What it is
A caseworker who asks the system whether a family is dangerous has already framed the answer. The research cited below finds that an AI system's tendency to agree grows when the person asking pushes back. It also finds that a system that keeps agreeing can lead even a user who reasons perfectly to a confident false belief. Framing hygiene is the discipline of asking without a position: stating the facts of the case, asking what they support, and leaving the conclusion out.
What it pushes on in the Lab
In the Lab, this lever weakens the pathway by which a person's framing biases the model's answer.
You can also pull this lever at a strong tier, which costs more. At the strong tier, each effect below the Lab's strongest setting pushes harder.
In the modes that offer aiming, you can aim this lever at particular parts and pathways of a network. Otherwise it applies to the whole network.
Its pattern in the Practice Library
The Practice Library describes the pattern behind this lever:Framing and mirroring reduction
The pressures it answers
These pressures list this lever among the levers that answer them:
A lever answers a pressure when it pushes the other way on something the pressure pushes on.
Where you can pull it
Networks in the Lab that offer this lever:57
- A commercial code assistant across three enterprises
- Albert France Services
- Allegheny Family Screening Tool
- Ambient scribe RCT + monitoring playbook
- An ambient AI scribe at a multi-specialty health system
- Benefits Data Trust wind-down
- Burokratt
- Caddy adviser copilot at Citizens Advice
Every network that offers it
- Calgary Drop-In Centre
- CDTFA Axyom Assist
- CHAI (chronic-homelessness prediction)
- Character.AI crisis-safety stack
- Colorado Family Safety and Risk Assessments
- CORA, the DC CFSA policy assistant
- Crisis Text Line & Loris.ai
- CVS Health's Massachusetts applicant video-interview screen
- Dave ExtraCash (CashAI)
- Douglas County Decision Aide
- EDD Virtual Assistant
- Frida (NAV Norway)
- Gaggle Safety Management
- GDS Microsoft 365 Copilot cross-government experiment
- GetCalFresh
- Google ML code completion
- GOV.UK Chat
- Illinois Rapid Safety Feedback
- Imagine LA Benefit Navigator copilot
- Justice Transcribe
- Kaiser Permanente ambient AI scribe
- LA's coordinated-entry triage revision
- Learned Hand AI clerk pilot (LA and Riverside courts)
- London's Strategic Insights Tool
- Magic Notes (Beam)
- Mass.gov Virtual Assistant
- Meta employment-ad targeting and delivery optimization
- Minute / Local Transcribe
- Nava assistive benefits chatbot
- Netherlands childcare-benefits scandal (Toeslagenaffaire)
- nH Predict Utilization Review
- NJ AI Assistant
- NYC MyCity business chatbot
- ODMAP overdose spike alerts
- Propel in-app SNAP benefits assistant
- REACH VET
- Rotterdam welfare-fraud risk model
- San Jose's camera car
- Santa Clara County Homelessness Prevention System
- Singapore's chatbot fleet refresh
- SSA Insight
- The same AI running hands off: the agentic office
- The same AI under full guardrails: the professional office
- The same AI with a human checking: the supervised office
- Trelleborg's Welfare Robot
- UK DWP Universal Credit Advances fraud model
- UK Home Office asylum AI copilots
- United Behavioral Health's Level of Care Guidelines
- VA claims automation (automated survivor-benefit decisions)
The evidence behind its effects
The Lab cites these claims from the evidence registry for this lever's effects.
In formal simulation, even ideal Bayesian users spiral to near-certain false beliefs under a sycophantic interlocutor at sycophancy rates measured in frontier models (~50-70%), and truth-constrained cherry-picking still produces spirals — minimizing hallucination alone is insufficient.[2]
Chandra, K., Kleiman-Weiner, M., Ragan-Kelley, J., & Tenenbaum, J. B. (2026). Sycophantic chatbots cause delusional spiraling, even in ideal Bayesians [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2602.19141
doi.org/10.48550/arXiv.2602.19141
Appears in: Paramerge authored research
Topics: ai-safety
Sharma, M., Tong, M., Korbak, T., et al. (2024). Towards Understanding Sycophancy in Language Models. In International Conference on Learning Representations (ICLR 2024). https://doi.org/10.48550/arXiv.2310.13548
doi.org/10.48550/arXiv.2310.13548
Appears in: Evidence reverification (2026)
Topics: ai-safety, human-ai-interaction
Research on AI sycophancy describes it as a fragmented construct — a family of distinct agreement-seeking behaviors that share a label but differ in form, mechanism, measurement, and required mitigation — and finds it intensifies under user pushback and across multi-turn interaction.[2]
Ye, M., Ibrahim, L., Bo, J. Y., et al. (2026). What Counts as AI Sycophancy? A Taxonomy and Expert Survey of a Fragmented Construct [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2605.21778
doi.org/10.48550/arXiv.2605.21778
Appears in: PAN framework development
Topics: ai-safety
Sharma, M., Tong, M., Korbak, T., et al. (2024). Towards Understanding Sycophancy in Language Models. In International Conference on Learning Representations (ICLR 2024). https://doi.org/10.48550/arXiv.2310.13548
doi.org/10.48550/arXiv.2310.13548
Appears in: Evidence reverification (2026)
Topics: ai-safety, human-ai-interaction