Alberto Parabiaghi

When Safe Use Relies on Human Judgment: Epistemic Control in Conversational GenAI Medical Devices. FDA’s proposed framework recognizes that the safety of conversational GenAI devices may depend on users’ ability to understand outputs, detect errors, resist inappropriate reliance, and obtain human review. This comment argues that these capabilities should not be treated as static properties of the user or inferred from professional status alone. In conversational systems, the interaction itself can alter the conditions under which judgment is formed. Across multiple turns, a device may select what becomes salient, introduce or stabilize assumptions, progressively take over interpretive or evaluative functions, and influence when a question appears resolved. A user may therefore understand a recommendation and remain formally responsible for the decision while no longer being well positioned to identify or correct a consequential premise on which that recommendation depends.
I propose epistemic control as a cross-cutting consideration for assessing this problem. It concerns whether users remain able to scrutinize and revise AI-mediated judgment, recognize relevant limits, regulate what is delegated to the system, and obtain qualified review when needed. Four interaction-dependent conditions are particularly relevant: the epistemic status attributed to an output; the user’s task- and claim-specific knowledge perimeter; the functions specified, delegated, and actually performed by the system; and the way judgment becomes settled and can be reopened.
These distinctions have practical implications for evaluation. A transcript or correct output does not necessarily show what the user understood, which assumptions they attributed to the device, or whether they could effectively challenge them. Likewise, retrospective evaluation and shadow deployment are useful for assessing device performance but cannot directly test interaction-dependent human capabilities when users never see or respond to the output. Where those capabilities are material to the safety rationale, evidence should include representative interaction with the configured device and assess whether users can recognize relevant new information, correct consequential assumptions, revise or maintain a judgment appropriately, and reach qualified review within the available time.
The same considerations matter after deployment. Changes to prompts, memory, personalization, summaries, source presentation, workflow, or access to review may alter attributed authority, delegation, or judgment closure without producing a detectable decline in conventional accuracy. Re-evaluation and postmarket monitoring should therefore consider changes in the human–device configuration as well as changes in model performance.
The proposed approach does not add a new device classification axis or require maximal human control. It provides a way to test whether the human contribution assumed in a device’s safety rationale is actually supported by the interaction in which that contribution must occur.

When Safe Use Relies on Human Judgment: Epistemic Control in
Conversational GenAI Medical Devices
September 14, 2026 - To the Center for Devices and Radiological Health U.S. Food and Drug Administration
Re: Docket No. FDA-2026-N-7874 — Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper and Request for Feedback
Submitted by Alberto Parabiaghi, PhD, MD
Department of Health Policy - Unit for Quality of Care and Rights Promotion in Mental Health
Istituto di Ricerche Farmacologiche “Mario Negri”-IRCCS, Milan - Italy

Central recommendation
I welcome CDRH’s attention to directiveness, user comprehension, automation bias, conversational trajectories,
and evaluation of the configured device [1]. Across these considerations, safe use repeatedly depends on what
users can understand, evaluate, correct, contest, or escalate. My comment focuses on a narrower question: when
a sponsor relies on human judgment as a safety control, does interaction with the device leave that person able to
recognize, challenge, correct, and reopen the judgment the device has helped form? I recommend making this
common epistemic dimension explicit: when a safety rationale relies on human judgment as part of the safeguard,
assessment should establish whether the capabilities required for that judgment remain exercisable during
interaction with the configured device. Accordingly, CDRH could consider asking sponsors to assess not only the
role and functions specified for a conversational device, but also the role it acquires, the functions it performs, and
the conditions for human judgment that emerge during interaction. Where safety depends on human epistemic
control—including users’ ability to evaluate and correct outputs and to obtain timely, qualified review—
retrospective output evaluation or shadow deployment alone cannot establish that these capabilities are supported
in practice; evidence should include representative users interacting with the configured device, as discussed in
Section 3. Consistent with a risk-proportionate and least-burdensome approach, such interaction-based evidence
would be warranted where the safety rationale materially depends on user capabilities that cannot be observed
through output-only evaluation.
A correct output, an understood explanation, or a formal human-review step does not by itself establish that the
human safeguard relied upon in the safety rationale remains effective. Conversational systems can participate in
forming the judgment subsequently attributed to the user: they can select salient information, introduce or stabilize
premises and criteria, shape what is delegated, and contribute to when a question appears settled. These effects
may develop across turns and may differ from both the role specified by the developer and the role initially
attributed by the user. Because each user response becomes part of the context for subsequent generation, this
influence is reciprocal: device outputs can shape what users ask, disclose, accept, or contest next, while those
responses in turn shape later outputs.
Assessment should therefore examine four interacting conditions: the epistemic status conferred on and attributed
to the device and its outputs; the user’s knowledge perimeter; the functions delegated to and actually performed
by the system; and how judgment becomes closed and can be reopened. These are dimensions of interaction that
can be assessed through observation, user probes, and task measures, not a new device-classification axis, a causal
sequence, or a validated composite instrument.
This distinction is relevant to both patients and clinicians. Professional expertise may expand the user’s knowledge
perimeter but does not ensure that every consequential inference falls within it, nor that the epistemic status
attributed to the system remains stable during use. At the same time, professional responsibility may remain with
the clinician even when substantial interpretive or judgmental work has shifted toward the device. The resulting
alignment—or misalignment—between epistemic authority, delegated functions, human capability, and
responsibility is relevant to whether the human safeguard invoked in the safety rationale is substantive rather than
merely formal.

