AVATAR therapy for voices (Leff, Craig)

1. Introduction

Version followed: the AVATAR2 trial protocol (Garety et al., Trials 2021) and trial report (Garety et al., Nature Medicine 2024), with AVATAR-Extended (initial session plus 12 sessions) as the main version because its authors recommend that future provision be guided mainly by it; AVATAR-Brief (initial session plus 6 sessions) and the AVATAR1 trial manual (Craig et al., Lancet Psychiatry 2018; adapted from a brief guide by Julian Leff) are noted wherever they differ. The original proof-of-concept procedure (Leff et al., British Journal of Psychiatry 2013) is noted where it adds detail.

AVATAR therapy is a blended digital psychological therapy for people with psychosis who hear distressing voices (auditory verbal hallucinations), usually voices that have persisted despite antipsychotic medication. With the therapist, the person builds a computerised face and a matched voice (the avatar) representing the entity they believe is behind their main distressing voice. In sessions the person sits in one room facing the avatar on a screen; the therapist sits in another room, voices the avatar in real time through voice-transforming software, and switches to their own voice to coach and support the person. [source: Leff et al. 2013; Craig et al. 2018; Garety et al. 2024]

Core idea of change: the person faces the voice's actual words in a setting where it cannot hurt them, practises standing up to it, and experiences the formerly all-powerful voice gradually conceding power, acknowledging the person's strengths and becoming less hostile. Proposed mechanisms in the developers' accounts include exposure with expectancy violation and dropping of safety behaviours, a shift in perceived power and control, strengthening self-esteem, linking voice content to the person's history and self-beliefs, and (Leff's account) re-integrating an externalised, rejected part of the self once the avatar becomes benign. [source: Leff et al. 2013; Leff, HKU presentation slides; Ward et al. 2020; Craig et al. 2018]

This is a licensed, software-based treatment, not a technique that can be improvised. Its software (the Avatar Therapy System) is described by the AVATAR2 authors as registered as a Class 1 medical device, and NICE says the technologies in its 2024 early value assessment must be delivered or supported by a mental health professional trained in the technology. A psychotherapist wanting to start needs access to the system, AVATAR training and supervised cases (see sections 3 and 4). [source: Garety et al. 2024; NICE HTE17]

Regulatory and evidence status (as of October 2026 per the NICE HTE17 recommendations page and the KCL AVATAR-RW project page):

NICE early value assessment HTE17 (reported in March 2024 [uncertain: exact publication date not verified]) says AVATAR Therapy can be used in the NHS for adults with psychosis to help manage auditory verbal hallucinations as an option while more evidence is generated, for 3 years; conditions include delivery or support by a trained mental health professional, appropriate regulatory approval including NHS England's DTAC (Digital Technology Assessment Criteria), evidence-generation agreements with annual confirmation to NICE, suitability assessment, and people's right to decline digital options. For young people, NICE says more research is needed and access should be through research, company or non-core NHS funding. NICE notes it may withdraw the guidance if evidence-generation conditions are not met; the page reviewed showed no withdrawal. [source: NICE HTE17, Recommendations; NICE HTE17, The technologies]
NICE describes the technology as 6 to 12 sessions in which a mental health professional voices the avatar in a three-way conversation, usable alone or alongside standard care such as CBT for psychosis (CBTp). [source: NICE HTE17, The technologies]
Evidence, attributed: AVATAR1 (150 participants, South London and Maudsley; versus supportive counselling) found a larger reduction in overall voice severity (PSYRATS auditory hallucinations total) at 12 weeks, reported by the authors as a large effect (d = 0.8), with the between-group difference no longer significant at 24 weeks as the counselling group continued to improve. AVATAR2 (345 participants, four UK centres; versus treatment as usual) found reduced voice-related distress at 16 weeks for both versions, with only AVATAR-Extended meeting the authors' prespecified clinically significant threshold; differences in distress were not significant at 28 weeks, while AVATAR-Extended reduced voice frequency at both 16 and 28 weeks. AVATAR has not yet been compared head to head with CBTp. [source: Craig et al. 2018; Garety et al. 2024]
Real-world evaluation: King's College London lists AVATAR-RW, a Wellcome-funded project to test AVATAR in routine NHS pathways (September 2026 to August 2029, ongoing), with NHS partners in Manchester, South London, North East London and Glasgow. [source: KCL AVATAR-RW page]
Developer: the therapy was developed by researchers at the King's College London Institute of Psychiatry, Psychology and Neuroscience, UCL and Avatar Therapy Ltd, a company whose CEO is reported as Professor Mark Huckvale; AVATAR2 authors declared shareholding or unpaid advisory roles with the company. [source: KCL news, 26 April 2024; Garety et al. 2024]
Outside the UK: the AVATAR2 authors report that immersive virtual-reality variants are being tested in Germany, Denmark and Canada; these are separate research versions. [source: Garety et al. 2024] No FDA, DiGA or other non-UK approval was found [uncertain: not verified].

2. Framework

Who it is for (trial criteria, the best available indication of suitability). Adults (18 and over) under a specialist mental health team, with frequent, distressing voices for at least 6 months (AVATAR2; at least 12 months and medication-unresponsive or partly responsive in AVATAR1), and a diagnosis of schizophrenia-spectrum disorder or an affective disorder with psychotic symptoms. Trial exclusions: primary substance, personality or learning disorder; lacking capacity to consent; profound visual or hearing impairment; insufficient English for therapy (voices also had to be in a language the therapist speaks in AVATAR1); currently receiving individual psychological therapy for voices; current acute mental health crisis. AVATAR1 also excluded organic brain disease and refusal of medication. [source: Garety et al. 2021; Garety et al. 2024; Craig et al. 2018] Age: the trials were adults only; NICE limits young people's access to research settings. [source: NICE HTE17]

Format. Individual sessions, weekly, about 45 to 60 minutes, typically in a clinic; the person and therapist sit in separate private rooms linked by the AVATAR software, with the therapist seeing the person by webcam. In AVATAR2, 87% received in-person treatment; a bespoke end-to-end encrypted video version was used for remote delivery during COVID-19 restrictions. The live avatar dialogue itself is short: about 5 minutes early on, rising to 10 to 15 minutes; mean dialogue length in AVATAR2 was about 10 minutes. [source: Garety et al. 2021; Garety et al. 2024; Ward et al. 2020]

