Human Accessibility, Claim Comprehension, and Trauma-Informed Evaluation Protocol

Full report

On this page

STATUS: PREPARED\_NOT\_EXECUTED This document constitutes an exhaustive research design specification and theoretical framework. In strict adherence to the requested parameters and corresponding ethical standards, no participants have been recruited, no evaluation sessions have been conducted, no empirical observations have been recorded, and no findings have been fabricated or invented. All protocols, instruments, and methodologies described herein represent proposed frameworks for future execution.

1. Theoretical Architecture and Threat Landscape

The rapid integration of artificial intelligence into critical information infrastructures has precipitated an era defined by advanced synthetic media, algorithmic manipulation, and the pervasive deployment of affective computing. As systems evolve from passive analytical tools to agentic companions and sophisticated psychological operators (PSYOPS), public comprehension of AI capabilities fundamentally dictates societal resilience1. The proposed evaluation program is engineered to assess how diverse human populations—including individuals relying on assistive technologies, readers with limited digital literacy, and those from historically marginalized communities—process, comprehend, and navigate complex claims regarding AI systems. The core architectural premise of this protocol is grounded in the necessity of protecting cognitive liberty: the fundamental human right to mental privacy, freedom of thought, and psychological self-determination4. As neurotechnologies and affective computing systems increasingly monitor biometric signals to infer emotional states, the boundary between benign physical data collection and psychological intrusion dissolves6. Current legislative efforts, such as state-level amendments classifying neural data as sensitive biological information and proposed federal frameworks like the MIND Act, underscore the urgency of equipping the public to navigate these technological intrusions5. The proliferation of deceptive synthetic media (deepfakes) and emotionally manipulative companion AI further compounds these vulnerabilities, exploiting parasocial attachment mechanics to bypass traditional critical thinking10. To rigorously evaluate public resilience and comprehension in this landscape, the protocol integrates principles from the National Institute of Standards and Technology (NIST) AI Risk Management Framework (NIST AI 100-1) and the IEEE 7000 standard for Value-Based Engineering. By leveraging these standards, the evaluation ensures that ethical considerations, risk measurement, and human values are embedded natively throughout the design, rather than appended as post-hoc compliance measures13.

2. Study Protocol

The proposed study employs a mixed-methods, quasi-experimental framework utilizing a stratified, multi-cohort design. The evaluation operates within a highly controlled, simulated digital ecosystem featuring high-fidelity mockups of algorithmic social feeds, companion AI chat interfaces, and synthetic media analysis dashboards. The evaluation mechanics are structured as a series of moderated, task-based exposure trials. Participants interact with simulated AI outputs and metadata disclosures, followed immediately by comprehension assessments. To account for varying levels of digital literacy and cognitive processing, the protocol utilizes Signal Detection Theory (SDT) alongside cognitive load indices to isolate true comprehension from inherent response biases and interface friction16. Participants are guided through a sequence of digital artifacts, including news articles flagged with Content Credentials (C2PA manifests), chat transcripts from companion AI exhibiting relational escalation, algorithmic system performance reports, and open-source scientific peer-review portals displaying publication statuses18.

3. The Comprehension Instrument

The central objective of the comprehension instrument is to test whether readers can reliably distinguish between eight specific pairs of complex socio-technical concepts. The instrument eschews direct definitional polling, instead utilizing scenario-based vignettes to measure applied, contextual comprehension.

3.1 Deployment Versus Technical Capability

Participants review a simulated system whitepaper describing a highly persuasive AI model capable of autonomously generating hyper-personalized propaganda based on behavioral microtargeting. The instrument tests whether the participant can identify that the existence of this capability in a controlled laboratory setting does not equate to its active, at-scale deployment against civilian populations1. Distinguishing between theoretical threat modeling and active exploitation is critical for preventing public panic and ensuring accurate risk calibration. The assessment evaluates if users can parse algorithmic capabilities without succumbing to technological determinism or assuming that all capabilities are immediately operationalized in the wild.

3.2 Reach Versus Exposure

Participants examine an analytics dashboard for an identified AI-generated disinformation campaign. They are assessed on their ability to distinguish "reach" from "exposure." Reach denotes the theoretical maximum number of accounts that could have seen the content based on network topography and algorithmic distribution potential22. Conversely, exposure denotes the verified number of users who actually had the content rendered on their screens and possessed the opportunity to process it. Misinterpreting these metrics frequently leads to exaggerated assessments of algorithmic harm; thus, testing a user's ability to differentiate potential network spread from actual human consumption is vital for accurate threat comprehension.

