Interviewing
Pasted Patient Records Are a Finding About the Last Employer
Patient records pasted into a chatbot at a previous job say something about that employer's training before they say anything about the candidate who volunteers it. Treat the admission as an opening rather than a verdict. Ask what the rule was there, who they would ask about it now, and what they would do on the day the approved tool cannot do the job. Those three answers are evidence.
The takeMaking this disqualifying is self-defeating, and the reason is arithmetic rather than principle. People who volunteer it are only some of the people who did it. Make the admission fatal and the behavior stays in your pipeline while the reporting of it leaves, so you end up hiring the same habit from someone who learned to keep quiet about it. A round that punishes candor selects for polish, and polish is the one quality that tells you nothing about how somebody handles a record they should not have moved.
Where Olive fits
Open a role and see what the work shows
Olive produces no composite, and no decision about a candidate is made by a machine alone: a person writes or edits each of the six findings, and every finding carries the timestamped excerpt it rests on, with the rubric, scorer and bank versions exported alongside the released report. For a policy conversation about how a candidate works with AI, that is a record of what they did during one session rather than an account of what they say they would do.
Assess your shortlistIs this your breach to report?
No. The federal breach notification rule puts the duty on the covered entity that holds the information. It must notify each affected individual after discovering a breach of unsecured protected health information, without unreasonable delay and no later than 60 calendar days 1. A company interviewing the person who did it is not that entity.
Confirm your own position with counsel, since this reading covers the US federal rule alone.
What you owe is a hiring decision, and the sentence is one input to it.
The reflex in the room runs the other way. An admission like that arrives with a verdict already attached. The next ten minutes get spent deciding whether the candidate is trustworthy, on evidence that will never be better than the sentence you just heard.
But the sentence is thin evidence about a person and thick evidence about a place. Ask where the approved tool was at that employer, and the answer may be that nobody can name one.
Read the four questions the breach rule asks first
The breach notification rule, published on January 25, 2013, treats an impermissible disclosure as something to be assessed rather than pronounced on. A disclosure not permitted under the privacy rule is presumed to be a breach unless the covered entity shows a low probability that the information was compromised, based on a risk assessment of at least four factors 2.
The four, in the rule's own order:
- The nature and extent of the information involved, including the types of identifiers and the likelihood of re-identification.
- The unauthorized person who used the information, or to whom the disclosure was made.
- Whether the information was actually acquired or viewed.
- The extent to which the risk has been mitigated.
None of the four asks whether the person who made the disclosure can be trusted. One is about the material, one about who received it, and two about what happened to it afterward. A candidate can speak to each of them.
The second factor also does real work in an interview. A hospital may disclose protected health information to a business associate when it obtains satisfactory assurance that the associate will appropriately safeguard it, documented through a written contract or other written agreement 4.
So a deployment covered by a written agreement with the vendor and a personal account are not the same event, even when the screen looks identical.
Borrow the rule's posture and the questions nearly transfer to the interview. What was in the paste, whether it carried identifiers, where it went, who else could see it, and what happened next.
The rule is more curious than the reflex is.
Ask three questions that separate a gap from a habit
Ask what the rule was where they worked, who they would ask about it now, and what they would do when the approved tool cannot do the job. The first question describes the environment they came from, and the second describes escalation. The third comes closest to the job, because it puts them in a situation your own company will hand them by March.
A strong answer is specific about the boundary and vague about nothing. The candidate names the class of material that never goes into a general tool, names the sanctioned one, and names the person they would ask when the two collide.
A weak answer stays up at the level of the principle.
"I would never put confidential data in" is a sentence anybody can say, and it is the sentence a coached candidate says.
The third question is where the two come apart. The hard version of this failure goes like this: the sanctioned tool cannot do the thing, the deadline is today, and somebody has to choose.
A person who has been there answers with a specific route that keeps the material inside the boundary. A person who has not stays general, with no named person to ask and no plan for the work still due.