Every session has three parts: pre-dialogue discussion (review of the week, agreeing the dialogue's focus and themes, role-play rehearsal of expected challenges); the active avatar dialogue; and a post-dialogue debrief (ratings, reflection on strengths and emerging content). [source: Garety et al. 2024; Ward et al. 2020; Garety et al. 2021]

Stages

Stage	Sessions (AVATAR2)	Goals	Move on when
Initial session: assessment and avatar creation	1 session before therapy	Detailed voice assessment (verbatim content, characterisation, the person's relationship with the voice, developmental history); build face and voice; orient to the setup	Person has a 'good enough' avatar of the main (most distressing) voice and consents to proceed
Phase 1: exposure and assertiveness	Brief: most of 6 sessions; Extended: early sessions (Ward et al. describe Phase 1 as typically sessions 1 to 2 in AVATAR1; Craig et al. 2018 as sessions 1 to 3)	Exposure to verbatim voice content; assertive responding; drop safety behaviours; reduce fear; begin power shift; self-esteem	The avatar has conceded ground in step with the person's growing sense of power; the concession signals the move to Phase 2
Phase 2 (Extended): formulation-driven work	Remaining sessions to 12	Developmental understanding of the voice within the person's life and relationships; flexible dialogues on power and malevolence beliefs, self-concept, trauma, the relationship with the voice; future focus	Final session reached; goals and future plans set
AVATAR-Brief later sessions	Up to 6	Standardised focus on exposure, assertiveness, power and control, self-esteem, session by session from a manual	End of 6 sessions
Ending	Final session	Future planning; goodbye to the avatar; therapy summary letter	Therapy complete

[source: Garety et al. 2021; Garety et al. 2024; Ward et al. 2020; Craig et al. 2018]

Session number. AVATAR2 allowed session numbers to vary by up to two, by therapist judgement in agreement with the person; AVATAR1 allowed up to three extra where benefit was likely. The AVATAR2 authors suggest a collaborative review of session number in Extended, to keep the person's sense of control during trauma-focused work. [source: Garety et al. 2021; Garety et al. 2024; Craig et al. 2018]

Progress criteria. In AVATAR1, progress through sessions was judged by discussing changes in the voices' severity, malevolence or frequency; therapy could end early if the person reported no voices for three consecutive sessions. [source: Craig et al. 2018] After each dialogue in AVATAR2, the person rates the avatar's hostility, their anxiety and the sense of the voice's presence. [source: Garety et al. 2021] Using these ratings to time the power shift is [synthesis].

Discontinuation. Therapy stops if the person asks, or if the therapist judges with the supervisory team that there has been significant mental-health deterioration (AVATAR2), or if it is linked to significantly increased distress or risk of harm to self or others (AVATAR1). In AVATAR1, non-attendance at three consecutive sessions also ended therapy. [source: Garety et al. 2021; Craig et al. 2018]

3. Guiding principles

[source for principles 1 to 9: Leff et al. 2013; Leff, HKU slides; Ward et al. 2020; Garety et al. 2021; Garety et al. 2024; Craig et al. 2018. Practical meanings are synthesis where not stated.]

Validation through co-creation. Building the avatar with the person acknowledges their experience as real to them. Practical meaning: take the description of the voice seriously and in detail; never argue about whether the voice exists.
Safe space for risk-taking. The person made the avatar, so it cannot harm them or their family. Practical meaning: say this explicitly before the first dialogue and return to it whenever fear of retaliation appears.
The person's power and control is the central target. This was clearly present in 100% of AVATAR1 completers' therapy records reviewed by Ward et al. Practical meaning: every dialogue should leave the person feeling more in charge than before it.
Calibrated concession. The avatar starts hostile, even exaggerating its power, and concedes only in step with the person's growing strength. Practical meaning: a concession must feel earned; too early feels false (one AVATAR1 participant found the early shift out of keeping with their voices), too late keeps the person stuck.
Use the voice's real words. Verbatim content gives practice in reclaiming power over the words that silenced the person; commands to harm self or others are excluded from what the avatar says (Ward et al. 2020). Practical meaning: collect exact phrases at assessment and agree which will be used.
End every dialogue on a win. Practical meaning: close each dialogue with a self-affirming statement from the person or a concession from the avatar, never mid-attack.
Two distinct roles, cleanly signalled. The therapist is both the voice and the person's ally. Practical meaning: always mark the switch (software channel and wording), and never leave the person unsure who is speaking.
Formulation guides Phase 2. Practical meaning: in Extended, link voice content to the person's life story (trauma, bullying, discrimination, exclusion, grief, self-beliefs) and choose dialogue targets from that formulation, tentatively and without invalidating.
Sensitivity to emotional hotspots. Ward et al. note that Phase 1 can be especially challenging and that invalidation must be avoided, especially in internal-attribution, compassion and avatar-transition work. Practical meaning: pace intensity to the person, check in often, and keep the dialogue short when needed.
Generalise to daily life. Recordings were given so the dialogue could be replayed when voices troubled the person; maintenance processes and future plans are explicit targets. Practical meaning: every session ends with a plan for using what was gained before the next session.
Embedded in the clinical team. Trial therapy ran alongside usual care from specialist mental health teams, with adverse-event monitoring and a summary letter to the responsible clinician. Practical meaning: AVATAR is never delivered in isolation from the person's care team, prescriber and crisis plan. [source: Craig et al. 2018; Garety et al. 2021]
Collaboration and choice. People can decline digital options (NICE); session number is reviewed collaboratively in Extended (Garety et al. 2024). Practical meaning: consent is revisited before each new phase, not taken once.