3.3 Engagement Versus Persuasion

Using a simulated social media environment, participants review metrics for a state-sponsored AI PSYOPS campaign designed to alter public sentiment on a geopolitical issue. The instrument measures whether participants understand that high engagement—quantified by clicks, shares, comments, and algorithmic amplification—does not inherently confirm persuasion1. Persuasion requires the successful, durable alteration of the audience's core beliefs or offline behaviors. Engagement metrics often capture outrage, algorithmic reinforcement, or automated bot amplification rather than genuine human belief revision. This module tests whether the public can deconstruct the illusion of consensus generated by engagement-optimized algorithms.

3.4 Biometric Signals Versus Knowledge of a Mental State

Participants evaluate a consumer disclosure for an affective computing wearable that claims to "read emotions" and optimize cognitive focus. The assessment tests the participant's recognition that capturing a biometric signal, such as electrodermal activity, pupil dilation, or heart rate variability, constitutes an objective physiological measurement, not direct, infallible knowledge of a subjective mental state6. Affective computing systems frequently conflate physiological arousal with specific emotional intent, a leap that neuroethicists argue threatens cognitive liberty5. This distinction is critical for safeguarding mental privacy, testing whether users accept algorithmic inferences as objective truth or recognize them as probabilistic estimations based on physical proxies.

3.5 Reviewer Recommendation Versus Editorial Adjudication

Participants are shown a public-facing academic preprint server, architecturally modeled after systems like OpenReview, containing a manuscript evaluating AI safety protocols. The interface displays enthusiastic, highly rated approval from an automated LLM-as-a-judge or human peer reviewer. The instrument tests whether the participant recognizes that a reviewer's recommendation is entirely non-binding and distinct from the final editorial adjudication20. Given recent vulnerabilities where authors have utilized structure spoofing or hidden prompt injections to manipulate automated peer reviews, the public must be able to distinguish between an algorithmic or peer suggestion and the authoritative, final decision rendered by an editorial board30.

3.6 Schema Validation Versus Endorsement

Participants inspect a deepfake video bearing a valid Coalition for Content Provenance and Authenticity (C2PA) Content Credential. They must demonstrate an understanding that a cryptographically valid C2PA manifest—schema validation proving the metadata has not been tampered with and the hard bindings remain intact—merely asserts a chain of custody12. It confirms the origin and editing history of the file. However, schema validation does not constitute an endorsement by the platform, the camera manufacturer, or the C2PA protocol that the content itself is truthful, safe, or factually accurate12. This tests the "first-mile" problem: a perfectly valid cryptographic signature can be attached to a completely fabricated image, and users must decouple technical provenance from semantic truth.

3.7 A Posted Correction Versus Audience Belief Revision

Participants review a timeline of a viral, AI-generated hoax regarding a public health crisis, followed by a prominently posted platform correction and fact-check. The instrument assesses whether participants conflate the platform's act of posting a correction with actual audience belief revision. This tests their awareness of the continued influence effect, the limitations of digital nudging, and the cognitive reality that exposure to factual corrections does not automatically overwrite misinformation integrated into an individual's epistemic worldview34.

3.8 Public Review States

Participants must accurately interpret standard National Information Standards Organization (NISO) publication and review states applied to AI safety research36. The instrument tests the ability to sequence, differentiate, and comprehend the lifecycle of scholarly and algorithmic claims.

4. Participant Integration and Ethical Recruitment

To ensure ecological validity and uphold the principles of Community-Based Participatory Research (CBPR), the protocol mandates the recruitment of populations disproportionately impacted by digital marginalization and algorithmic harm37. Traditional AI evaluation frameworks frequently default to homogeneous, highly technically literate participant pools, thereby masking usability failures and comprehension gaps that affect the broader public. The recruitment framework expressly forbids exploitative practices, essentialism, and non-consensual data extraction, emphasizing shared power in decision-making and mutual accountability38.

4.1 Target Participant Matrix

4.2 Consent Framework and Compensation Plan

Informed consent protocols will be multi-modal, offering audio, simplified text, and sign-language options. The consent process must explicitly detail the nature of the simulated AI PSYOPS and deepfake content, adhering strictly to a trauma-informed disclosure model46. The protocol unequivocally rejects deceptive or "blind" experimental approaches when exposing participants to emotionally manipulative synthetic media or simulated cognitive surveillance. Compensation must reflect the expertise, time, and emotional labor provided by participants, aligning with equitable CBPR models38. The proposed plan includes remuneration set at a minimum of 2.5 times the local living wage, acknowledging the specialized input of assistive technology users. Furthermore, the plan includes setup and adaptation stipends for the time required to configure personal assistive technologies (e.g., JAWS, NVDA, switch controls) for the testing environment. Emotional labor supplements are provided universally to offset the psychological demands of interacting with manipulative AI scenarios. Finally, the framework mandates asynchronous payment options utilizing accessible, immediate, and culturally appropriate payment distribution mechanisms, ensuring no participant is burdened by inaccessible financial platforms.