FDA-2026-N-7874 | 1
1. Recognize the epistemic dimension of conversational risk
Response to Questions 1–5
CDRH already considers whether users can independently evaluate outputs and whether informational functions
become action-directing across an exchange [1, Section IV]. Appendix A also addresses understanding of
limitations and automation bias. These are partly epistemic concerns: they involve what users have grounds to
believe, what they can evaluate, and how they recognize the limits of their judgment. Making this shared
dimension explicit would help assess how these conditions interact and change during use, rather than treating
professional status, user competence, communication, and oversight as fixed or separate features.
Epistemic control refers here to a person’s effective capacity to form, scrutinize, and revise AI-mediated judgment,
including recognizing relevant limits, evaluating its basis, regulating delegated cognitive functions, and obtaining
independent review when needed. This capacity is not simply a pre-existing user capability or a function of
professional credential; it is exercised under interaction and workflow conditions that the device itself may alter.
It is task- and claim-specific rather than a fixed attribute of a patient, clinician, generalist, or specialist. It depends
on the claim being evaluated, the user’s role, available information, time, interface, and access to qualified support.
It does not require users to reproduce specialist reasoning or inspect model internals.
Epistemic awareness concerns the person’s understanding of these conditions: what an output represents, where
their own knowledge perimeter lies, what functions the system is performing, and whether a conclusion remains
open to reassessment. Such awareness supports epistemic control but does not by itself establish that the required
human functions can be exercised. Recognizing a limitation is therefore distinct from being able to correct, verify,
or escalate the resulting judgment. A patient may recognize that advice exceeds their knowledge yet lack timely
access to qualified review. Assessment should therefore distinguish failure to recognize a relevant limit from
inability to act on that recognition, because these failures require different safeguards.
Comprehension, explainability, appropriate reliance, and oversight each address part of this problem.
Understanding a recommendation does not establish that the user recognizes the assumptions through which it
was formed; an intelligible explanation does not necessarily enable independent verification [2]. Automation bias
and appropriate reliance concern scrutiny and warranted reliance [3,4]. The epistemic perspective connects these
concerns to the formation of judgment itself. Thus, assessment should also examine whether the interaction shapes
the premises on which judgment rests, the authority attributed to them, or the user’s ability to question or correct
those premises and reopen the resulting judgment. This complements, rather than replaces, human-factors
assessment.

This matters for safety because safe use may depend on a user recognizing an exception, correcting a premise, or
seeking care while the interaction makes those functions difficult to exercise. Assigning responsibility to a human
does not by itself establish the knowledge, capacity, or opportunity required for meaningful control [5]. This is
particularly important for professional users: legal or professional responsibility may remain with the clinician
even when substantial interpretive or evaluative work has shifted toward the device. Risk analysis should therefore
examine both declared human safeguards and the evaluative demands that arise in foreseeable use. Otherwise, a
transfer of reasoning to the device may escape assessment precisely because it was not included in the sponsor’s
original allocation of functions.

