaucrats, who are more accessible than MPs).
• Randomized stimuli and multiple angles of a single policy theme to clarify targets of affect.
• Baseline mood checks and, when feasible, controlled Prospect Theory tasks embedded in interviews.
• Longitudinal designs to track how affect and advocacy evolve.
• Attention to gender and other demographic moderators of affect–risk links.

The bottom line
Emotions aren’t just rhetorical flourishes in foreign policy—they can shift how leaders judge probabilities and calibrate risk. In a tightly focused German case, positive affect aligned with fewer constraints on Ukraine arms deliveries and negative affect with tighter guardrails. That pattern doesn’t rewrite IR theory, but it sharpens it: micro-level affect may help explain when and why risk preferences diverge within governments, and how those divergences shape the compromises that become state policy.

About the study

• Title: “Emotion and International Relations. The affective state of individuals and decision-making under risk in International Relations.”

• Author: Laurence Matthias Kroese

• Program: Master’s thesis in Peace and Conflict Studies

• Institution: Center for Geopolitics, Peace, and Security

• Date: May 2026

• Supervisor: Prof. Ana Luisa Sanchez Laws

Immersive journalism’s emotional edge: new chapter maps what we know, what’s muddled, and how to fix it

A new scholarly chapter, “Emotions in Immersive Journalism” by Kamilla Bergsnev and Ana Luisa Sánchez Laws, takes stock of how virtual, augmented, and mixed reality are reshaping the emotional experience of news—and what researchers and newsrooms must do to study and use those emotions responsibly.

The big picture

• Immersive journalism places audiences “in the middle” of events using extended reality (XR). Early works like Emblematic’s Hunger in L.A. showed that even animated scenes can elicit strong reactions—viewers cried and tried to intervene.

• The format rose rapidly from 2015 as newsrooms sought to reconnect with fragmented, distrustful audiences. But it magnifies a longstanding tension: the profession’s objectivity norm versus an “emotional turn” in journalism.

• The key questions now are empirical: Do immersive formats evoke different emotions than text or video? How? And do audiences want that added emotional connection?

How the authors assessed the field

• Design: A narrative overview anchored in a PRISMA‑guided search (fall 2023) for “immersive journalism” AND “emotion” in English.

• Screening: From 17 Scopus entries and a filtered sample of Google Scholar hits, the authors retained eight empirical studies for in‑depth review (excluding non‑XR, non‑empirical, and review pieces).

• Frame: Concepts and measures are examined through the lens of affective science, particularly debates between “basic/discrete” emotion theories and dimensional, constructionist models (notably Russell’s two‑axis circumplex of valence and arousal).

What the studies show

1. Immersion can heighten “being there,” but emotion measurement is inconsistent

• A foundational experiment (Sundar et al., 2017) compared VR, 360° video, and text for two New York Times stories. It linked presence, interaction, and realism to credibility, recall, and sharing intentions.

• Critique: The study conflated emotional intensity with negative valence and drew on scales not designed for emotion (borrowed from relational communication), blurring core distinctions well established in affective science (valence ≠ arousal).

2. More rigorous designs are emerging—but constructs still get mixed

• Greber et al. (2023) used sophisticated statistics and open science practices to test how inclusion, interactivity, and narrative immersiveness shape responses. They found:

  • Inclusion and interaction increased intensity.

  • Narrative immersiveness influenced valence.

  • Empathy tendency moderated effects.

• Caveats: Intensity was used as a stand‑in for arousal (collapsing a bi‑directional “activation” axis into a one‑way scale), and PANAS—built to capture mood, not momentary emotion—assessed post‑exposure state.

3. Emotional Personalization boosts presence and recall; effects on valence and empathy vary

• Li & Lee (2019/2022) compared a BBC 360° war/conflict piece with and without emotionally rich personal testimonies, across VR and desktop:

  • Emotional Personalization increased presence and recall.

  • VR produced lower emotional valence (more negative feelings) and stronger empathy than desktop.

• The study replicates earlier patterns but, like others, offers limited theoretical unpacking of emotion constructs.

4. Empathy isn’t guaranteed—even when immersion shifts attitudes

• Steinfeld (2020) reconstructed workplace sexual harassment scenes across written, conventional video, and 360° VR conditions:

  • Immersion helped reduce stereotypical views (opinion change).

  • No clear increase in empathetic reactions (vicarious emotion) emerged.

• The result challenges the early “empathy machine” narrative and suggests empathy may function more as an emotion regulation mechanism than a direct output of immersion.

5. Audiences report “affect and empathy” gratifications—but theory links are thin

• Nielsen & Sheets (2021) used Uses & Gratifications to show participants value personal experience and affect in immersive journalism and feel it helps them understand their emotions. However, an explicit affect theory framework was absent.

6. Haptics and thermal cues may backfire on trust

• Ooms et al. (2023) added somatosensory feedback (thermal/vibration) to news videos, grounded in embodied cognition theories:

  • No reliable boosts in perceived emotional intensity or valence (N=20).

  • Viewers preferred no somatosensory add‑ons, citing a need for agency over bodily reactions to maintain trust.

Why this matters

• Journalism use: Immersive formats can deepen presence and recall and shape perceptions—but the pathway from “more immersive” to “more empathetic” isn’t automatic.

• Research quality: Without clear, theory‑consistent measures (e.g., treating valence and arousal as distinct), findings are hard to compare or cumulate.

• Ethics and trust: Over‑engineering audience emotions (e.g., via haptics) may undermine credibility if audiences perceive manipulation.

Key takeaways from affective science the field should adopt

• Use dimensional models precisely: Valence (pleasant–unpleasant) and arousal (activated–deactivated) are separate axes; “intensity” is not a synonym for “negative.”

• Distinguish emotion from mood: Instruments like PANAS capture mood; short‑lived, stimulus‑linked emotions need other tools (e.g., valence–arousal grids, momentary ratings).

• Be explicit about empathy: Treat empathy as a process (often emotion regulation), not a guaranteed outcome of presence.

• Clarify related constructs: Presence is a feeling (the subjective “being there”); it is not a proxy for empathy or emotion strength.

What the authors recommend next

• Sharpen concepts and measures: Align operationalization with contemporary affective science; avoid collapsing constructs; justify scale choices.

• Build comparative, cumulative evidence: Replicate with shared measures across labs; report valence and arousal separately; include manipulation checks for presence and agency.

• Go transdisciplinary: Create multi‑lab collaborations that bring together journalists, media scholars, engineers, and affective scientists to co‑design studies and tools.

• Keep audiences’ agency central: Design immersive experiences that inform and engage without overstepping into perceived emotional coercion—especially when adding haptics or other embodied cues.

The bottom line
Immersive journalism can powerfully shape how audiences feel and what they remember. But the field is still sorting out how to measure those feelings—and when, or whether, they translate into empathy and action. By importing the best of affective science and collaborating across disciplines, researchers and newsrooms can harness immersion’s strengths while guarding against conceptual muddle and erosion of trust.

Antigone Heraclidou  –  Postdoctoral Fellow

Laurence Kroese

Ana Luisa Sanchez Laws  –  Professor

Kamilla Bergsnev

Bernhard Spanlang