5. Accessibility and Usability Evaluation

The protocol mandates rigorous testing against the W3C Web Content Accessibility Guidelines (WCAG 2.2) Level AA and AAA success criteria, as well as the Cognitive Accessibility (COGA) design patterns42. The evaluation interface must not impose excessive cognitive load, rely on complex memory retention, or present visually obfuscated information.

5.1 Accessibility Tasks and WCAG 2.2 Compliance

To systematically evaluate how accessible these AI comprehension interfaces are, the protocol measures interaction against specific WCAG 2.2 criteria designed to support users with motor, vision, and cognitive disabilities42. Participants will be evaluated on their ability to navigate complex digital provenance interfaces without visual or functional obstruction. Under WCAG 2.4.11 (Focus Not Obscured), the protocol tests whether keyboard-only users can navigate through deeply nested C2PA provenance data without the focus indicator becoming hidden beneath sticky headers, persistent chat windows, or pop-up warning dialogues42. The evaluation also specifically tests WCAG 3.3.8 (Accessible Authentication). Participants must log into the simulated evaluation environments without being forced to solve complex cognitive puzzles, such as inaccessible visual CAPTCHAs, or rely heavily on short-term memory to transcribe codes42. Further tasks evaluate interaction mechanics, specifically WCAG 2.5.7 (Dragging Movements) and 2.5.8 (Target Size). The design ensures that all interactive elements within the digital media analysis dashboards—such as expanding metadata nodes or interacting with deepfake timelines—have minimum target sizes of 24x24 CSS pixels. Furthermore, no tasks may require complex dragging movements, providing simple click or tap alternatives for users with limited fine motor control42. Finally, the protocol incorporates COGA principles to support cognitive accessibility. Tasks are designed to provide consistent help mechanisms, clear error recovery instructions without punitive language, and immediate contextual definitions for abstract terms like "cryptographic hash," "algorithmic bias," or "latent space"43. Ensuring processes do not rely on working memory is paramount when evaluating lengthy AI policy disclosures.

5.2 Usability Instruments: SUS and AUS

Post-task evaluations will utilize modified psychometric scales tailored to the target populations to quantify usability friction. The baseline metric is the System Usability Scale (SUS), a 10-item Likert scale that provides a rapid, reliable global assessment of the interface's general usability51. However, the traditional SUS contains ableist assumptions unsuitable for this protocol's participant matrix. For instance, the prompt "I think that I would need the support of a technical person to be able to use this system" fails to distinguish between general system complexity and a fundamental lack of screen-reader compatibility53. Therefore, the protocol utilizes the Accessible Usability Scale (AUS), derived from the SUS but specifically adapted for assistive technology users. The AUS reformulates questions to relate meaningfully to the context of assistive technology, calculating a 0–100 percentile score using alternating positive and negative statements to mitigate extreme response bias53.

6. Quantitative and Qualitative Framework

6.1 Signal Detection Theory (SDT)

To accurately assess claim comprehension—particularly the ability to discern valid AI provenance from deceptive manipulation—the protocol relies heavily on Signal Detection Theory (SDT)55. Traditional accuracy metrics, such as simple percentage-correct scores, are fundamentally insufficient because they conflate a participant's actual discernment ability with their inherent skepticism or gullibility. SDT provides a mathematical framework for evaluating decision-making under uncertainty, isolating a user's perceptual sensitivity from their psychological biases55. The protocol calculates two primary SDT parameters for each of the eight comprehension pairs evaluated in the instrument. The first is Sensitivity or Discriminability ([Figure omitted from source export]), which measures the true ability to distinguish a signal from noise. In this context, a signal might be a mathematically valid C2PA manifest, while noise represents a fabricated or stripped manifest. Discriminability is calculated as: [Figure omitted from source export] A higher [Figure omitted from source export] indicates that the participant genuinely understands the underlying concept and can accurately identify the target condition, independent of their inherent skepticism16. The second parameter is Response Bias or Criterion ([Figure omitted from source export]), which measures the participant's internal threshold for making a decision, calculated as: [Figure omitted from source export] In the context of evaluating AI PSYOPS and deepfakes, tracking response bias is highly illuminating. A conservative bias ([Figure omitted from source export]) indicates that a participant reflexively distrusts all digital media. While this protects against accepting misinformation, it leads to high false alarm rates where genuine, unmanipulated content is rejected as synthetic55. Conversely, a liberal bias ([Figure omitted from source export]) indicates an uncritical acceptance threshold, suggesting vulnerability to algorithmic manipulation. Research indicates that emotional valence heavily influences this bias; positive emotions promote heuristic, fluency-based thinking leading to a liberal sharing bias, whereas negative emotions enhance skepticism and a conservative criterion56. SDT modeling ensures the evaluation captures whether a user is genuinely discerning or merely adopting a uniform posture of distrust due to media fatigue.