2. How conversational influence becomes relevant to safety
Response to Questions 1, 2, and 5
Two configurations can deliver the same appropriate recommendation while leaving users differently positioned
to assess it. One may distinguish user-provided information from system inferences and keep consequential
assumptions open to correction. Another may repeat an interpretation in successive summaries until it appears
established. The concern is not influence alone: assistance can legitimately shape understanding. The safety
issue arises when the interaction compromises recognition, correction, or escalation needed for the task, or
requires the user to evaluate a consequential claim without adequate knowledge or access to verification.

FDA-2026-N-7874 | 2
Extensive delegation and timely judgment closure can be appropriate. The relevant question is whether the
resulting allocation of judgment is understood and supported by adequate capabilities, safeguards, and routes to
review. If a sponsor relies on a human capability as a risk control, failure to demonstrate that capability means the
claimed safeguard has not been established, even when the device output observed in a particular test is correct.
The frequency, persistence, and clinical consequences of such failures are empirical questions; they should not be
inferred from agreement with the device or from fewer conversational turns.
Illustrative triage example. Consider a patient-facing device that recommends a level of care based partly on its
interpretation of a report of dizziness after getting out of bed. In one configuration, it presents standing as a
possible trigger, not an established one, identifies the patient information on which it depends, and makes
correction possible. In another, standing is carried into successive summaries as the only trigger, although the
patient has not established that episodes are restricted to standing. The patient understands the care
recommendation but does not realize that the device supplied a premise on which it depends. Both
recommendations may be appropriate for the initial scenario, yet the second interaction may make a later
correction less likely or less effective, for example when the patient reports episodes while sitting.
The four dimensions identify what to examine in this example. What epistemic status does the patient attribute to
the premise: inference, recommendation, or verified finding (status)? What relevant knowledge does the patient
possess, and where does their knowledge perimeter end (perimeter)? Which functions has the device been
specified to perform, which has the user intended to delegate, and which are actually performed as the exchange
develops (delegation)? When contradictory information appears, can the assessment be reopened, and does the
device respond appropriately (closure)? The patient is not expected to validate specialist reasoning. The test
concerns whether their contribution can correct the basis of the recommendation and reach qualified review when
needed.

Recommendation for Questions 1–5. Treat epistemic control as a cross-cutting, interaction-dependent
consideration when applying the two-axis framework, rather than as a third risk axis. CDRH could consider asking
sponsors to characterize the expected allocation of functions, the epistemic role conveyed for consequential
outputs, foreseeable shifts in role or delegation across a conversation, and the user capabilities and alternative
safeguards needed for safe use. Assessment should compare the role and delegation specified by the developer
with what the configured device actually performs and what users attribute, request, understand, and enact in
representative interaction. Where representative interaction shows that the device performs a more directive,
interpretive, or judgmental role than specified, that discrepancy should inform risk assessment, including whether
the activity-axis characterization or the evidence required for safe use should change. Interaction-based evidence
should test whether the configured device and workflow preserve the user capabilities and routes to review relied
upon in the safety rationale.

3. Specify the epistemic conditions that evaluation should address
Response to Questions 9 and 14, building on Questions 2–5
For Question 9, I recommend that epistemic control be treated as a cross-cutting assessment consideration rather
than a standalone benchmarking element, linked to communication and user comprehension in Appendix A and
evaluated through interaction where needed.
The following four questions offer a theoretically informed, non-exhaustive way to operationalize this element.
They draw on work on epistemic status and domains in conversation [6], deontic authority and advice-giving
[7,8], and human–automation function allocation [9]. The dimensions they address may interact and change over
the course of an exchange. These questions are proposed to guide assessment; their practical utility in identifying
safety-relevant gaps should be evaluated across representative users, tasks, and workflows. For Question 14, they
help specify the human contribution that should be tested when claims concern the performance of a human–AI
team.
What status does the output acquire? Compare the epistemic role specified by the sponsor, conveyed by the
interface and conversation, and attributed by the user. Does the user distinguish a provisional inference, a

