Lesson objective
Distinguish the refusal failure from the unread-acceptance failure, and state honestly what the automation-bias and gaze research does and does not support about either.
Plain-language explanation
Why this matters now
Every dental AI session is aimed at two failures that pull in opposite directions, and an instructor who forgets one produces the other.
The first failure is refusal: a hygienist who will not use a drafting tool for a recall note, so the note is late or never written; an assistant who dismisses a marked-up radiograph before the dentist sees it; a practice that keeps a manual process because the tool feels like a threat. Refusal costs quietly, in work not done and time not recovered.
The second failure is unread acceptance: a drafted message sent without being read against the record; a narrative submitted because it looked right; an overlay taken as a finding by someone whose job is not to make findings. Unread acceptance costs loudly, because it reaches a patient.
An instructor who spends thirty minutes on the dangers produces refusal. An instructor who spends thirty minutes on the gains produces unread acceptance. The session has to hold both, and the honest statement about the second is uncomfortable.
Research (not binding): Parasuraman and Manzey, "Complacency and Bias in Human Use of Automation: An Attentional Integration", Human Factors, vol. 52, issue 3, pp. 381-410, June 2010, a review and integrated theoretical model. The authors report that automation complacency "occurs under conditions of multiple-task load, when manual tasks compete with the automated task for the operator's attention", that it "is found in both naive and expert participants and cannot be overcome with simple practice", that "Automation bias results in making both omission and commission errors when decision aids are imperfect", and that automation bias "occurs in both naive and expert participants, cannot be prevented by training or instructions, and can affect decision making in individuals as well as in teams."
Research (not binding): Goddard, Roudsari and Wyatt, "Automation bias: a systematic review of frequency, effect mediators, and mitigators", Journal of the American Medical Informatics Association, vol. 19, issue 1, pp. 121-127, 2012. It defines automation bias as "the tendency to over-rely on automation". Of 13,821 retrieved papers, 74 met the inclusion criteria. Mediators included user factors such as cognitive style and task-specific experience, trust and confidence, and environmental factors: "workload, task complexity, and time constraint". Mitigators included "training and emphasizing user accountability" and design factors such as position of advice on screen, updated confidence levels, and "the provision of information versus recommendation".
Both concern automated decision aids in fields such as aviation, process control and clinical decision support. Both predate large language models. Neither is a study of teaching. Neither measured a session, a curriculum, a handout or an instructor. They reach a dental practice only by analogy, and a lesson must say so out loud.
A question to sit with
Your closing slide says "Always check the output." Nobody in the room disagrees. Three weeks later, at 4:40 on a Friday with two patients waiting and a claim due, a coordinator sends a drafted narrative unread. Besides your sentence, what is in place at that desk?
What goes wrong
1. The instruction carries the whole load. "Always verify" is true, necessary and insufficient. Parasuraman and Manzey report that automation bias could not be prevented by training or instructions.
2. "Our team are experienced, so this does not apply." Both reviews report the effect in expert participants as well as naive ones. A dentist of twenty-five years reading an AI-generated history summary is doing a new task, and task-specific experience is a mediator, not immunity.
3. The research read backwards. Someone concludes the studies prove training is money wasted. They do not. Goddard and colleagues list training and emphasising user accountability among the mitigators. The honest statement is narrower: training or instructions alone did not prevent automation bias.
4. The gaze study turned into a teaching result. Research (not binding): Arsiwala-Scheppach and colleagues, "Impact of artificial intelligence on dentists' gaze during caries detection: A randomized controlled trial", Journal of Dentistry, 2024;140:104793. Twenty-two dentists assessed a median of 18 bitewings each, yielding 170 datasets without AI and 179 with AI after excluding poor gaze recordings. Dentists using AI had more fixations on teeth with restorations, median 68 with an interquartile range of 31 to 116, than dentists without AI, median 47 with an interquartile range of 19 to 100, p = 0.01. The visual search strategy of dentists using AI was less systematic, with a lower proportion of lateral tooth-by-tooth transitions. That is an eye-tracking mechanism study. It measures where dentists looked. It does not measure patient outcomes, and most between-group differences other than fixation count were not statistically significant. It is not a teaching-method result and must never be presented as one. What it is good for in a session is one narrow, honest point: assistance can change how a clinician searches, so the habit of a systematic look does not maintain itself.
5. Only one kind of error rehearsed. Learners practise catching a wrong statement. They never practise noticing what the draft left out. Automation bias produces both omission and commission errors, and in a dental practice the omission is often the one that surfaces later.
6. Refusal treated as an obstacle to overcome. The instructor sets out to convert the sceptic. But doubt pointed at a confident output is the disposition the work needs. The aim is to aim it, not to remove it.