4. Intervention inventory

The published sources describe the structure, targets and roles, but the session-by-session manuals are not public. Steps below follow the published descriptions; detailed sequencing within a step is marked [synthesis]. All example language is illustrative unless quoted and cited. AVATAR trial therapists were mostly clinical or counselling psychologists, plus psychiatrists and nurse therapists, with substantial psychosis experience, trained in AVATAR and supervised (see 4.20). [source: Garety et al. 2024]

4.1 Referral, eligibility and suitability check

Purpose / when: Before any assessment, confirm the person fits the population in which AVATAR was tested and that it fits their care plan. [source: Garety et al. 2021; NICE HTE17]
Steps:
1. Confirm the person is an adult under a specialist mental health team, with persistent frequent distressing voices and a psychosis diagnosis as in section 2.
2. Check exclusions used in trials: lack of capacity to consent, acute crisis, primary substance, personality or learning disorder, profound sensory impairment, language mismatch, concurrent individual therapy for voices. [source: Garety et al. 2021]
3. Liaise with the care coordinator and responsible clinician: current risk history, crisis plan, medication is the prescriber's domain and is not altered as part of AVATAR. [synthesis]
4. Tell the person that digital therapy is an option they may decline, and what the alternatives are (for example CBTp). [source: NICE HTE17 on declining; alternatives are synthesis]
Example language (illustrative): "This is a talking therapy where you get to face the voice on a screen, in a safe setting, with me helping you. It is one option; CBT for voices is another. You can say no, or stop at any point, and your usual care carries on."
Watch for: current crisis, recent admission, active risk to others linked to voice content, or the person agreeing only to please the team; these call for postponing and team discussion. [synthesis]

4.2 Voice assessment

Purpose / when: Initial session. Gathers what the avatar will say and how it will sound, and starts the formulation. [source: Garety et al. 2021; Craig et al. 2018]
Steps:
1. Map all voices; agree to focus on the dominant or most distressing voice (or the one the person most wants to be rid of). [source: Leff et al. 2013; Ward et al. 2020]
2. Record verbatim content: typical phrases, threats, insults, commands. Note which phrases are most distressing. [source: Craig et al. 2018; Garety et al. 2021]
3. Characterise the voice: identity the person attributes to it, gender, age, accent, tone, perceived power and intent. [source: Garety et al. 2021]
4. Explore the voice-hearer relationship: how the person responds now (complying, appeasing, arguing, avoiding), what they fear will happen if they stand up to it. [source: Garety et al. 2021 for relationship; specific questions are synthesis]
5. Take a developmental history relevant to the voice: onset, trauma, bullying, discrimination, losses, people it resembles. [source: Garety et al. 2021; Garety et al. 2024]
6. Use standard measures as the service requires; the trials used validated voice-rating instruments (PSYRATS, BAVQ-R, VAAS); describe and use these under their own licences. [source: Craig et al. 2018; Garety et al. 2024]
Example language (illustrative): "I'd like to hear the actual words the voice uses, even though they are unpleasant. Saying them here is safe. Which line upsets you most? What do you think would happen if you answered back?"
Watch for: distress when repeating content (slow down, allow writing it instead), fear the voice will punish disclosure, content that indicates current risk (commands to harm self or others: assess and share per local risk policy; such commands are not voiced by the avatar). [source for not voicing commands: Ward et al. 2020; risk steps are synthesis]

4.3 Orientation, rationale and consent

Purpose / when: Initial session and before the first dialogue. Informed consent, and preparing the person for an unusual setup. [source: Garety et al. 2021; Leff, HKU slides]
Steps:
1. Explain the two rooms, the screen, the two voices the therapist will use, and that sessions are recorded (for therapy and supervision) and recordings will be given to them. [source: Garety et al. 2021; Craig et al. 2018]
2. Explain the plan: early dialogues with a hostile avatar, which changes as the person gains control. [source: Garety et al. 2021]
3. Stress the safe-space point: the avatar is made by the person and cannot harm them or anyone they love. [source: Leff et al. 2013; Leff, HKU slides]
4. Explain that they can stop any dialogue at any moment, and how. [synthesis; the 2013 pilot had a stress button, see 4.17]
5. Obtain and document consent, including to recording and to sharing a summary with the care team. [source: Garety et al. 2021; Craig et al. 2018]
Example language (illustrative): "You'll hear me in two ways: as the avatar, through the computer, and as myself, coaching you. I'll always tell you when I'm switching. If at any point you want to stop, just say 'stop' and I'll come straight back to you as myself."
Watch for: belief that the avatar is the actual entity or will give it access to them; fear of repercussion. Re-explain, and agree a stop signal. [synthesis]

4.4 Creating the avatar with the person

Purpose / when: Initial session. Produces a face and voice close enough to the person's experience to evoke the voice, and validates their experience by building it together. [source: Leff et al. 2013; Garety et al. 2024; Leff, HKU slides]
Steps:
1. Face: Using the software's face generator (FaceGen in the published systems), the person chooses and adjusts face shape, colour, texture and hair; non-human characters are available. If the person cannot picture a face, they choose one they feel able to talk to. [source: Garety et al. 2024; Leff et al. 2013]
2. Voice: Record a sample; the software alters pitch and other voice characteristics (in AVATAR2, pitch, vocal tract size, spectral tilt, roughness) and the person adjusts sliders until the match is acceptable. In the 2013 pilot, face choice took about 15 minutes and voice about 25; in AVATAR2 avatar building took about 20 minutes. [source: Leff et al. 2013; Garety et al. 2024]
3. Aim for adequacy, not perfection: the AVATAR2 authors aim for "a good enough representation of the voice for the person" (Garety et al. 2024), about 70% being acceptable.
4. Test a few neutral phrases through the avatar so the person hears it live before any hostile content. [synthesis]
5. Check comfort: if the face itself is intolerable (in the pilot, one man who had been sexually abused could not look at it), the face can be removed and the dialogue run by voice alone. [source: Leff et al. 2013]
Example language (illustrative): "Let's get it close enough that it feels like talking to the voice, but it doesn't have to be perfect. Tell me when the eyes look right. Now listen: is the voice deeper, rougher, older?"
Watch for: rising fear as the likeness improves (pause, ground, remind the person they control it); perfectionism that stalls the session; the person's sense of achievement in having built it, which is itself part of the validating work. [source for validation: Leff, HKU slides; rest synthesis]

