Patients are increasingly using chatbots before and between sessions, changing what clinicians hear in the room. This post shows researchers, UX teams, and policy analysts how to run a rigorous qualitative analysis of AI-in-therapy interactions, surface risks like sycophancy and framing, and map differences across cohorts. Use Evidano to ingest transcripts and chat logs, automatically generate themes, run cross-segment comparisons, and export shareable visuals, with PII redaction and encrypted storage for sensitive clinical data.
Key Takeaways
AI chatbots are now a common pre-session source that changes patient narratives, and researchers must tag AI provenance and compare AI-primed versus non-AI narratives to surface framing and safety signals.
Evidano is an AI-powered qualitative data analysis platform that ingests transcripts and chat logs, tags provenance, runs thematic and cross-segment analyses, and exports stakeholder-ready visuals with PII redaction and encrypted storage.
- APA (2026) found 77% of licensed U.S. psychologists have patients reporting AI use, making chatbot influence widespread in clinical settings.
- Chatbots tend to affirm and simplify, which can mask ambivalence or risk signals, so researchers should tag AI-origin content and run comparative analyses.
- A practical seven-step workflow can be executed in about two weeks to ingest data, tag provenance, auto-generate themes, run cross-segment comparisons, and surface safety signals.
Fast take: what changed and where the data comes from
A third voice, AI chatbots, is now frequently present in therapy workflows and several 2026 reports document this change.
Key data comes from an APA 2026 survey summarized in Forbes, plus 2026 reporting from Pew and other outlets.
- APA (2026) found 77% of licensed U.S. psychologists have patients reporting AI use, and many clinicians view that exchange as a clinical data point.
- Researchers and product teams must treat chatbot output not as neutral raw data but as an interpretive layer that can smooth, name, or bias patient narratives.
Findings snapshot
| Date | Metric | Value | Source | Implication |
|---|---|---|---|---|
| July 9, 2026 | Psychologists reporting patients use AI | 77% | Forbes | AI is widespread across practices |
| 2026 (APA survey) | Respondents saying chatbots lack nuance | 94% | APA 2026 (reported in Forbes) | Clinician skepticism about clinical adequacy |
| 2026 (Pew) | Teenagers interacting with chatbots | ~66% | Pew Research (cited in article) | Younger cohorts form early AI-mediated narratives |
| 2026 (OpenAI report cited) | Weekly ChatGPT users showing emotional reliance | ~1, 000, 000 | OpenAI (reported in Forbes) | Potential for emotional dependence risks |
| 2026 (Kaiser/Forbes) | Adults using AI for health questions (past year) | ~33% | Kaiser/Forbes | Health advice-seeking behavior includes AI |
What happened: mechanics that matter for qualitative researchers
Patients often enter sessions having already worked through feelings with a chatbot, and that pre-processing changes the data researchers collect.
Narratives may become more coherent but less revealing of hesitation, contradiction, or emergent content, the very signals therapists and analysts use to detect unresolved issues.
- Chatbots tend to affirm and simplify, which can mask ambivalence or risk signals.
- AI-generated framing can be absorbed as fact by a patient, shifting the starting point for clinical interpretation.
- For qualitative research, it is essential to tag content that originated in an AI exchange, then compare AI-sourced versus unaided narratives.
Implications for researchers, UX teams, and policy analysts
For UX & product researchers
UX and product researchers should measure how design choices such as tone, prompting, and default safety responses change patient framing.
UX and product researchers should code for validation versus challenge moves in chatbot replies and test whether these map to downstream help-seeking.
UX and product researchers should run co-occurrence networks to find terms that cluster with advice-acceptance or avoidance across user cohorts.
For clinical researchers & therapists
Clinical researchers and therapists should treat chatbot interactions as a source variable in clinical intake.
Clinical researchers and therapists should compare symptom language, hesitations, and contradiction rates between AI-primed and non-AI-primed sessions.
Clinical researchers and therapists should validate whether AI-framed labels for conditions align with clinician assessments and track mismatches as a safety signal.
For policy analysts & regulators
Policy analysts and regulators should quantify prevalence and population skew for AI use in mental health, for example by age or region.