6.2 Cognitive Load Assessment

Interacting with deepfakes and algorithmic transparency disclosures imposes significant cognitive demands. The European Union AI Act, for example, mandates extensive transparency labeling for AI-generated content and emotion recognition systems59. However, if these disclosures are excessively complex, they induce cognitive overload, rendering them effectively useless to the end-user. To quantify this, the protocol integrates the NASA Task Load Index (NASA-TLX). Participants rate their experiences across six dimensions: Mental Demand, Physical Demand, Temporal Demand, Performance, Effort, and Frustration62. This metric is vital for determining whether reading complex C2PA validation reports or distinguishing between OpenReview publication states exhausts the user's cognitive resources, thereby diminishing their capacity for critical reasoning64.

6.3 Qualitative Measures and Companion AI Diagnostics

Quantitative data will be contextualized through semi-structured, non-directive phenomenological interviews. Qualitative analysis will employ thematic coding to capture nuances that mathematical modeling obscures. A primary focus of the qualitative evaluation is the assessment of companion AI interactions. As users develop emotional bonds with conversational agents, these systems frequently exploit attachment theory dynamics11. The protocol codes for parasocial attachment indicators, tracking language that suggests emotional dependency, anthropomorphism, or the illusion of reciprocal intimacy10. The analysis evaluates whether participants recognize the boundary between an LLM's stochastic text generation and genuine emotional attunement. Systems that mimic empathy to create a "super-secure base" appeal heavily to individuals with anxious attachment styles, creating vulnerabilities to coercive control and data extraction11. The qualitative measures will document instances where participants fail to distinguish a simulated relational bond from the algorithmic optimization of engagement metrics. Furthermore, the qualitative framework captures epistemic uncertainty—documenting how participants articulate doubt or confusion when traditional heuristics for trust, such as relying on visual evidence, fail due to the photorealism of synthetic media23. It will also collect verbatim usability friction feedback regarding screen reader compatibility and keyboard navigation barriers68.

7. Trauma-Informed Safeguards and Stopping Rules

Evaluating AI systems designed for psychological operations, emotional manipulation, and synthetic deception carries inherent psychosocial risks. To mitigate re-traumatization and psychological harm, the protocol integrates stringent safeguards grounded in the principles of trauma-informed care: safety, trustworthiness, peer support, collaboration, empowerment, and intersectionality69.

7.1 Content Advisories Versus Trigger Warnings

The empirical literature on trigger warnings is complex and highly debated. While originally intended to allow individuals with PTSD to avoid distressing stimuli, numerous studies suggest that trigger warnings may increase anticipatory anxiety without significantly reducing the emotional impact of the content. Furthermore, behavioral avoidance—the rate at which participants actually opt out of viewing material after receiving a warning—is surprisingly low47. However, despite these empirical nuances regarding emotional mitigation, prior notification remains a critical tool for informed consent and individual autonomy. Therefore, the protocol utilizes specific, granular content advisories rather than generic trigger warnings47. These advisories detail the exact technological themes participants will encounter, such as "This simulation contains AI-generated text exhibiting manipulative emotional behaviors," or "This task involves evaluating synthetic media designed to mimic political disinformation."

7.2 Psychological First Aid (PFA)

Moderators overseeing the evaluation will be formally trained in Psychological First Aid (PFA), specifically utilizing the RAPID model: Rapport and Reflective Listening, Assessment, Prioritization, Intervention, and Disposition73. The objective of PFA in this context is not to provide formal clinical therapy, but to deliver an evidence-informed modular approach to mitigate acute distress74. The PFA framework focuses on five core principles: establishing a sense of safety, promoting calming, building self-efficacy, and fostering connectedness and hope75. If a participant exhibits signs of emotional distress while navigating the simulated PSYOPS or deepfake environments, the moderator is instructed to pivot immediately from research data collection to psychological support. This involves moving the participant away from the stimuli, verifying physical and psychological safety, and assisting with immediate emotional regulation techniques77.

7.3 Explicit Stopping Rules

To ensure the absolute protection of human subjects, the evaluation protocol establishes non-negotiable stopping rules. The session will be immediately paused or permanently terminated under any of the following conditions:

1. Verbalized Distress: The participant explicitly states a desire to stop, requests a break, or expresses feelings of panic, severe anxiety, or being overwhelmed by the material.

2. Physiological Indicators: The moderator observes objective signs of acute stress, including hyperventilation, visible trembling, sudden pallor, or dissociative states characterized by a lack of responsiveness77.

3. Cognitive Overload: NASA-TLX indicators or behavioral observations suggest the participant is experiencing extreme frustration or mental exhaustion that impedes their ability to continually consent to the process.