4.5 Pre-dialogue planning and rehearsal

Purpose / when: Every session, before the dialogue. Reviews the week and prepares the person for what the avatar will say. [source: Garety et al. 2024; Ward et al. 2020]
Steps:
1. Review the past week: voice activity, use of the recording, any change in how they respond to the voice. [source: Garety et al. 2024]
2. Agree the dialogue's focus and which verbatim lines will be used. [source: Ward et al. 2020; Garety et al. 2024]
3. Role-play the challenges the person expects (the therapist plays the voice in the same room, briefly), and draft responses together. [source: Garety et al. 2024 for role-play; drafting is synthesis]
4. Agree how this dialogue will end (the win) and the stop signal. [synthesis based on Ward et al. 2020's 'end on a win']
Example language (illustrative): "Last week the voice kept saying you are worthless. Today, what do you want to say back? Let's try it here first: I'll say the line, you answer."
Watch for: the person scripting only appeasing responses; help them find a firmer stance. Overlong planning that leaves no time for the dialogue. [synthesis]

4.6 Phase 1 dialogue: exposure to verbatim content

Purpose / when: Early sessions. Exposure to the voice's actual words in a safe setting, producing expectancy violation (the threatened catastrophe does not happen), dropping safety behaviours and reducing conditioned fear. [source: Ward et al. 2020; Craig et al. 2018]
Steps:
1. The therapist goes to the other room and opens the dialogue as the avatar, using agreed verbatim content (critical, abusive, hostile), voicing the avatar as a characterised entity; commands to harm self or others are excluded. [source: Ward et al. 2020]
2. In early sessions keep the dialogue short (around 5 minutes), lengthening to 10 to 15 minutes as tolerance grows. [source: Ward et al. 2020]
3. The avatar initially presents itself as powerful, even exaggerating its power. [source: Ward et al. 2020]
4. Watch the person by webcam throughout and switch to your own voice whenever coaching or support is needed (4.7). [source: Garety et al. 2024; Leff et al. 2013]
Example language (illustrative, as the avatar, using the person's agreed content): "You know you're useless. Nobody wants you around."
Watch for: signs of being overwhelmed (freezing, looking away, dissociative signs, tearfulness): switch to own voice, lower intensity or end on a small win. Too little arousal (person laughing it off) may mean the content is not the real one. All withdrawals in AVATAR1 occurred in Phase 1. [source for withdrawals: Ward et al. 2020; responses are synthesis]

4.7 Therapist coaching in own voice: assertive responding

Purpose / when: During every dialogue. The therapist's second role: ally and coach who helps the person stand up to the avatar. [source: Leff et al. 2013; Craig et al. 2018; Ward et al. 2020]
Steps:
1. Switch channel to your own voice (the console switches between avatar and therapist voice; name the switch). [source: Garety et al. 2021; naming is synthesis]
2. Prompt assertive replies: telling the avatar they will no longer accept its threats, challenging its claims and apparent misconceptions, calling its bluff about its power. [source: Craig et al. 2018; Ward et al. 2020]
3. Reinforce wins and reflect the person's strong statements back. [source: Ward et al. 2020]
4. Switch back to the avatar and let it respond to what the person said, adapting in real time. [source: Craig et al. 2018]
Example language (illustrative, own voice): "That's it, you said it clearly. Say it again, louder: you won't take orders from it any more." (illustrative, as the avatar responding) "I didn't expect you to talk to me like that."
Watch for: the person looking to the therapist for permission (gradually hand over: prompt less); responses becoming aggressive or abusive rather than assertive (coach toward calm firmness); the person growing visibly taller in posture and voice, a sign it is working. [synthesis]

4.8 Ending each dialogue on a win

Purpose / when: End of every dialogue. Consolidates a sense of power. [source: Ward et al. 2020]
Steps: before closing, let the person make a self-affirming statement, or have the avatar concede a point; then close the dialogue and return to the person's room. [source: Ward et al. 2020; sequence synthesis]
Example language (illustrative, own voice): "Before we finish, tell it one thing that is true about you."
Watch for: ending on an attack when time runs out; protect time for the win. [synthesis]

4.9 The power shift: calibrated concession of the avatar

Purpose / when: Toward the end of Phase 1 and through Phase 2. The avatar gradually concedes ground, acknowledges the person's strengths and becomes less hostile; in Leff's original description it becomes supportive and offers advice, its expression changing from grim to smiling. The concession marks the move to Phase 2. [source: Ward et al. 2020; Garety et al. 2021; Leff, HKU slides; Leff et al. 2013]
Steps:
1. Time the shift to the person's growing power, using observation and the post-dialogue ratings. [source: Ward et al. 2020; ratings Garety et al. 2021]
2. Concede in small steps: first acknowledge the person's resistance, then reduce threats, then acknowledge strengths. [step sizes synthesis]
3. Keep the character consistent: the avatar remains recognisably the same entity as it changes. [synthesis]
Example language (illustrative, as the avatar): "You're not as easy to push around as I thought." Later: "Maybe I've been wrong about you."
Watch for: a shift that feels fake to the person (one AVATAR1 participant found it out of keeping with their voices); if so, step back to earlier stance and let the person earn more. [source: Ward et al. 2020; response synthesis]

4.10 Post-dialogue ratings and debrief

Purpose / when: After every dialogue. Monitors response and consolidates learning. [source: Garety et al. 2021; Garety et al. 2024]
Steps:
1. Return to the person's room and ask them to rate the avatar's hostility, their own anxiety and the sense of the voice's presence (on screen or in a booklet). [source: Garety et al. 2021]
2. Discuss what they did well and what content emerged. [source: Garety et al. 2024]
3. Check current distress and safety before they leave. [synthesis]
4. Agree between-session use of the recording (4.15). [synthesis]
Example language (illustrative): "How hostile did it feel, from 0 to 10? And your anxiety? What did you notice when you told it to stop?"
Watch for: high residual anxiety or a sense of the voice still present in the room: stay longer, ground, and arrange contact before the next session. [synthesis]