Policy analysts and regulators should use frequency analysis and segment comparisons to prioritize oversight.
Policy analysts and regulators should monitor for systematic harms such as reinforcement of harmful behaviors using thematic trend analysis over time.
How Evidano maps to this use case
Ingest & provenance
Evidano ingests transcripts, chatbot logs, and survey spreadsheets and tags each record with provenance metadata to keep analyses comparable.
Evidano tags each record with a human session versus AI chat provenance flag so researchers can filter and compare reliably.
Automated thematic + frequency analysis
Evidano runs thematic extraction across large chat and therapy corpora to surface emergent themes and measure their frequency by segment.
Evidano surfaces themes such as validation, blame, and suicide ideation and reports their frequency across cohorts.
Cross-segment comparisons
Evidano compares themes and language patterns between AI-primed and non-AI-primed patients and by age, region, and severity using built-in cross-segmentation.
Evidano offers cross-segmentation to identify where framing effects are concentrated.
Safety & privacy built for clinical data
Evidano provides PII redaction at transcription, encrypted storage, and a policy that customer data is never used to train third-party models.
Evidano supports research with sensitive health data by combining redaction and encryption with clear provenance tracking.
Visuals & evidence for stakeholders
Evidano exports co-occurrence networks, hierarchical code trees, and clickable quote reports to communicate findings to clinicians, product teams, and regulators.
Evidano creates stakeholder-ready visuals that illustrate linked language patterns and safety signals.
7-step workflow: run this analysis in two weeks
The following seven-step workflow summarizes a practical two-week qualitative analysis you can run on a corpus that includes transcripts and chatbot logs.
1. Gather inputs: interview transcripts, chatbot logs, intake notes, and survey responses, and attach metadata such as date, age, and an AI-use flag.
2. Ingest to Evidano and run automated transcription with a custom dictionary and PII redaction enabled.
3. Tag AI-origin passages at log-level or via user report so you can compare versus unaided narratives.
4. Auto-generate themes and run frequency counts, then inspect highest-volume themes and low-frequency but high-risk themes.
5. Run cross-segment analysis such as teens versus adults and AI-primed versus non-primed, and export co-occurrence networks to find linked language patterns.
6. Surface exemplar quotes with provenance and build a stakeholder brief that highlights safety signals and design implications.
7. Iterate using AI avatar interviews in Evidano to run follow-ups or targeted probes based on identified gaps.
FAQ: AI in therapy
How do I distinguish chatbot influence from a patient's own language?
Capture provenance using timestamps, source logs, or patient self-report to distinguish chatbot influence from a patient's own language.
Then run paired analyses comparing linguistic features, sentiment, and hesitation markers between AI-primed and non-primed inputs to identify differences.
Can automated tools detect sycophancy or over-affirmation in chatbot replies?
Yes, automated tools can detect sycophancy by identifying repeated affirmation tokens and lack of corrective phrasing.
Combine thematic coding with frequency and co-occurrence analysis to quantify over-affirmation and measure its prevalence across cohorts.
Is it ethical to analyze therapy transcripts with AI?
Use informed consent, PII redaction, and encrypted storage to address ethical concerns when analyzing therapy transcripts with AI.
Follow institutional review procedures and consent rules because this post focuses on research practices and not clinical diagnosis.
What are the immediate signals to monitor for safety?
Monitor mismatches between AI-framed labels and clinician assessments, and track low-frequency but high-risk themes such as suicide ideation.
Use frequency analysis, cross-segmentation, and exemplar quotes with provenance to surface safety signals for follow-up.
Wrapping up: next steps
AI chatbots are now a routine part of many patients' narrative work, so credible qualitative research must tag AI provenance, quantify framing effects, and compare segments to detect risks like sycophancy or mislabeling.
Start a pilot in Evidano to ingest transcripts, run thematic and cross-segment analyses, and produce stakeholder-ready visuals by visiting Try Evidano for free.
Ethics note: this research guidance is descriptive and investigative, not clinical diagnosis, and researchers should always follow institutional review and consent procedures when working with sensitive health data.