Self-report only reaches so far. When the role is worth the extra half hour, the alternative is testing the data boundary with a case rather than a policy question. Hand over a request and some material with something in it that should never be sent, then read what comes back.
Should a disclosure like this cost someone the job?
Rarely, and the reason is who ends up telling you about it. In Microsoft and LinkedIn's 2024 Work Trend Index, 78% of the people who use AI at work said they bring their own tools to it 3. The index is a survey of 31,000 knowledge workers across 31 markets, fielded between February 15 and March 28, 2024.
An AI tool of their own is the common case, and the candidate in front of you may be describing one out loud. The figure says nothing about what goes into it.
Saying it out loud is harder. In the same survey, 52% of people who use AI at work said they are reluctant to admit using it on their most important tasks 3.
Disqualify on the admission and you have selected against the people willing to say the awkward thing.
A firm's account of its AI use and its workers' accounts are different measurements, too. A Federal Reserve note comparing the three main US adoption surveys found they disagree by construction.
About 18% of firms reported adopting AI at the end of 2025, against about 41% of the workforce reporting generative AI use at work in November 2025 6. Different units, different questions.
"Nobody here does that" is a statement about the firm.
For a privacy officer, a compliance lead, or a lone coordinator handling protected health information with no second reader, a candidate who would still move it through an unapproved tool today has given the whole answer. That reading is narrow, and it is real. The previous employer's training failure changes nothing about what would happen on your side.
Whatever you land on, the wording of the note matters.
A note recording that a candidate described pasting patient records into a consumer chatbot at a previous job is a fact. A note calling them a compliance risk is a conclusion somebody will ask you to defend, which is the case for writing what you heard, never why you think it happened.
Check whether your own onboarding names the approved path
Check what your company would have told this person in week one, because the rule that governs the incident puts the duty to train on the covered entity, not the individual. A covered entity must train all members of its workforce on its privacy policies and procedures, as necessary and appropriate for them to carry out their functions, and must have and apply appropriate sanctions against workforce members who fail to comply with them 5.
Picture guidance that fails that test. It sits on a wiki page written by legal in 2023, names no tool anybody can actually open, and is never repeated by the manager handing out the work.
Name the sanctioned tool, the class of material that never goes anywhere else, and the person to ask.
Three sentences from a manager in a new hire's first week beat the policy.
The same boundary applies to your own process, since everything handed to a candidate during an exercise meets it. Assume anything you send a candidate goes into a model, and send the material that can survive the trip.
The sentence stays in the room either way. What changes is what sits beside it in the file: the rule the candidate could name, the person they would ask, and what they said about the day the sanctioned tool could not do the job.
Common questions
Do you have to report what a candidate says about a previous employer?
For patient records, the US federal breach notification duty in 45 CFR 164.404 runs to the covered entity that holds them, which must notify affected individuals without unreasonable delay and no later than 60 calendar days after discovering a breach of unsecured protected health information. An employer interviewing the candidate is not that entity and is not the one discovering a breach in its own records. Ask counsel about your jurisdiction before acting on any of this. The practical rule for the room is narrower: keep the detail inside the hiring file, and do not repeat it to a reference, a vendor or a mutual contact.
What does a strong answer to "what was the rule where you worked" sound like?
Specific, and sometimes unflattering to the old employer. A candidate with sound judgment about regulated data names the sanctioned AI tool, names the class of material that never goes into any other tool, and names the person they would ask when a deadline and a boundary collide. A candidate without it describes a training module they clicked through and a policy nobody enforced. Both answers are usable. The first says the habit is already built; the second says the person has never been given the question, which is a gap onboarding can close in a week.
Does it matter which chatbot it was?
Yes, and the federal breach notification rule says so. The identity of the person or service that received the information is one of the four factors in the risk assessment a covered entity runs after an impermissible disclosure. A licensed deployment covered by a written agreement with the vendor and a personal consumer account are different events, even when the interface on screen is identical. Ask which one it was. A candidate who does not know the difference has told you where the real gap is, and that gap is teachable.