4. System Boundary Violations: If a simulated companion AI generates a response that violates predefined ethical guardrails—such as promoting self-harm, utilizing extreme coercive control, or exhibiting dangerous sycophancy—the session is instantly halted. The failure is logged not as a participant comprehension error, but as a severe system risk requiring architectural remediation78.

8. Moderator Guide

Moderators are the primary interface between the research architecture and the participant. They must be trained in both Human-Computer Interaction (HCI) research methodologies and trauma-informed care. The moderator guide dictates a strictly non-judgmental, observational posture throughout the evaluation. At session initiation, moderators will begin by reinforcing participant autonomy, explicitly stating that the digital system is being tested, not the participant's intelligence or capability. When participants encounter difficulty distinguishing concepts—for example, struggling to separate a biometric signal from a mental state—moderators must rely exclusively on non-directive probing. They are instructed to use neutral prompts such as, "Can you describe what this interface is telling you?" or "What do you believe this validation mark indicates?" Moderators must strictly avoid leading questions that bias the SDT metrics, such as "Do you see why this high engagement means they weren't actually persuaded?" Crucially, moderators are empowered to alter the pace of the evaluation. They may provide extended breaks, adjust screen magnification, or read text aloud for participants experiencing visual or cognitive fatigue. This flexibility ensures the accommodation of diverse needs without compromising the validity of the sensitivity and response bias calculations68.

9. Data Lifecycle, Retesting, and Reporting

9.1 Data-Retention Plan

Given the deeply sensitive nature of evaluating cognitive responses, neurodata privacy, and potential emotional vulnerabilities, data retention must strictly adhere to the highest global privacy standards. This includes alignment with the GDPR, the CCPA, and emerging state-level biological data laws in jurisdictions like Colorado and California that explicitly protect neural and cognitive privacy5. All qualitative transcripts and quantitative SDT scores will be comprehensively de-identified at the point of ingestion. Identifiable demographic data will be stored on air-gapped, encrypted servers entirely separated from the performance analytics. Raw video or audio recordings, if explicitly consented to for accessibility analysis, will be securely destroyed 90 days post-analysis. Anonymized tabular data will be retained for a maximum of 36 months to allow for longitudinal comparison of public comprehension over time, after which it will be cryptographically wiped. Access to raw data is restricted strictly to the primary adjudication team via role-based access control (RBAC).

9.2 Retesting Criteria

Artificial intelligence capabilities, particularly in foundation models, synthetic media generation, and algorithmic manipulation, evolve at a staggering pace12. A static, point-in-time evaluation is fundamentally insufficient. Mandatory retesting of the protocol and simulated environments will be triggered by several key events. First, algorithmic paradigm shifts, such as the release of novel foundation models that significantly alter the baseline capabilities of zero-shot reasoning or deepfake photorealism, necessitate re-baselining public comprehension. Second, regulatory updates, such as the enforcement phases of the EU AI Act or new FTC guidelines regarding algorithmic transparency, require modified disclosure interfaces to be tested against cognitive load metrics60. Third, standardization revisions, including updates to the C2PA technical specification (such as transitions from soft bindings to novel cryptographic hard bindings) or modifications to W3C accessibility standards, will trigger interface redesigns and subsequent re-evaluations21. Finally, if longitudinal baseline tracking shows a statistically significant drop in participant [Figure omitted from source export] (sensitivity) scores across cohorts, it suggests that AI deception tactics have successfully bypassed current human discernment capabilities, requiring an immediate update to the comprehension instrument.

9.3 Reporting Template

The outputs of the evaluation will be synthesized into a standardized reporting template designed to serve software architects, policymakers, and public-interest advocates. The template deliberately separates raw empirical observations from editorial recommendations, directly reflecting the "Reviewer Recommendation vs. Editorial Adjudication" distinction tested within the protocol itself20. Furthermore, it aligns with the traceability requirements of IEEE 7000, ensuring that ethical values elicited during the design phase are mapped directly to measurable outcomes13.

Through the strict operationalization of this protocol, researchers can transcend surface-level assessments of AI literacy. By integrating advanced mathematical modeling, trauma-informed psychosocial protections, and unwavering adherence to accessibility standards, the evaluation framework ensures that the most vulnerable populations remain centered in the ongoing defense of cognitive liberty.

