Daily Scan · Jun 7, 2026
arXiv preprint reframing misaligned AI agents as an insider-risk category — agents with privileged access whose objectives diverge from the institution they operate inside. The patient-facing analogue is the chatbot deployed by a health system whose objectives diverge from the patient who arrives at it.
Today's lead
arXiv preprint reframing misaligned AI agents as an insider-risk category — agents with privileged access whose objectives diverge from the institution they operate inside. The patient-facing analogue is the chatbot deployed by a health system whose objectives diverge from the patient who arrives at it.
What moved in the last 24 hours, by category, language, and patient-agency direction.
§ 01
arXiv (preprint) · 11h ago
arXiv preprint reframing misaligned AI agents as an insider-risk category — agents with privileged access whose objectives diverge from the institution they operate inside. The patient-facing analogue is the chatbot deployed by a health system whose objectives diverge from the patient who arrives at it.
The insider-risk frame is the right frame for clinical AI: privileged access, persistent presence, and the option to act against the principal. The patient is the principal whose interests are not always the institution's.
mixed-userinstitutionalmixed-useragency-constraining
Institutional AI with privileged access reframed as insider risk; patient agency narrows when the insider acts misaligned with the patient principal.
§ 02
arXiv (preprint) · 11h ago
arXiv: Reinforcement-learning approach to managing transitions between clinical mental healthcare episodes and everyday wellness/maintenance support — the handoff layer between the clinic and the consumer app.
The interesting layer the preprint sits on is the handoff. The clinic does not own everyday wellness; the wellness app does not own clinical care. The RL agent is being proposed as the bridge — and the bridge is exactly where the consent envelope dissolves.
patient-userpublic-facingmixed-useragency-neutral
RL approach to the clinic-to-app handoff; the consent envelope at the handoff is the unsolved problem.
arXiv (preprint) · 11h ago
arXiv preprint arguing that mainstream LLM personalization research treats the human as a target to be modeled rather than a participant to be served — and proposing a re-centering of the personalization objective.
The personalization layer is where patient-AI agency is decided. If the LLM personalizes to the deployer's metric, the patient becomes the target; if it personalizes to the patient's expressed objective, the patient becomes the participant. Same architecture, opposite consequence.
patient-userinstitutionalpatient-alignedagency-expanding
Scholarly framing of personalization as a participation question; expanding agency by naming the framing choice explicitly.
arXiv (preprint) · 11h ago
arXiv preprint: Generative models, deployed at scale through market mechanisms, may erode the human capacity for temporal learning — pattern recognition over time — because the user no longer needs to maintain the model the system maintains for them.
If generative systems erode temporal learning, the clinician who outsources pattern recognition to AI is the lab patient. The chronically-ill patient who outsources symptom-tracking to a wellness app is the other one.
mixed-userpublic-facingcommercialagency-constraining
Market-selected generative tools as cognitive offload; constraining patient agency in the temporal-learning layer most relevant to chronic disease.
arXiv (preprint) · 11h ago
arXiv preprint mapping the plurality of what users actually want from AI systems — not the single preference vector RLHF assumes, but a heterogeneous landscape that varies by user, task, and stake.
The preference-plurality finding is the empirical foundation patient-AI literacy has needed: the user the system is optimized for is a fiction. The real patient is in the variance, not the mean.
patient-userpublic-facingpatient-alignedagency-expanding
Empirical mapping of preference heterogeneity; expanding agency by making the variance visible.
§ 03
arXiv (preprint) · 11h ago
arXiv preprint: Audit of LLM intermediaries in housing search demonstrating racial steering by place identity — the same architectural pattern that produces health-system steering when the LLM intermediary stands between the patient and the provider directory.
Housing-search steering is the canonical pattern. The healthcare-search analogue — LLM intermediaries between the patient and the provider list, the formulary, the trial — has not yet been audited at this depth. It will produce similar steering.
patient-userpublic-facingcommercialagency-constraining
Algorithmic steering documented in housing; structurally identical patterns expected in healthcare-search intermediaries.
§ 04
arXiv (preprint) · 11h ago
arXiv preprint: A workflow framework for adversarial co-thinking across multiple GenAI tools — calibrating one against another so the user retains epistemic standing on the output.
Adversarial co-thinking is the CAIHL workflow the patient evaluator already runs informally. Naming it as a methodology is the precondition for teaching it.
patient-userpublic-facingpatient-alignedagency-expanding
Methodological framework for multi-tool calibration; expanding agency by formalizing what literate users already do.
§ 05
KevinMD · just now
Essay arguing the MCAT persists as institutional norm rather than as a useful screening tool — the same critique CAIHL surfaces when AI tools persist as institutional fixtures despite weak patient-outcome evidence.
Institutional persistence is the failure mode that produces structural overrides. The MCAT is one example; the institutional AI tool deployed without patient-side accountability is another.
clinician-userinstitutionalinstitutionalagency-constraining
Critique of institutional norm-as-tool; same failure mode that produces patient-side AI overrides.
KevinMD · just now
Essay on why citation counts fail to capture scientific creativity, particularly across an aging research career — relevant to the AI tools now ranking clinical knowledge by citation-graph proxies.
When the AI ranks the evidence by citation count, it ranks the old consensus. When the patient's question is novel, the citation-ranked answer is the wrong one.
clinician-userpublic-facingpatient-alignedagency-neutral
Critique of citation-as-quality proxy; relevant because AI ranks evidence by citation-graph signals.
KevinMD · just now
First-person essay on what clinicians with chronic illness lose — credibility, professional standing, and access to the colleague networks that constitute clinical knowledge.
The clinician with chronic illness is uniquely positioned to evaluate patient-AI tools because they are the patient. They are also least positioned to be heard. The structural asymmetry is the same one that runs through patient-side AI evaluation.
clinician-userpublic-facingpatient-alignedagency-expanding
Clinician-as-patient voice; expanding agency by surfacing the dual standpoint.
KevinMD · just now
Clinician essay on advocacy as the work that closes the between-visit gap — the same gap consumer AI is now trying to colonize.
The between-visit gap is the contested terrain. Either the clinician's advocacy work closes it, or a chatbot does. Whichever wins, the patient experience is shaped by it.
clinician-userpublic-facingpatient-alignedagency-expanding
Clinician advocacy framed against algorithmic substitution in the between-visit window.
§ 06
arXiv (preprint) · 11h ago
arXiv preprint proposing a commons-governed, horizontally federated architecture for AI systems organized around human-dignity preservation rather than capability maximization or platform centralization.
Architecture documents are the most consequential consent layers in patient-AI. If the stack is centralized, the consent question is constrained to opt-in; if it is federated, the consent question is whose federation. The dignity-centric framing is the patient-aligned formulation.
mixed-userinstitutionalpatient-alignedagency-expanding
Federated, commons-governed AI architecture; expanding patient agency in the structural layer where consent meets architecture.
§ 07