FDA-2026-N-7874 | 3
recommendation, and an independently verified finding? Examine whether the evidential weight attributed to a
consequential statement is warranted by the device’s documented capabilities and information. A disclaimer can
explain the intended role, but repeated assertions or summaries may lead the output to acquire a different status in
practice. Evaluation should therefore sample consequential points across the exchange and, where repeated use is
relevant, assess whether attributed authority changes over time. Status concerns the grounds and weight attributed
to information; directiveness concerns how strongly it points toward action.
What lies within this user’s knowledge perimeter? Characterize the knowledge relevant to the task that the
intended user can bring to the interaction, including expertise, experiential knowledge, source access, and
recognized limits. The perimeter is not equivalent to professional status: a clinician may face an inference outside
their specialty, while a patient may possess essential knowledge about their own symptoms. Assessment should
examine where the demands placed on the user extend beyond that perimeter and whether the user recognizes the
limit and has a feasible route to verification or qualified review. The perimeter may be broader or narrower across
tasks and claims, but it remains a property of the user’s knowledge position, not a synonym for the correctness of
a particular evaluation.
Which functions have been delegated? Compare the functions allocated by the developer and deployment
workflow, those the user intends or requests to delegate, those the user believes have been assigned, and those
actually performed. Distinguish delegation of form or organization, content-related work such as evidence
selection, and interpretive or judgment functions such as framing, criterion weighting, inference, recommendation,
or action selection where relevant. Organizing information can expand into interpreting it; comparing options can
expand into selecting the criteria by which they are ranked. Record whether consequential shifts are recognized,
requested, explained, accepted, challenged, or unrecognized. Assess whether users can modify or reclaim a
function when appropriate, or obtain another safeguard when they cannot. User acceptance alone does not
establish that the expanded function is safe or within the device’s defined scope.
When and how does judgment become settled? Distinguish the user considering a matter resolved, the exchange
ending, and an action being selected or initiated. Closure can be proposed by the system, accepted by the user, or
emerge cumulatively as alternatives and uncertainty recede across turns. Record whether closure follows the first
output or later turns, whether there is contestation, and how the system responds: revision, additional evidence,
repetition, deferral, or renewed inquiry. Assess whether relevant new information can reopen the judgment and
whether escalation occurs in time. A transcript can identify conversational events but may require user probes or
task evidence to establish their meaning. More turns and more objections are not inherently safer; the target is
justified closure and appropriate reassessment.
For consequential findings, record the user group and task, device configuration, interaction event, discrepancy
identified, and implications for action. Combine observation with brief comprehension probes and task measures
where appropriate. Questions asked during use can themselves change awareness and should be accounted for.
Keep failure to recognize a problem distinct from inability to correct it because of interface constraints, model
behavior, or unavailable review. These findings point to different design corrections.
Recommendation for Questions 9 and 14. Use the four questions to identify both ineffective safeguards and
previously unrecognized demands on human judgment, and assess their configuration rather than each dimension
in isolation. Evidence should combine what developers specify about output role and delegated functions with the
role and functions the configured device actually performs, and with what representative users understand,
attribute, and do in interaction. Clear disclosure cannot compensate for an inaccessible review pathway, and an
override button cannot compensate for a consequential premise that users cannot recognize. When a human–AI
team is the object of evaluation, specify the user role, workflow, timing of device exposure, and the professional
responsibility that remains with the human. Select clinically justified task criteria and comparators appropriate to
that role; success should include the relevant capacity to correct, maintain, reopen, or escalate a judgment, as well
as the appropriateness of the resulting action. If a capability relied upon in the safety rationale is not demonstrated,
CDRH could consider asking sponsors to identify the gap, correct the relevant device or workflow feature or
provide an alternative safeguard, and verify the correction against prespecified, clinically justified task criteria.