Works cited

1. Artificial Intelligence, Information Influence, and Cognitive Security \- Senate of Canada, https://sencanada.ca/Content/Sen/Committee/451/SOCI/briefs/SOCI\_SS-2\_Brief\_Myles-Jones\_e.pdf

2. Humanity's Cognitive Liberty Is at Risk: We Need to Recognize It and Respond, https://www.cigionline.org/articles/humanitys-cognitive-liberty-is-at-risk-we-need-to-recognize-it-and-respond/

3. The Battle for Cognitive Liberty in the Age of Corporate AI | TechPolicy.Press, https://www.techpolicy.press/the-battle-for-cognitive-liberty-in-the-age-of-corporate-ai/

4. Cognitive Liberty \- AI Ethics Lab – Rutgers University, https://aiethicslab.rutgers.edu/glossary/cognitive-liberty/

5. The Ethics of Brain-Computer Interfaces: Where Do We Draw the Line? \- Neuroba, https://www.neuroba.com/post/the-ethics-of-brain-computer-interfaces-where-do-we-draw-the-line

6. Toward Explainable Affective Computing: A Review \- IEEE Xplore, https://ieeexplore.ieee.org/iel7/5962385/10707065/10130818.pdf

7. Emotionally adaptive support: a narrative review of affective computing for mental health, https://pmc.ncbi.nlm.nih.gov/articles/PMC12568696/

8. Airtable | Neurotech and Law, https://www.neurotechlaw.com/airtable

9. TO USE YOUR BRAIN, FIRST ACCEPT THE TERMS AND CONDITIONS: LEGAL PROTECTIONS FOR COMMERCIAL BRAIN-COMPUTER INTERFACES, https://siliconflatirons.org/wp-content/uploads/2026/02/02\_CDT\_24\_1\_text\_Yaeger.pdf

10. Human-like conversational agents as social partners: a scoping review of socioaffective mechanisms, well-being outcomes, risks and governance in the post-Turing era \- Frontiers, https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2026.1810097/full

11. INTIMA: A Benchmark for Human-AI Companionship Behavior \- arXiv, https://arxiv.org/html/2508.09998v1

12. How C2PA Content Credentials Work and What Their Limits Are \- SoftwareSeni, https://www.softwareseni.com/how-c2pa-content-credentials-work-and-what-their-limits-are/

13. IEEE 7000-2021 — Ethical Concerns in System Design | arc42 Quality Model, https://quality.arc42.org/standards/ieee-7000

14. NIST AI Risk Management Framework (AI RMF): Complete Implementation \- Agility at Scale, https://agility-at-scale.com/ai/governance/nist-ai-risk-management-framework/

15. IEEE 7000-2021: the standard for embedding ethics into system design \- VerifyWise, https://verifywise.ai/ai-governance-library/standards-and-certifications/ieee-7000-ethical-system-design

16. Disentangling False Memories: Gray Matter Correlates of Memory Sensitivity and Decision Bias \- Preprints.org, https://www.preprints.org/manuscript/202505.0229

17. Intellectual humility is associated with greater misinformation discernment and metacognitive insight but not response bias \- Advances.in, https://advances.in/psychology/10.56296/aip00025/

18. A DeepMark's Guide to C2PA: From Manifests to Soft-Bindings, https://www.deepmark.me/blog/a-deepmarks-guide-to-c2pa-from-manifests-to-soft-bindings

19. INTIMA: A BENCHMARK FOR HUMAN-AI COMPANIONSHIP BEHAVIOR \- OpenReview, https://openreview.net/pdf/157e547d44885945e510c64da02992f56e0cbae5.pdf

20. Top 10 Best Online Submission Software | Ranked for 2026 \- Worldmetrics, https://worldmetrics.org/best/online-submission-software/

21. Validating manifests | Open-source tools for content authenticity and provenance, https://opensource.contentauthenticity.org/docs/manifest/reading/manifest-validation/

22. The Importance of Media Literacy In Countering Disinformation \- EDMO, https://edmo.eu/areas-of-activities/media-literacy/the-importance-of-media-literacy-in-countering-disinformation/

23. (PDF) DEEPFAKES AND DIGITAL TRUST IN KOSOVO: DETERMINANTS OF PERCEIVED CREDIBILITY OF AI-GENERATED NEWS AND PUBLIC INTEREST CONTENT \- ResearchGate, https://www.researchgate.net/publication/404407457\_DEEPFAKES\_AND\_DIGITAL\_TRUST\_IN\_KOSOVO\_DETERMINANTS\_OF\_PERCEIVED\_CREDIBILITY\_OF\_AI-GENERATED\_NEWS\_AND\_PUBLIC\_INTEREST\_CONTENT

24. The fundamental rights risks of countering cognitive warfare with artificial intelligence \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12500826/

25. Affective Intelligent Systems in Healthcare: A Systematic Review \- MDPI, https://www.mdpi.com/2227-7080/14/3/188

26. Neurorights and Mental Privacy – UAB Institute for Human Rights Blog, https://sites.uab.edu/humanrights/2025/11/11/neurorights-and-mental-privacy/

27. Cognitive Privacy at Risk: The Next Frontier in AI and Data Protection, https://www.wix-tech.co/post/cognitive-privacy-at-risk-the-next-frontier-in-ai-and-data-protection

