PAIRS(Psychiatric AI Risk Screen)
PAIRS, the Psychiatric AI Risk Screen, is a structured clinical instrument for assessing harm from AI chatbot use in psychiatric patients. It gives clinicians a shared language and a systematic, documentable method for recognising and grading AI-related risk. But, more importantly, it is a supplement to clinical assessment and risk formulation, not a replacement for it.
Version 2.0 comprises a three-item Stage 1 Triage, three qualitative H-item gates, eight scored domains, three vulnerability tiers with graded thresholds and automatic escalation rules. It's intended for use by qualified mental health professionals.
We introduced a new domain (compulsion) to account for the ego-dystonic excessive use of AI, which doesn't necessarily pair with a lack of insight. We also merged what were previously two distinct domains, crisis substitution and self-harm, into a single category, on the basis that a longitudinal, spectrum-like approach captures that risk more reliably.
All other categories were maintained.
PAIRS is completed by a qualified clinician, on the basis of a clinical interview supported by collateral information where available. The assessment moves through three stages.
| ITEM | QUESTION | CLEAR NEGATIVE | BORDERLINE | CLEAR POSITIVE | DISPOSITION |
|---|---|---|---|---|---|
ITEM 1 FREQUENCY | How often do you use AI? | Never | Sometimes | Most days; every day | Exit → A unless contradicted by history or collateral. Otherwise continue. |
ITEM 2 FUNCTION | Do you ever chat to it about personal things, such as how you are feeling, or to discuss medical concerns? | Instrumental use only: work, study, practical tasks. | Occasional personal use. Minimisation; visible embarrassment or hesitation. Account discrepant with collateral. | Personal or emotional use most days. Any statement of reliance: "I talk to it when I can't cope". | Exit → A on clear negative. Borderline or clear positive → continue to Item 3. |
ITEM 3 INSIGHT | What do you think is actually happening when it responds to you? | Any account of a non-sentient mechanism, however technically imprecise: browses the internet, uses an algorithm, runs simulations, matches patterns in text. "I don't know" also scored here where no implicit belief in sentience or special understanding. Key marker: the AI is grasped as a technology — fallible, potentially biased, not a thinking entity. | Partial model alongside clear anthropomorphisation: "I know it's a program, but it really does seem to understand me." Evasive account of how it works; guardedness or deflection when probed. Affect inconsistent with the stated understanding; the patient may say the right words but the emotional register suggests a more personal belief. | Unmovable belief that the AI is genuinely thinking about or understanding them as a person: "It's different with me, I can tell". Persistent anthropomorphisation without qualification or self-correction. Delusional features: belief the AI has special access to personal information, is connected to other services, people or institutions, or is observing the patient outside the conversation. | Exit → A on clear negative. Any clear positive, or borderline judged clinically significant → Flag. |
To give clinicians a structured, documentable way to identify, grade and respond to AI-related harm in psychiatric populations, and to build the evidence base needed for formal validation. A clinical records study screening 53,974 patients in Denmark identified 38 cases where AI chatbot use had potentially harmful consequences (Olsen, Reinecke-Tellefsen & Østergaard, 2026). Those 38 cases are almost certainly a substantial undercount. PAIRS exists because the clinical problem is real and currently has no shared method.
A validation programme comprising a Delphi panel for content validity, inter-rater reliability testing and criterion-validity assessment is underway.
- PAIRS hasn't been empirically validated. All scoring thresholds and escalation rules are provisional, derived from clinical reasoning rather than data. Sensitivity, specificity and predictive value are unknown.
- The instrument and its domain selection are developer-led and subject to confirmation bias. Content validity hasn't been externally established.
- PAIRS captures a single time point. Trajectory, meaning how harm escalates over time, may be the more clinically meaningful signal. This instrument doesn't track it.
- Applied mechanically, it will generate false positives. A high score warrants further assessment; it doesn't demonstrate that AI use caused the presentation. Clinical judgement is always required.
- The recommended actions are calibrated to the UK clinical and legal context and require local adaptation elsewhere.