What if the candidate seems completely untroubled by it?
Treat the calm as a reason to ask rather than as a read on character, then test it. Someone who describes the paste with no discomfort may have worked somewhere the practice was normal and sanctioned, or may genuinely not see a boundary. The two can look identical. Ask what they would do now, in a job where the rule is written down and the deadline has not moved. A person who has thought about it answers with a way to get the work done inside the rule. A person who has not treats the question as rhetorical.
How do you keep this consistent across candidates?
Consistency comes from asking everyone in the role the same three questions (what the rule was where they worked, who they would ask now, and what they would do when the approved tool cannot do the job), not only the person who volunteered something. A question asked once, of the candidate whose disclosure unsettled you, produces a record showing one person was interviewed differently from the rest, and that record is the part that becomes expensive later. Put the questions in the interview guide, give every interviewer the same wording, and record the answers in the candidate's own words.
Should this go on the scorecard as its own line?
Only if the role handles regulated material, and then as a behavior rather than a trait. Score what the candidate can name: the boundary, the sanctioned path, the escalation, and the workaround for the day the two conflict. Do not score whether the disclosure happened, because that fact belongs to a different employer's controls and says almost nothing about the person on their own. A line that reads "named the escalation path" can be marked the same way by two interviewers. A line reading "trustworthy with data" cannot.
References
- 1. 45 CFR 164.404 - Notification to individuals ✓ govinfo.gov Supports the claim that the notification duty sits with the covered entity holding the information: it 'shall, following the discovery of a breach of unsecured protected health information, notify each individual' and must do so 'without unreasonable delay and in no case later than 60 calendar days after discovery of a breach.'
- 2. 45 CFR 164.402 - Definitions (breach notification) ✓ govinfo.gov Supports the presumption and the four factors quoted in the second section: an impermissible disclosure 'is presumed to be a breach unless the covered entity or business associate, as applicable, demonstrates that there is a low probability that the protected health information has been compromised based on a risk assessment of at least the following factors,' being the nature and extent of the information and identifiers, the unauthorized person involved, whether the information was actually acquired or viewed, and the extent of mitigation. Section dated [78 FR 5695, Jan. 25, 2013].
- 3. AI at Work Is Here. Now Comes the Hard Part (2024 Work Trend Index Annual Report) ✓ microsoft.com Supports both figures in the fourth section: '78% of AI users are bringing their own AI tools to work (BYOAI)' and '52% of people who use AI at work are reluctant to admit to using it for their most important tasks.' Survey of 31,000 knowledge workers across 31 markets, fielded February 15 to March 28, 2024 by Edelman Data and Intelligence.
- 4. 45 CFR 164.502 - Uses and disclosures of protected health information: general rules ✓ govinfo.gov Supports the claim that a licensed deployment and a personal account are different disclosures: a covered entity may disclose protected health information to a business associate 'if the covered entity obtains satisfactory assurance that the business associate will appropriately safeguard the information,' and those assurances 'must be documented through a written contract or other written agreement or arrangement.'
- 5. 45 CFR 164.530 - Administrative requirements ✓ govinfo.gov Supports the claim in the final section that training and sanctions are employer duties: 'A covered entity must train all members of its workforce on the policies and procedures with respect to protected health information required by this subpart and subpart D of this part, as necessary and appropriate for the members of the workforce to carry out their functions within the covered entity' (164.530(b)(1)), and 'A covered entity must have and apply appropriate sanctions against members of its workforce who fail to comply with the privacy policies and procedures of the covered entity.'
- 6. Monitoring AI Adoption in the US Economy (FEDS Notes) ✓ federalreserve.gov Supports the point that firm-level and worker-level adoption figures are different units rather than conflicting readings: about 18% of firms had adopted AI at the end of 2025 (Census BTOS) against about 41% of the workforce reporting generative AI use at work in November 2025 (Real-Time Population Survey), with the note attributing the spread to sampling distributions and units of analysis.
6 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.