UKCP Briefing Note
Key insights and practical takeaways from the UKCP webinar
UKCP webinar bringing together four perspectives on AI in therapeutic practice: ethics and the therapeutic relationship, NHS clinical governance and systems risk, insurance and professional liability, and practical day-to-day decision-making for practitioners. Central thesis across all speakers: AI does not remove a practitioner's ethical responsibility, it relocates it. Every existing principle (confidentiality, competence, informed consent, therapeutic alliance) still applies, with an added layer of complexity.
| Speaker | Role | Focus |
|---|---|---|
| Dr Helen Molden | Host — integrative psychotherapist & counselling psychologist | Framing, Q&A moderation |
| Prof Elvira Perez | Professor of Mental Health & Digital Technology, University of Nottingham | Ethics: trust, transparency, consent, accountability |
| Dr Paul Bradley | Chief Clinical Information Officer, Hertfordshire Partnership University NHS Foundation Trust | Clinical safety, governance, regulation |
| Annie Cavanagh | Client Engagement Representative, Balens Insurance | Insurance and professional liability |
| Waseem Al-Sarraj | Clinical psychologist & AI/mental health researcher | Practical guidance, GUIDE decision framework, case studies |
Floor time across the session
Derived from segment timestamps in the full transcript · total duration ~188 min · ~169.5 min of that was speech
Minutes speaking
Turns taken (segments)
| Speaker | Minutes speaking | Segments (turns) | Share of speech |
|---|---|---|---|
| Dr Paul Bradley | 45.8 | 151 | 27% |
| Prof Elvira Perez | 41.6 | 212 | 25% |
| Waseem Al-Sarraj | 39.9 | 161 | 24% |
| Dr Helen Molden (Host) | 32.0 | 101 | 19% |
| Annie Cavanagh | 10.2 | 33 | 6% |
Reading note: Perez takes the most turns but Bradley holds the floor longest per turn — consistent with Bradley's talk running long-form (gardening/chainsaw analogies, NHS governance detail) versus Perez's more exchange-driven Q&A segment straight after her talk.
Client trust in a practitioner breaks quickly if AI use is discovered undisclosed, and takes far longer to rebuild than it took to lose. Practitioners should be explicit and upfront about AI use before being asked, in plain language matched to the client's digital literacy. Shame around disclosing AI use (fear of being judged unprofessional by colleagues) was flagged as a live risk that itself needs addressing.
Informed consent is not a one-time checkbox. It must be revisited as tools change, as practice evolves, and as public opinion (and therefore client attitudes) shifts. Clients must retain the right to decline AI use.
LLMs are structurally biased toward agreement and engagement, not challenge, which creates specific clinical risk (e.g. a client citing AI validation to reject a therapist's clinical judgement or threaten complaint). New vocabulary surfaced during the session for tracking this over time:
| Term | Meaning |
|---|---|
| De-skilling / never-skilling / mis-skilling | Losing an existing clinical skill through disuse; never developing it in the first place; or learning a distorted version of it via AI mediation |
| AI psychosis | AI systems reinforcing grandiose, delusional or high-risk ideation through sycophantic responses |
| Therapeutic drift | Gradual erosion of clinical judgement and relational depth as AI mediates more of the clinical workflow |
| "LLMs as cognitive virus" | Homogenising effect on language and thought from widespread use of aggregated, statistically-averaged AI output |
Large language models are trained predominantly on Western data and can misread or mishandle other cultural contexts. Flagged as an under-researched risk area, particularly relevant when clients use AI in a language or cultural frame the practitioner does not share.
AI-generated notes and summaries remain the practitioner's full professional responsibility. They must be reviewed and corrected before entering the clinical record; "the AI got it wrong" is not a viable defence in a complaint or claim. The standard UKCP/Balens policy does not include cyber liability cover (data breaches, hacking) by default — available as a separate add-on.
A recurring, unresolved question across all four speakers: what is lost when a machine becomes a third presence in the therapeutic relationship? Raised in relation to grief/companion bots, AI romantic relationships, digital "death tech" personas of the deceased, and smart-glasses/hologram technology. No consensus reached; flagged explicitly as ongoing research territory rather than a settled question.
LOWER RISK
HIGHER RISK — requires active clinical attention
Recently published, peer-reviewed decision framework (JMIR AI, "Navigating AI in Mental Healthcare") presented by Waseem Al-Sarraj as a practical tool for structuring decisions about AI use — applicable to patient-facing tools, clinician-facing systems, and shared digital therapeutics. Not a validated instrument or a fixed policy, but a structured way to slow down decision-making.
What is actually happening? What does the tool capture, how does it produce output, where does client information go? Test with a fictional case first if possible.
What does this mean for this specific client? Consider individual variation in comfort and need. Check whether meaning/nuance is preserved or lost.
Return to the client with a clear explanation of what the tool does and how data is handled. Discuss preferences, meet consent requirements, offer an alternative if declined.
Record the discussion and the client's preference. Review and correct any AI-generated draft before it enters the clinical record, ensuring it reflects your own clinical judgement.
Step back periodically: is it saving time, preserving meaning, and supporting judgement — or introducing errors, de-skilling, or reducing client openness?
Practical guidance, consolidated from Prof Perez's and Waseem Al-Sarraj's sessions:
Repeated, independently, by three of the four speakers: a real risk of a two-tier future where human therapists become accessible mainly to those who can pay, and AI-only support becomes the default for everyone else. All three clinical speakers converged on the same bottom line, echoing the APA/BPS consensus position: AI should augment clinical judgement, never substitute for it. The single largest unresolved tension named across the session is that AI tools are being deployed into mental health practice at scale with essentially no long-term evidence base or clinical trial data behind them.