The practical answer
What an instructor may say, and may not say.
| You may say | You may not say |
|---|---|
| Automation bias is "the tendency to over-rely on automation" (Goddard and colleagues, 2012). | "Dental staff over-rely on AI 40% of the time." No such figure exists here. |
| It appeared in naive and expert participants, and training or instructions alone did not prevent it. | "Training does not work." |
| Workload, task complexity and time constraint were named environmental mediators. | "A busy hygiene day doubles the error rate." |
| Training and emphasising user accountability were named mitigators. | "Accountability eliminates automation bias." |
| Dentists using AI fixated more on restored teeth and searched less systematically (Arsiwala-Scheppach and colleagues, 2024). | "The study shows AI makes dentists worse", or anything about patient outcomes. |
| This research is pre-LLM, from other fields, and applies to dental work by analogy. | "This research was done on dental AI training sessions." |
Best practice: teach both kinds of error with dental material. A commission error is acting on something the draft got wrong: a fictional recall message states that a fictional patient is due for a three-month periodontal maintenance visit when the record says six, and the scheduler sends it. An omission error is missing what the draft never raised: a fictional history summary lists two medications and silently omits a third that appears on the intake form, and the reader, seeing a tidy paragraph, does not go back to the form. A session that shows only the first teaches half the problem.
Best practice: make the conditions of failure explicit and design around them. Multiple-task load, time constraint and task complexity are the named mediators, and they describe a dental practice exactly: an assistant turning a room while the next patient is seated, a coordinator working a claim deadline, a hygienist finishing a chart between columns. Where those conditions cannot be removed, the check belongs at the moment with the least load, which is usually earlier than anyone has placed it.
Best practice: make accountability visible during the session. Goddard and colleagues name emphasising user accountability among the mitigators. In a dental practice that has a concrete shape: a named person reviews the output and the review is recorded. Show the practice's actual review record in the room, and have every learner complete one on a fictional item before they leave.
Best practice: rehearse under something like real conditions. An untimed exercise in a quiet break room rehearses a situation the learner will never be in. Run one timed round on fictional material with one realistic interruption, then debrief which items were accepted too fast. Describe the result as what happened in that room, never as a measurement of anything.
AI Coalition Network recommendation: design the session as one part of a system. The session carries part of the weight. The practice's written review rule, the record of who reviewed what, the tool's settings and the minutes allowed for the check carry the rest. When a design is finished, write the list of what else must be true at the chair and at the front desk, and hand each item to the person who owns it. This is our educational standard, not a rule from any board, association or regulator.
Remember this
- Two failures, opposite directions: refusal costs quietly, unread acceptance reaches a patient.
- Automation bias appeared in naive and expert participants and was not prevented by training or instructions alone; training with accountability is among the named mitigators.
- The gaze trial measured where dentists looked, not outcomes and not teaching; most other between-group differences were not statistically significant.
- Teach both omission and commission errors, using dental drafts, charts and messages to do it.
- Name what else must be true at the desk, and hand each item to its owner.
Dental practice example
Sunil Abernathy, a fictional practice manager and trainer at the fictional Brightfield Pediatric Dentistry, inherits a session whose last slide reads "AI makes mistakes. Always check." He rebuilds it. He adds one slide quoting Parasuraman and Manzey directly, with the caveat printed beneath it that the review concerns automated decision aids in other fields before large language models, reaches dental work by analogy, and did not study teaching. He then builds a timed round. Each learner receives a fictional AI-drafted recall batch for ten fictional patients at a fictional practice, generated from a fictional recall report. The batch contains 24 factual statements: names, last visit dates, the interval each patient is on, and the procedure each is due for. Three statements are wrong: a patient placed on a three-month interval whose record says six, a last-visit date a year adrift, and a procedure named for the wrong sibling in a family of three. The batch also omits one thing entirely, which is the planted omission: one patient's record carries a note that all contact must go to a guardian at a second address, and no drafted message reflects it. Three wrong statements plus one omission makes four planted problems in twenty-four statements. Learners get eleven minutes and are interrupted once by a fictional phone message handed across the table, which is how Sunil rehearses multiple-task load. In the debrief, most learners found the interval error, fewer found the date, two found the sibling mix-up, and only one mentioned the guardian note. Sunil says plainly that this describes the nine people in that break room on that Wednesday and measures nothing. He then shows the practice's real review record, which names the reviewer and the items checked, and every learner completes one for the fictional batch. Afterwards he writes to the owner dentist, Dr Fenella Achebe, who owns the tool settings: the drafting tool shows its message without the recall record beside it, and he recommends she look at whether it can show both. The decision stays with her.