4.11 Self-esteem work: qualities from a significant other

Purpose / when: Phase 1 onward (AVATAR1 explicitly shifted to self-esteem in later sessions; AVATAR-Brief keeps it as a standard focus). Counters the avatar's denigrating messages; clearly present in 96% of AVATAR1 completers' therapy. [source: Ward et al. 2020; Craig et al. 2018; Garety et al. 2021]
Steps:
1. Between sessions or in pre-dialogue, the person gathers a list of their qualities as seen by someone who values them. [source: Ward et al. 2020]
2. In dialogue, the avatar voices the person's negative self-beliefs; the person answers using the list. [source: Ward et al. 2020]
3. The avatar's stance changes as the person draws on others' views, eventually acknowledging those qualities. [source: Ward et al. 2020]
Example language (illustrative, own voice): "Your sister said you are loyal and funny. Tell it that." (illustrative, as the avatar) "Well... perhaps you do care about people."
Watch for: the person unable to name anyone (use staff, past teachers, peers, or the therapist's honest observations [synthesis]); the person dismissing the list as untrue, which suggests the need for slower work on self-concept. [synthesis]

4.12 Formulation-driven Phase 2 (AVATAR-Extended)

Purpose / when: After the power shift, in Extended. Builds a developmental understanding of the voice within the person's life and relationships and uses it to choose dialogue targets. [source: Garety et al. 2021; Garety et al. 2024; Ward et al. 2020]
Steps:
1. In pre-dialogue time, map with the person how the voice's content relates to their history (abuse, bullying, discrimination, social exclusion, losses) and to their beliefs about self, others and the voice's power and malevolence. [source: Garety et al. 2024; Garety et al. 2021]
2. Agree, from that map, which of the Phase 2 targets (4.13) to address next, and the order. [synthesis]
3. Continue the focus on power, control and autonomy throughout. [source: Ward et al. 2020]
4. Review session number collaboratively, to preserve the person's sense of control in trauma-related work. [source: Garety et al. 2024]
Example language (illustrative): "The voice says you're dirty and to blame, and those were your stepfather's words too. Would it help to use our dialogues to answer that?"
Watch for: links offered too early, which can feel invalidating; affective changes linked to trauma work (the AVATAR2 authors attribute part of Extended's higher adverse-event count to these, as well as more sessions and monitoring). [source: Ward et al. 2020; Garety et al. 2024]

4.13 Phase 2 dialogue targets

Ward et al. (2020) coded these targets in AVATAR1 therapy records; frequencies are the proportion of 53 completers with clear evidence. Each is delivered in the same pre-dialogue / dialogue / debrief frame.

a. Maintenance processes (51%)

Purpose: Address what keeps distress going, such as inactivity, avoidance, isolation and worry. [source: Ward et al. 2020]
Steps: the avatar names these drivers in dialogue; the debrief turns them into real-world changes to try. [source: Ward et al. 2020; debrief planning synthesis]
Example language (illustrative, as the avatar): "I'm stronger when you stay in your room all day." (own voice) "What would it mean to go to the café anyway this week?"
Watch for: planned changes being too large; set small, specific steps. [synthesis]

b. Internal attribution (26% clear, 19% partial)

Purpose: Where the person is open to it, link the voice to the self, for example to self-critical thoughts or fears. [source: Ward et al. 2020]
Steps: tentatively, the avatar presents itself as a part of the person that reflects their fears; Ward et al. quote an avatar line, "I voice your deepest fears". Proceed only as the person accepts it. [source: Ward et al. 2020]
Example language (illustrative, own voice): "Some people find their voice says the things they secretly fear about themselves. Does that fit at all, or not for you?"
Watch for: any sign the person feels disbelieved or blamed; drop it. Leff's slides report one participant saying "what the voice says is what I think about myself." [source: Ward et al. 2020; Leff, HKU slides]

c. Identity and social inclusion (51%)

Purpose: Address being made to feel 'other' and chronic shame, including discrimination. [source: Ward et al. 2020]
Steps: in dialogue, rescript experiences of discrimination or exclusion, with the person asserting their place and worth. [source: Ward et al. 2020]
Example language (illustrative, own voice): "It's using the same words people used at school about your skin colour. What do you want to say to that now?"
Watch for: the therapist's own cultural assumptions; let the person define their identity. [synthesis]

d. Compassion toward the voice (19%)

Purpose: Identify positive functions the voice may serve (for example protection) and allow empathic responses to it. [source: Ward et al. 2020]
Steps: explore in pre-dialogue whether the voice has ever warned or protected; in dialogue the avatar can acknowledge that intent; the person can respond with understanding while keeping boundaries. [source: Ward et al. 2020; sequence synthesis]
Example language (illustrative, as the avatar): "I was trying to keep you safe, back then." (own voice) "You can thank it and still tell it you're in charge now."
Watch for: compassion work being experienced as minimising abuse; Ward et al. flag invalidation risk here. [source: Ward et al. 2020]

e. Experiential disengagement (36%)

Purpose: Practise staying calm and not responding to taunts, then apply this in daily life. [source: Ward et al. 2020]
Steps: the avatar provokes; the person responds neutrally or with silence; the debrief plans using the same stance with the real voice. [source: Ward et al. 2020]
Example language (illustrative, own voice): "This time, see if you can just notice it and not get pulled in. Breathe, and let it talk."
Watch for: disengagement sliding into avoidance or dissociation; check that the person remains present. [synthesis]

f. Working with grief (11% clear, 9% partial)

Purpose: For voices tied to someone who has died: address unresolved grief, rescript conflicts, express love and forgiveness. [source: Ward et al. 2020]
Steps: formulation of the relationship; dialogues in which conflicts are rescripted and unsaid things said. Ward et al. describe a man hearing his deceased father's voice, whose dialogue reframed kindness as strength; he later visited his father's grave and reported no longer hearing the voice. [source: Ward et al. 2020]
Example language (illustrative, own voice): "Is there something you never got to say to him? You can say it now."
Watch for: intense grief reactions; allow time after the dialogue and involve the care team if mood drops. [synthesis]

g. Working with trauma (36% clear, 38% partial)

Purpose: Address shame and self-blame tied to trauma, using 'then-now' comparisons. [source: Ward et al. 2020]
Steps: a phrase from a past abuser, if it is part of voice content, may be voiced, with the person responding from present-day strength; then-now comparisons are drawn in debrief. [source: Ward et al. 2020] Trauma processing beyond this is not described in the sources [uncertain]; if the person has prominent PTSD symptoms, discuss with the team whether trauma-focused therapy is needed separately. [synthesis]
Example language (illustrative, own voice): "Then you were a child and had no choice. Now you're an adult in this room. Tell it what you know now."
Watch for: flashbacks, dissociation, shame spikes; stop the dialogue, ground, debrief; see section 5. AVATAR2 found no effect on trauma symptom scores (ITQ). [source: Garety et al. 2024; responses synthesis]

h. Future focus (85%)