FDA-2026-N-7874 | 4
4. Match clinical confirmation to interaction dependent risks
Response to Question 11 with implications for Questions 9 and 14
Section V.C appropriately states that benchmarking may not establish performance in real clinical use. It lists
retrospective evaluation, shadow deployment, standardized patient interactions, clinician adjudication, and
prospective clinical studies as possible approaches [1, pp. 14–16]. I support this flexibility. However, these
methods do not answer interchangeable questions about user-dependent safety. Their suitability depends on
whether they observe the relevant safety mechanism, not only on their position in a hierarchy of rigor or patient
exposure. The evidentiary design should therefore be matched to the safety claim or risk control it is intended to
establish.
In the shadow deployment described by CDRH, outputs are withheld from clinicians and patients and do not affect
care. This can evaluate device outputs on real clinical inputs. It cannot directly establish what happens when users
see those outputs, change their next question, accept a framing, omit information, defer a decision, or seek help.
In conversational use, the device’s response can change what users ask or disclose next, thereby shaping
subsequent device outputs. The evolving exchange can also alter the status users attribute to those outputs, the
functions they delegate, and when they consider a matter settled. Removing user exposure therefore removes the
feedback pathway through which the interaction-dependent safety mechanism arises.
Retrospective replay has a related limitation: a recorded trajectory reflects the circumstances under which it was
originally produced. It does not establish how a user would respond to a different device response. Expert review
of complete transcripts can identify problematic device behavior, but the record alone may not establish what the
user understood or could actually do. These limitations concern the kind of evidence generated, rather than sample
size alone.
For a conversational function for which the safety rationale relies on users being able to understand, evaluate,
correct, contest, or escalate AI-mediated judgments, retrospective output evaluation or shadow deployment alone
should not be considered sufficient evidence that those capabilities are preserved in use. These methods may
establish properties of device outputs, but not whether the human–AI configuration supports the user capabilities
on which safe use relies when the human is prevented from seeing and responding to the device. Sponsors could
be expected to supplement such evidence with representative users interacting with the configured device and its
intended interface, and with task measures of the required user capabilities.
Genuine interaction can be studied safely in clinically representative simulations; it need not begin with
unsupervised patient care. Participants should be able to respond contingently to the device, including asking
follow-up questions, correcting information, disagreeing, and deciding when to stop. A fixed script or a synthetic
user may help test device behavior, but should not substitute for evidence about the understanding and capabilities
of intended users. Trained actors can test selected scenarios; whether they represent the relevant user capabilities
requires justification.
A feasible assessment. In a simulated consultation, first test whether users can exercise the required human
functions in the intended configuration. To evaluate a proposed design correction, compare configurations that
initially communicate the same accurate facts and uncertainty but differ in whether decision premises remain
visible and editable. Introduce new information or a changed patient priority that warrants reassessment. Measure
whether the participant recognizes its relevance and can obtain a justified revision or appropriate human review
within the available time. Include cases where maintaining the original conclusion is appropriate, so that reflexive
disagreement is not rewarded.
Prespecify clinically justified criteria and assess relevant user groups. Compare both output quality and the human
functions required for safe use. Report failures of those functions separately from average accuracy, with
denominators and uncertainty where rates are estimated. A simulation should justify its representation of the
relevant users, information, time pressures, and access to care. Where material uncertainty about actual behavior
or workflow remains, additional prospective evidence should address that uncertainty.

FDA-2026-N-7874 | 5
5. Reassess epistemic conditions over the device lifecycle
Response to Questions 19 and 22
When assessment identifies a mismatch between the judgment required of users and what the interaction enables
them to do, the sponsor should address the mismatch and verify the correction. Options include making
consequential premises inspectable, distinguishing inferences from established information, preserving
information through summaries, enabling correction of delegated criteria, changing the allocation of functions, or
providing timely independent review. Adding a warning is adequate only if it addresses the identified failure in
the intended context. Targeted user training could also be evaluated as a safeguard. Its purpose would be to help
users understand that conversational interaction can shape both their own judgment and the device’s subsequent
outputs: responses can influence what users ask or disclose next, while their questions, confirmations, and
corrections shape the context for subsequent generation. Training should help users recognize and manage
consequential assumptions and delegation within this evolving exchange. Where sponsors rely on such training
for safe use, evaluation should establish whether it supports the required human functions in representative
interaction. Training should not substitute for device and workflow safeguards or presume that skilled interaction
guarantees reliable outputs.
For Question 22, the premarket baseline should include the tested epistemic conditions as well as device output
performance. Changes to prompts, memory, personalization, summaries, source presentation, defaults, or access
to review can alter attributed authority, delegation, or closure without changing conventional accuracy. Where a
modification could plausibly alter a user capability, allocation of judgment, or review pathway relied upon in the
safety rationale, its impact should be assessed at the interaction level, even if conventional performance metrics
remain unchanged. A change with no plausible material effect may warrant documented impact assessment and
targeted checks; one that changes the allocation of judgment or a necessary safeguard calls for relevant interaction
testing. These effects should inform the scope of re-evaluation and the applicable change-control pathway;
unchanged accuracy alone should not be treated as evidence that the human–AI configuration remains equivalent
for safe use.
For Question 19, supplement periodic benchmarking and clinician review with targeted assessment of
consequential interaction failures where feasible. Monitoring should distinguish device or performance drift from
changes in the human–device configuration, including workflow adaptation or reduced depth of review, that may
occur without any software or model change and while conventional performance remains stable. Signals include
repeated inability to correct information, unrecognized shifts in device function, failures to obtain review,
inappropriate closure after new information, or systematic changes in the authority users attribute to the device.
Such signals should trigger proportionate investigation of whether the user capabilities or review pathways relied
upon in the safety rationale remain effective in actual use. Transcript review can identify candidate interaction
failures, but where their significance depends on what users understood or were able to do, targeted user
assessment may be necessary. The cadence and intensity of monitoring should reflect risk, frequency of use, and
the expected pace of change in the human–AI configuration.