Responsible-use warning
Never let a closing instruction carry a session, and never tell a dental team that the research proves training is pointless. Parasuraman and Manzey report that training or instructions alone did not prevent automation bias, and Goddard and colleagues list training with emphasis on user accountability among the mitigators. Never present the gaze trial as a teaching-method result or as a finding about patient outcomes, never report what happened in your room as a measurement, and never let a session aimed at over-reliance produce blanket refusal instead.
Human-review requirement
The instructor owns the accuracy of every statement about the research. Before delivery, a second reviewer checks each quotation against the published paper, confirms each is labelled as pre-LLM research from other fields applying by analogy, confirms the gaze trial is labelled as an eye-tracking mechanism study rather than a teaching or outcome result, and confirms no rate or percentage has been invented. The practice owner of the review rule confirms that the review record used in the session is the one the practice actually uses, and the licensed dentist owns any clinical content that appears in the practice material.
Practical exercise
Design a timed round that rehearses both kinds of error. Use a fictional practice, fictional patients, fictional records and a fictional tool throughout.
- Choose one administrative output a dental team produces with AI help: a recall batch, a history summary, a benefits note or a plan presentation draft. Say which role owns it.
- Write the fictional source record the draft should have been built from, with at least twelve stated values: names, dates, intervals, procedures, contact preferences or medications. Mark every value as fictional.
- Write the fictional AI-produced draft containing twenty to thirty factual statements. Count them and record the total.
- Plant at least two commission errors: statements that contradict the source record. Record exactly where each is and what the record actually says.
- Plant at least one omission: something in the source record the draft never reflects. Record it, and write one sentence on how it could reach a patient.
- Confirm your planted problems and your statement count add up, and write the addition down.
- Set a time limit that reflects a working day rather than a quiet classroom, and add one realistic interruption drawn from practice life. Write one sentence explaining the choice as instructional practice, not as a research finding.
- Write the review habit you will teach: which items are always checked against the record, and how a sample of the rest is chosen under the practice's own written rule.
- Draft the review record the learners will complete, naming the reviewer, the items checked against the source and the date, and say which outputs on your list must instead be owned by the licensed dentist.
- Write one sentence each about Parasuraman and Manzey, about Goddard and colleagues, and about the gaze trial, each carrying its caveat and each stating that none of them studied teaching.
- Write your debrief questions, including one about the planted omission, and make sure none asks a learner to describe a real patient.
Knowledge check
-
What did Parasuraman and Manzey report about automation bias and training?
Answer guidance: That automation bias occurs in naive and expert participants, cannot be prevented by training or instructions, and can affect decision making by individuals and by teams; complacency also arises under multiple-task load.
-
What did the 2024 dental gaze trial actually measure, and what may it not be used for?
Answer guidance: It measured where dentists looked while assessing bitewings, finding more fixations on restored teeth and a less systematic search with AI; it measured no patient outcome and no teaching method, so it supports neither.
-
What is the difference between an omission error and a commission error in a drafted recall message?
Answer guidance: A commission error is acting on a statement the draft got wrong, such as the wrong recall interval; an omission error is missing something the draft never raised, such as a record note about who may be contacted.
-
Why is refusing a useful tool treated as a failure alongside accepting output unread?
Answer guidance: Because refusal costs quietly in work not done and time not recovered, while unread acceptance costs loudly by reaching a patient; a session weighted to either extreme reliably produces the other.
Sources and further reading
- Raja Parasuraman and Dietrich H. Manzey; Human Factors (Human Factors and Ergonomics Society) — Complacency and Bias in Human Use of Automation: An Attentional Integration (2010-06-01). Accessed 2026-09-15.
- Kate Goddard, Abdul Roudsari and Jeremy C. Wyatt; Journal of the American Medical Informatics Association — Automation bias: a systematic review of frequency, effect mediators, and mitigators (2012-01-01). Accessed 2026-09-15.
- Journal of Dentistry (Elsevier) — Impact of artificial intelligence on dentists’ gaze during caries detection: A randomized controlled trial (2024-01-01). Accessed 2026-09-15.
- Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon and Karim R. Lakhani; Organization Science — Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality (2026-03-01). Accessed 2026-09-15.
- American Dental Association, House of Delegates — Policy Statement on the Use of Augmented Intelligence in Dentistry (Trans.2021:282), as printed in ADA Current Policies, Adopted 1954-2025. Accessed 2026-09-15.
- Complacency and Bias in Human Use of Automation: An Attentional Integration, Human Factors (Human Factors and Ergonomics Society), 2010
- Impact of artificial intelligence on dentists' gaze during caries detection: A randomized controlled trial, Journal of Dentistry (Elsevier), 2024