Purpose: Plan work, study and relationships; usually the theme of the final session. [source: Ward et al. 2020]
Steps: the person states their goals to an avatar that is now supportive; the avatar may endorse them. [source: Ward et al. 2020; Leff, HKU slides for the supportive, advice-giving avatar]
Example language (illustrative, as the avatar): "You've shown you can stand up to me. What are you going to do with that?"
Watch for: goals that are other people's rather than the person's own. [synthesis]

4.14 AVATAR-Brief session plan (shorter version)

Purpose / when: Where the Brief version is used: six manualised sessions after the initial session, with a standardised focus on exposure to verbatim content, assertiveness, power and control, and self-esteem (4.5 to 4.11), without the formulation-driven Phase 2. [source: Garety et al. 2021; Garety et al. 2024]
Steps: follow 4.5 to 4.11 across the six sessions; the AVATAR1 version typically placed exposure and assertiveness in sessions 1 to 3 and the conceding avatar and self-esteem in sessions 4 to 6. [source: Craig et al. 2018]
Example language (illustrative): "In these six sessions we'll focus on you taking power back from the voice and on how you see yourself."
Watch for: the authors note Brief did not meet their clinical-significance threshold for distress, while completion was higher (82% versus 58% for Extended); choice of version is a clinical and service decision. [source: Garety et al. 2024]

4.15 Session recordings and between-session listening

Purpose / when: After every session. Gives the person 'the therapist in their pocket' to replay when voices trouble them or when they feel bad about themselves; the 2013 authors describe it as "a therapist in their pocket" and suggest it may help explain continued improvement after therapy, though use was not measured. [source: Leff et al. 2013; Leff, HKU slides]
Steps:
1. Record each session or dialogue (with consent) and give the person the recording; AVATAR1 provided an MP3 player, AVATAR2 provided a listening device if needed. [source: Craig et al. 2018; Garety et al. 2021]
2. Agree when to listen: especially when the voice is active. [source: Craig et al. 2018]
3. Suggest listening to the parts where they win and the avatar concedes, rather than the early hostile content [synthesis].
4. Review use at the next pre-dialogue. [source: Garety et al. 2024 for weekly review; recording review synthesis]
Example language (illustrative): "When the voice starts up this week, play the bit where you told it you're not afraid any more."
Watch for: the AVATAR2 authors report engagement with recordings was variable and could be improved. Listening to hostile sections alone may raise distress; who else could hear the recording at home (privacy). [source: Garety et al. 2024; rest synthesis]

4.16 Between-session contact

Purpose / when: As needed. AVATAR2 recorded phone or message contact between sessions as a therapy process activity. [source: Garety et al. 2021]
Steps: agree in advance what contact is available, how quickly it is answered, and that it is not a crisis line; give the crisis pathway (4.19). [synthesis]
Example language (illustrative): "You can text me to rearrange or if something important comes up. I reply within working hours. If you feel unsafe, call the crisis number on your plan."
Watch for: contact replacing sessions, or signs of deterioration in messages. [synthesis]

4.17 In-session safety exit

Purpose / when: Any dialogue. In the 2013 pilot, a red stress button made the avatar disappear and replaced it with a calming beach scene and music. [source: Leff et al. 2013] Whether current systems include an equivalent is not stated in the sources reviewed [uncertain: not verified].
Steps: show the person the exit (button, if available, and the agreed stop word); when used, switch immediately to own voice, return to the person's room if needed, ground, and decide together whether to resume. [synthesis]
Example language (illustrative, own voice): "I've stopped the avatar. You're in the clinic with me. Let's breathe for a moment, then you decide what's next."
Watch for: frequent use early on: shorten dialogues and soften content; never push past a stop. [synthesis]

4.18 Ending therapy

Purpose / when: Final session and planned early endings. Consolidates gains and hands over to the care team. [source: Ward et al. 2020; Craig et al. 2018]
Steps:
1. Final session focuses on the future (4.13h). [source: Ward et al. 2020]
2. Allow for feelings about ending: Ward et al. note some people feel sad saying goodbye to the avatar. [source: Ward et al. 2020]
3. Write a therapy summary letter to the person, copied to the responsible clinician (AVATAR1). [source: Craig et al. 2018]
4. Plan how the person will keep using the recordings and their new responses to the voice; agree what happens if voices worsen. [synthesis]
5. Early ending criteria used in AVATAR1: no voices for three consecutive sessions; recommendation to stop by the person, therapist or supervision group; three consecutive missed sessions. [source: Craig et al. 2018]
Example language (illustrative): "Today is our last dialogue. What do you want to say to the avatar before we finish? And what will you do if the voice tries to take over again?"
Watch for: fading gains after therapy: in both trials, between-group differences were smaller and non-significant at later follow-up for most outcomes; AVATAR1 had no monitoring between end of therapy and follow-up. Arranging follow-up through the care team, or booster contact, is [synthesis]. [source: Craig et al. 2018; Garety et al. 2024]

4.19 Risk, adverse-event monitoring and crisis pathway

Purpose / when: Throughout. The trials used structured adverse-event monitoring; services adopting AVATAR need an equivalent. [source: Garety et al. 2021; Craig et al. 2018]
Steps:
1. Before starting, confirm the person's crisis plan and contacts in their specialist team; check they are not in acute crisis. [source: Garety et al. 2021 for exclusion; crisis plan synthesis]
2. Monitor and record adverse events at each session; trial categories included hospital admission, home treatment or crisis team involvement, self-harm, suicide attempts, violent incidents, and (AVATAR2) events related to remote delivery. [source: Garety et al. 2021; Craig et al. 2018]
3. Grade events (mild, moderate, severe) and report serious events promptly to the clinical team and governance route; the trials reported them to the chief investigator and an independent data monitoring and ethics committee for relatedness judgements. [source: Garety et al. 2021]
4. Stop therapy if the person asks, or if, with supervisors, you judge there has been significant deterioration or increased risk to self or others. [source: Garety et al. 2021; Craig et al. 2018]
5. Use standard suicide and violence risk assessment per local policy whenever voice content, mood or behaviour change; share with the care team the same day where risk is raised. [synthesis]
Example language (illustrative): "Since we started, have the voices pushed you to hurt yourself or anyone else? Have you had thoughts of ending your life? I'll need to share this with your care coordinator so we can keep you safe."
Watch for, as reported: AVATAR1 recorded no adverse events attributable to therapy over 24 weeks. In AVATAR2, 58 serious adverse events occurred in 56 participants across all arms, mostly hospital admissions distributed similarly across arms; two deaths (one suicide) in the Extended arm were judged unrelated to treatment; six events in five participants were rated possibly related, mostly admissions near the end of treatment. Extended had more adverse events overall, which the authors attribute partly to more sessions and monitoring and partly to affective changes with trauma-focused work. Risk around the end of therapy and during trauma work therefore merits particular attention [synthesis]. [source: Craig et al. 2018; Garety et al. 2024]