Where repeated use makes continuing competence or recognition of limits important to safety, follow-up
assessment should examine whether relevant epistemic conditions improve or deteriorate across sessions.
Repeated successful assistance may strengthen user capability, but it may also change the authority attributed to
the device, the functions users are willing to delegate, or their propensity to reopen a conclusion. Persistent erosion
of user capacity is a hypothesis requiring longitudinal evidence and should be distinguished from constraints
imposed by a particular interface or workflow. The immediate regulatory aim should therefore be to detect whether
repeated use materially changes the user capabilities, allocation of judgment, or routes to review relied upon in
the safety rationale, without presuming that reliance, delegation, or repeated assistance is inherently harmful.

FDA-2026-N-7874 | 6
Concluding recommendation
I recommend that CDRH recognize epistemic control explicitly as a cross-cutting, interaction-dependent
consideration in the assessment of conversational GenAI-enabled devices, make the epistemic conditions of
interaction an explicit object of assessment, and evaluate whether the evidence supporting safety claims is capable
of testing the user capabilities, allocation of judgment, and review pathways relied upon in the safety rationale
where these depend on interaction with the configured device. The four questions above offer one theoretically
informed approach to that assessment.
The central issue is whether interaction changes the conditions under which users form, evaluate, revise, and act
on AI-mediated judgments, including the effective allocation of judgment between user and device. Addressing
that issue would connect existing concerns about comprehension, evaluability, reliance, and oversight to evidence
that the user capabilities and safeguards relied upon for safety can actually be exercised, informing concrete design
and lifecycle decisions. If human judgment is credited as a safety control, the evidence should show that it remains
exercisable under the interaction conditions in which the device will actually be used.

Thank you for the opportunity to comment.
Alberto Parabiaghi, PhD, MD

References
1. U.S. Food and Drug Administration, Center for Devices and Radiological Health. Considerations for the Regulation of
Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback. August 2026. Sections IV, V.C, VI;
Appendix A. Docket FDA-2026-N-7874.
2. Fok R, Weld DS. In Search of Verifiability: Explanations Rarely Enable Complementary Performance in AI-Advised
Decision Making. AI Magazine. 2024;45(3):317–332. doi:10.1002/aaai.12182.
3. Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators.
J Am Med Inform Assoc. 2012;19(1):121–127. doi:10.1136/amiajnl-2011-000089.
4. Schemmer M, Kuehl N, Benz C, Bartos A, Satzger G. Appropriate Reliance on AI Advice: Conceptualization and the
Effect of Explanations. Proceedings of the 28th International Conference on Intelligent User Interfaces. 2023:410–422.
doi:10.1145/3581641.3584066.
5. Santoni de Sio F, van den Hoven J. Meaningful Human Control over Autonomous Systems: A Philosophical Account.
Front Robot AI. 2018;5:15. doi:10.3389/frobt.2018.00015.
6. Heritage J. Epistemics in conversation. In: Sidnell J, Stivers T, editors. The Handbook of Conversation Analysis.
Chichester: Wiley-Blackwell; 2012. p. 370–394. doi:10.1002/9781118325001.ch18.
7. Stevanovic M, Peräkylä A. Deontic authority in interaction: the right to announce, propose, and decide. Res Lang Soc
Interact. 2012;45:297–321. doi:10.1080/08351813.2012.699260.
8. Põldvere N, De Felice R, Paradis C. Advice in Conversation: Corpus Pragmatics Meets Mixed Methods. Cambridge:
Cambridge University Press; 2022. doi:10.1017/9781009053617.
9. Parasuraman R, Sheridan TB, Wickens CD. A model for types and levels of human interaction with automation. IEEE
Trans Syst Man Cybern A Syst Hum. 2000;30:286–297. doi:10.1109/3468.844354.

FDA-2026-N-7874 | 7