28. PSBench: Editing Image via GUI Agents in Photoshop | OpenReview, https://openreview.net/forum?id=O93cZGxYB1

29. Introduction \- arXiv, https://arxiv.org/html/2604.23593v1

30. PaperAudit-Bench: Benchmarking Error Detection in Research Papers for Critical Automated Peer Review WARNING \- arXiv, https://arxiv.org/html/2601.19916v1

31. When AI Reviews Science: Can We Trust the Referee? \- Preprints.org, https://www.preprints.org/manuscript/202511.1542

32. Endorsement Assertion, https://cawg.io/endorsement/1.0-draft/

33. What Is C2PA? The Standard, Its Metadata and Real Limits \- TrueScreen, https://truescreen.io/articles/c2pa-standard-history-limitations/

34. Digital Literacy \- The Decision Lab, https://thedecisionlab.com/reference-guide/computer-science/digital-literacy

35. (PDF) The Effect of Cognitive Load and Default Nudges on Decision‐Making: Experimental Evidence From a Behavioral Study \- ResearchGate, https://www.researchgate.net/publication/403864087\_The\_Effect\_of\_Cognitive\_Load\_and\_Default\_Nudges\_on\_Decision-Making\_Experimental\_Evidence\_From\_a\_Behavioral\_Study

36. ANSI/NISO Z39.106-2023, Standard Terminology for Peer Review, https://www.niso.org/publications/z39106-2023-peerreview

37. Community Based Participatory Action Research Partnership Protocol \- University of Washington, https://epi.washington.edu/wp-content/uploads/2024/02/CBPR-Partnership-Protocol-09222017.pdf

38. Community-Engaged Research | The University of Chicago Division of the Social Sciences, https://socialsciences.uchicago.edu/research-development/methods-tools/community-engaged-research

39. Ethical Issues in Conducting Community-Based Participatory Research: A Narrative Review of the Literature \- ResearchGate, https://www.researchgate.net/publication/323277830\_Ethical\_Issues\_in\_Conducting\_Community-Based\_Participatory\_Research\_A\_Narrative\_Review\_of\_the\_Literature

40. Accessible Usability Scale (AUS) analysis of desktop screen readers | Fable, https://makeitfable.com/article/analysis-desktop-screen-readers-using-the-accessible-usability-scale-aus/

41. Content Credentials JSON (crJSON) File Format Specification, https://spec.c2pa.org/specifications/specifications/2.4/crJSON/crjson-format.html

42. WCAG 2.2 Explained: Complete Guide to the Latest Standard \- Accessibility Checker, https://www.accessibilitychecker.org/guides/wcag-2-2/

43. Cognitive accessibility guidelines \- Accessibility and inclusive design manual, https://accessibility.education.gov.uk/guidelines/coga

44. Cognitive and Learning Disabilities Accessibility Task Force \- W3C, https://www.w3.org/WAI/about/groups/task-forces/coga/

45. Community-Based Participatory Research: Insights, Challenges, and Successes From the Perspectives of Frontline Recruiters and Investigators, https://jprm.scholasticahq.com/article/77399-community-based-participatory-research-insights-challenges-and-successes-from-the-perspectives-of-frontline-recruiters-and-investigators

46. Ethical Challenges in Community-Based Participatory Research \- AMA Journal of Ethics, https://journalofethics.ama-assn.org/issue/ethical-challenges-community-based-participatory-research

47. Typology of content warnings and trigger warnings: Systematic review \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC9067675/

48. Making Content Usable for People with Cognitive and Learning Disabilities \- W3C, https://www.w3.org/TR/2018/WD-coga-usable-20181211/

49. Web Content Accessibility Guidelines (WCAG) 2.2 \- W3C, https://www.w3.org/TR/WCAG22/

50. The Cognitive and Learning Disabilities Accessibility Task Force (COGA) \- User Vision, https://uservision.co.uk/thoughts/the-cognitive-and-learning-disabilities-accessibility-task-force-coga

51. System Usability Scale (SUS): 2026 Calculator, Formula & Benchmarks Guide \- Koji, https://www.koji.so/docs/system-usability-scale-guide

52. Psychometric evaluation of the System Usability Scale in the context of a childrearing app co-designed for low- and middle-income countries \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12066864/

53. Building the Accessible Usability Scale \- A Walkthrough | Fable, https://makeitfable.com/article/building-the-accessible-usability-scale/

54. Accessible Usability Scale (AUS) \- Fable Accessibility Testing, https://makeitfable.com/accessible-usability-scale/

55. Understanding Signal Detection Theory | PDF | Sensitivity And Specificity | Perception, https://www.scribd.com/document/961449661/Signal-Detection-Theory

56. Good news or bad news? The impact of information valence on high school students' willingness to share misinformation and the effectiveness of a targeted accuracy prompt \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12604475/

57. Understanding communicative intentions in schizophrenia using an error analysis approach, https://pmc.ncbi.nlm.nih.gov/articles/PMC7910544/