4.20 Training, supervision and fidelity

Purpose / when: Before and during delivery. NICE requires delivery or support by a professional trained in the technology. [source: NICE HTE17]
Steps (as published for trial therapists):
1. Background: AVATAR2 therapists were 12 clinical or counselling psychologists, 5 psychiatrists and 2 nurse therapists, with a mean of 11.6 years delivering psychological treatment and nearly all with over six months' psychosis-intervention experience. [source: Garety et al. 2024]
2. Training: manual orientation, teaching on rationale, targets and delivery, self-directed learning, and initial pilot cases closely supervised by an experienced AVATAR therapist. [source: Garety et al. 2021; Garety et al. 2024]
3. Supervision: typically weekly individual supervision and monthly group peer supervision with shared audio (AVATAR2); in AVATAR1, Julian Leff trained therapists and provided weekly group supervision. [source: Garety et al. 2024; Craig et al. 2018]
4. Fidelity: session checklists (an 80% threshold was predefined in AVATAR2) and expert competence ratings of recorded early, mid and late sessions. [source: Garety et al. 2021; Garety et al. 2024]
5. Access to training now: the sources reviewed do not name a public training route; contact the AVATAR team (avatar@kcl.ac.uk is listed on the KCL AVATAR2 page) or Avatar Therapy Ltd. [source: KCL AVATAR2 page; training route uncertain: not verified]
Example language (illustrative, for supervision): "In session 3 I conceded too early and she looked disbelieving. Can we listen to that bit?"
Watch for: therapist discomfort voicing abuse; drift toward general support (losing the power-shift focus that distinguishes AVATAR from the AVATAR1 counselling comparator). [source for the comparator's deliberate avoidance of power-shift targets: Craig et al. 2018; discomfort point synthesis]

4.21 Data privacy, security and digital governance

Purpose / when: Before adopting the system and before each person starts.
Steps:
1. Use only the approved software; NICE makes use conditional on appropriate regulatory approval including NHS England's DTAC. [source: NICE HTE17]
2. Store session recordings on secure organisational servers with access limited to the therapist and supervisor, as in AVATAR2. [source: Garety et al. 2021]
3. Explain to the person what is recorded, where it is stored, who hears it, how long it is kept, and that the copy they take home is theirs to protect. [synthesis]
4. For remote delivery use the encrypted system, not general video apps. [source: Garety et al. 2021 for encrypted system; synthesis for the caution]
Example language (illustrative): "Your recordings are stored securely in the NHS system and only I and my supervisor listen to them. Your own copy is yours; keep it where others can't play it by accident."
Watch for: paranoid concerns about recording or surveillance; address openly and offer choices (for example, audio only for supervision). [synthesis]

4.22 Remote delivery

Purpose / when: When attending a clinic is not possible. AVATAR2 adapted delivery by video link during COVID-19. [source: Garety et al. 2021; Garety et al. 2024]
Steps: provide hardware if needed (returned at end); hold a remote set-up session to explain and troubleshoot; agree a private space; keep close liaison with the care team to manage distress. [source: Garety et al. 2021]
Example language (illustrative): "Before we start, can you show me where you're sitting? Is anyone else at home who might hear? If the link drops, I'll phone you straight away."
Watch for: connectivity failures mid-dialogue and the therapist's lack of physical presence, both noted as barriers in the protocol; agree a reconnection and phone back-up plan. [source: Garety et al. 2021; plan synthesis]

4.23 Working with the care team and prescriber

Purpose / when: Throughout. AVATAR was delivered alongside usual specialist care; in AVATAR2 usual care included antipsychotic medication for 94% of participants. [source: Garety et al. 2024]
Steps: share the plan and summary letter; report changes in voices, mood or risk; refer medication questions (including side effects or wishes to change medication) to the prescriber, without advising any change. [source for letter: Craig et al. 2018; rest synthesis]
Example language (illustrative): "That's an important question about your tablets. Let's write it down for Dr X; I can let them know you want to talk about it."
Watch for: the person attributing improvement to stopping medication, or planning to stop; encourage a conversation with the prescriber. [synthesis]

4.24 Conversations about general-purpose AI and do-it-yourself voice tools

Purpose / when: If a person uses AI chatbots for support or is tempted to recreate the avatar with consumer voice or face tools. AVATAR is a therapist-voiced, supervised treatment; the AVATAR2 authors' planned AI conversational agents are research developments, not current practice. [source: Garety et al. 2024; rest synthesis]
Steps: ask about use without judgement; explain that unsupervised confrontation with a recreated voice lacks the safety, calibration and support built into AVATAR; agree to bring such experiences into sessions. [synthesis]
Example language (illustrative): "Some people try apps to talk to their voice. In AVATAR, I control what the avatar says so it can change at your pace. On your own, it could say anything. Can we talk about what you've tried?"
Watch for: increased distress or voice activity after such use; chatbot content echoing voice content. [synthesis]

5. Handling problems

Items tagged [source] are reported by the developers; items tagged [synthesis] are this guide's own suggestions drawn from general psychosis-therapy practice and are not part of the published AVATAR protocol.