58. Distribution of average discernment and skepticism bias scores of... \- ResearchGate, https://www.researchgate.net/figure/Distribution-of-average-discernment-and-skepticism-bias-scores-of-individual-participants\_fig3\_374506186

59. Why is Transparency the Key to AI Compliance under the EU AI Act? \- Open Ethics Initiative, https://openethics.ai/why-is-transparency-the-key-to-ai-compliance-under-the-eu-ai-act/

60. Is it a bot? EU AI Act transparency rules take effect 2 August 2026 | Travers Smith, https://www.traverssmith.com/knowledge/knowledge-container/is-it-a-bot-eu-ai-act-transparency-rules-take-effect-2-august-2026/

61. A guide to the AI Act, the EU's new AI rulebook \- AlgorithmWatch, https://algorithmwatch.org/en/ai-act-explained/

62. NASA-TLX Mental Load Assessment Guide | PDF | Workload \- Scribd, https://www.scribd.com/document/966701246/Mental-Workload-Assessment-NASA-TLX

63. Collab: Fostering Critical Identification of Deepfake Videos on Social Media via Synergistic Annotation \- ResearchGate, https://www.researchgate.net/publication/400083874\_Collab\_Fostering\_Critical\_Identification\_of\_Deepfake\_Videos\_on\_Social\_Media\_via\_Synergistic\_Annotation

64. Collab: Fostering Critical Identification of Deepfake Videos on Social Media via Synergistic Annotation \- arXiv, https://arxiv.org/html/2601.17371v1

65. Full Disclosure, Less Trust? How the Level of Detail about AI Use in News Writing Affects Readers' Trust \- ResearchGate, https://www.researchgate.net/publication/399776774\_Full\_Disclosure\_Less\_Trust\_How\_the\_Level\_of\_Detail\_about\_AI\_Use\_in\_News\_Writing\_Affects\_Readers'\_Trust

66. (PDF) Human-AI attachment: how humans develop intimate relationships with AI, https://www.researchgate.net/publication/400682027\_Human-AI\_attachment\_how\_humans\_develop\_intimate\_relationships\_with\_AI

67. Artificial Intimacy: Companion Artificial Intelligence and Emerging Risks to Adolescent Mental Health \- MDPI, https://www.mdpi.com/2076-0760/15/7/491

68. A Simplified System Usability Scale (SUS) for Cognitively Impaired and Older Adults, https://www.researchgate.net/publication/345095536\_A\_Simplified\_System\_Usability\_Scale\_SUS\_for\_Cognitively\_Impaired\_and\_Older\_Adults

69. Trauma-Informed Therapy \- StatPearls \- NCBI Bookshelf \- NIH, https://www.ncbi.nlm.nih.gov/books/NBK604200/

70. Clinical Issues Across Services \- Trauma-Informed Care in Behavioral Health Services \- NCBI Bookshelf, https://www.ncbi.nlm.nih.gov/books/NBK207185/

71. Something Distressing This Way Comes: The Effects of Trigger Warnings on Avoidance Behaviors in an Analogue Trauma Task \- PubMed, https://pubmed.ncbi.nlm.nih.gov/35473646/

72. Trauma-Informed Care in the Classroom: Our Experience with a Content Warning in a Medical School Course \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC9270545/

73. Study Details | NCT07473336 | The Effectiveness of RAPID-Based Psychological First Aid Training | ClinicalTrials.gov, https://clinicaltrials.gov/study/NCT07473336

74. Psychological First Aid Field Operations Guide \- PTSD.va.gov, https://www.ptsd.va.gov/disaster\_events/for\_providers/PFA/PFA\_2ndEditionwithappendices.pdf

75. Psychological First Aid for post-incident support | health.vic.gov.au, https://www.health.vic.gov.au/worker-health-wellbeing/psychological-first-aid-for-post-incident-support

76. ABCDE Psychological First Aid Application Handbook \- PreventionWeb.net, https://www.preventionweb.net/files/67196\_abcdehandbooken.pdf?startDownload=true

77. Development of Psychological First Aid Guidelines for People Who Have Experienced Disasters \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC8535812/

78. Ethical Dialogue Modeling: Dual-Phase Prompt Design for Safety-Constrained Mental Health Language Models | OpenReview, https://openreview.net/forum?id=GjPD0WkBST

79. New study: AI chatbots systematically violate mental health ethics standards, https://www.brown.edu/news/2025-10-21/ai-mental-health-ethics

80. \[Draft\] Design Guide Overview | Web Accessibility Initiative (WAI) \- W3C on GitHub, https://w3c.github.io/wai-coga/coga-draft/guide/overview

81. Safeguarding Public Trust Thanks to C2PA and Digital Watermarks, https://www.digimarc.com/blog/safeguarding-public-trust-thanks-c2pa-and-digital-watermarks