Fear of the voice's retaliation; refusing to start. In the 2013 pilot, most dropouts were attributed mainly to fear instilled by the voices, which often threaten people who disobey them. [source: Leff et al. 2013] Response: name the fear as understandable; repeat the safe-space rationale; offer a first dialogue with neutral or mildly critical content only, very short, with the stop word rehearsed; consider a session of building the avatar and hearing it say neutral things before any hostile content. [synthesis] (illustrative) "The voice has warned you not to come. That tells me it's worried you'll get stronger. We can go as slowly as you need."
Overwhelm in Phase 1. All AVATAR1 therapy-related withdrawals occurred in Phase 1, which the authors describe as especially challenging. [source: Ward et al. 2020] Response: shorten dialogues (about 5 minutes is the published starting length), use less distressing verbatim lines first, coach more in your own voice, end on a quick win, debrief longer. [source for short early dialogues: Ward et al. 2020; rest synthesis]
Other voices drowning out the dialogue. Two pilot participants with multiple voices could not concentrate because other voices were loud. [source: Leff et al. 2013] Response: reduce session demands, check whether the focus voice is the right one, consider coping strategies in pre-dialogue, and review suitability with the team. [synthesis]
Unable to look at the face. In the pilot, the face was removed for a man with a history of sexual abuse and the dialogue continued by voice. [source: Leff et al. 2013] Response: offer voice-only or a less realistic or non-human face (AVATAR2 offers non-human characters). [source: Garety et al. 2024]
'It doesn't look or sound right.' The aim is a good-enough match. [source: Garety et al. 2024] Response: adjust a little, then explain that the dialogue works through what is said and how the person answers, not a perfect copy. [synthesis]
Concession feels fake or too early. Reported by one AVATAR1 participant. [source: Ward et al. 2020] Response: return the avatar to a firmer stance, let the person win more ground, and ask the person what the voice would realistically say next. [synthesis]
Stuck: the person stays submissive or apologetic toward the avatar. [synthesis] Response: rehearse shorter, firmer lines in pre-dialogue; use their values or people they protect as motivation; have the avatar make a small concession after any assertive attempt to reward it; explore in debrief what makes standing up feel dangerous.
Voice unchanged or worse between sessions. [synthesis] Response: review use of recordings (listening to hostile parts may be counter-productive), check stressors, sleep, substance use and medication changes with the care team, reassess risk; consider slowing the pace. In AVATAR1, progress was judged from changes in voice severity, malevolence and frequency, discussed each session. [source for progress review: Craig et al. 2018]
Dissociation or flashbacks during trauma-related dialogue. Dissociation is not discussed in Ward et al. (2020) [uncertain: not addressed in sources reviewed]. Response: stop the avatar, switch to own voice, return to the room, use grounding (orientation to present time and place, sensory anchors), then debrief; resume with less intense content or postpone; consider trauma-focused work with the team. [synthesis] The AVATAR2 authors link some Extended-arm adverse events partly to affective changes with trauma work and recommend collaborative review of session number. [source: Garety et al. 2024]
Role confusion or rupture: the person feels the therapist is the abuser. [synthesis] Response: always signal switches; spend debrief time in your own voice affirming the alliance; acknowledge that voicing the content is uncomfortable for you too; if the person cannot separate roles, discuss with supervision whether to continue. (illustrative) "Those were the voice's words, not mine. I said them so you could practise beating them."
The avatar or therapist becomes incorporated into delusional beliefs (for example that the therapist now controls the voice). [synthesis] Response: explore the belief without argument, restate the setup, involve the care team, and consider pausing.
Assertiveness tipping into rage or threats toward others. [synthesis] Response: coach toward calm, firm statements; assess risk to others if anger generalises outside sessions; share with the care team. AVATAR1 recorded violent incidents in both arms with none attributed to therapy. [source: Craig et al. 2018]
Missed sessions and dropout. AVATAR-Extended had 58% completion versus 82% for Brief; the authors call engagement a key area for improvement. AVATAR1 ended therapy after three consecutive missed sessions. [source: Garety et al. 2024; Craig et al. 2018] Response: contact after a first missed session; ask about fear of the voice, practical barriers and session burden; offer remote delivery where appropriate; review whether a shorter version suits. [synthesis]
Not using the recordings. Engagement was variable in AVATAR2. [source: Garety et al. 2024] Response: make listening specific (which clip, when), check device access and privacy, and ask what gets in the way. [synthesis]
The person does not want to lose the voice, or the voice is partly benevolent. [synthesis] Response: clarify goals: AVATAR targets distress, power and control, not necessarily the voice's absence; compassion-toward-the-voice work may fit. [source for compassion target: Ward et al. 2020]
Religious or culturally specific voice identities (for example a devil or a spirit; the pilot included a man hearing a devil's voice for 16 years). [source: Leff et al. 2013] Response: explore the person's framework respectfully; agree what form of avatar and what kind of standing-up fits their beliefs; consider faith or cultural consultation. [synthesis]
Sadness or relapse fears at ending. Some people feel sad saying goodbye to the avatar. [source: Ward et al. 2020] Response: name and mark the ending, use the final dialogue for future goals, keep recordings, and agree with the care team what happens if voices return. [synthesis]
Technical failure during a dialogue (in-clinic or remote). Connectivity problems were noted as a remote-delivery barrier. [source: Garety et al. 2021] Response: agree a back-up (phone, returning to the room) in advance; never leave the person alone with an unfinished hostile exchange: close with a win in your own voice. [synthesis]
Worry about being recorded. [synthesis] Response: explain storage and access (4.21), offer choices, and do not proceed with recording without consent.
Therapist strain from voicing abuse. [synthesis] Response: use supervision with audio review (as in the trials), debrief after difficult sessions, and step out of role deliberately at the end of each session. [source for audio-based supervision: Garety et al. 2024]
Young people. [source: NICE HTE17] Use for people under 18 only within research or other non-core funded access, as NICE states.

Sources

Primary and official sources:

Further reading: Leff J, Williams G, Huckvale M, Arbuthnot M, Leff AP (2014). Avatar therapy for persecutory auditory hallucinations: what is it and how does it work? Psychosis 6(2):166-176 [uncertain: not verified; Leff's account of mechanisms here is taken from his presentation slides and the 2013 paper].

← Previous · IX-093VR therapy for psychosis (gameChange)Next · IX-095 →Mental-health and wellbeing apps