What You Need to Know
Accountability means being responsible for work when the work goes wrong. Disclosure means telling people that AI helped to make the work. A stakeholder is anyone whom the work affects. Domain 2 asks whether an output is correct. This objective asks who owns the output, who deserved to know how it was made, and whom the output describes.
This lesson uses one made-up example throughout: a recruitment agency that uses Claude for its client newsletter, its website chat for job applicants and the weekly job summary. A senior recruiter used to write the weekly job summary.
Accountability
The person whose name is on the work owns it: the email's sender, the report's publisher, the letter's signer. Saying that Claude drafted the work describes the production method. It is like saying that an assistant typed the work. The statement moves no responsibility. The Usage Policy includes this rule in its high-risk requirements. The policy names High-risk Areas, such as legal or medical advice. When Claude's output becomes advice or a decision in one of these areas, a qualified person must "remain accountable for what is delivered or decided". The AI Fluency framework, which Anthropic's Academy teaches, calls the same duty Diligence.
If the newsletter prints a wrong salary figure under the director's name, the director owes the correction. Blaming Claude, or the person who configured the Project, only explains how the error happened. Under pressure, that blame is tempting, so practise the rule until you apply it automatically.
When Disclosure Is Owed
The Usage Policy sets two fixed disclosure rules.
The first rule covers chatbots. Consumer-facing chatbots and AI agents that communicate directly with external users must tell users that they are interacting with AI rather than a human. This disclosure must happen at a minimum at the beginning of each chat session or in the product interface. The agency's website chat can open each conversation by saying that it is AI, or it can show a label in the chat window. A line in a privacy notice elsewhere on the site does neither of these things.
The second rule covers advice and decisions in the policy's High-risk Areas. The policy states: "An individual who receives advice or who is the subject of a decision based on a High-risk AI Recommendation must be clearly told that AI was used to produce it." Employment is one of these areas, and it includes disciplining a worker. Suppose the director asks Claude to assess a recruiter's late candidate lists and to recommend whether to give a warning. If she uses that recommendation, she must tell the recruiter that AI was used. The policy does not require the notice to name Anthropic, Claude or the model. Checked 29 October 2026: the version of the Usage Policy quoted here takes effect on 12 November 2026.
Outside these two rules, disclosure depends on proportion, and both extremes are wrong. Nobody declares AI assistance on a routine internal email, such as the weekly staff schedule. Declaring everything is over-caution. Disclosure is owed when:
- the reader's trust depends on who did the thinking or the judging: assessed work, a professional opinion
- a claim of authorship would be false
- a policy or client agreement requires it
The test is what the reader relies on. If the reader relies on the information, disclosure adds little. If the reader relies on the fact that you did the thinking, disclosure is owed. A personal message such as condolence, or thanks meant as your own feeling, is different. The reader of such a message relies on the feeling being your own. A disclosure note cannot restore that, so you write the message yourself. An organisation's routine thank-you letters are ordinary drafting work. The newsletter's salary figures need no note. The director writes the thank-you card for a retiring recruiter without Claude.
What a Disclosure Says
A vague note such as "AI helped with this" tells the reader little. Anthropic's Academy describes a fuller note, which it calls an AI diligence statement. A strong statement says what AI was used for. It describes the human review, and it makes clear who is accountable for the final product. For the recruiter's warning, the director could write: "Claude summarised the client records and suggested a warning. I checked each late candidate list against the records and removed two that were the client's fault. The decision to give the warning is mine."
Fairness Is a Different Review from Accuracy
An output can pass every accuracy check and still be unfair. Descriptions of groups, communities, applicants or service users can repeat assumptions that the facts never required. Suppose that the newsletter's applicants section presents one district as a problem, one town's workforce as declining and older applicants as resistant to change. Every figure is correct, yet the people that the section describes still see the unfairness at once. Checking the numbers misses the unfairness, because the unfairness is in how the section presents people. This unfairness needs a different reviewer: someone who knows those applicants and reads for how the section describes them. The Usage Policy also prohibits using Claude to critique anyone's body shape or size. The policy treats such critique as promoting unhealthy body image.
Domain 2 covers spotting bias as an evaluation skill. This objective covers the responsibility half: whose job it is to catch unfairness in work about people before publishing. It also covers what the publisher owes when unfairness is published.
Colleagues Are Stakeholders Too
The stakeholder that most candidates never consider is the person whose work changes. When Claude starts to do part of a process, the colleagues who do that work need to know what changed and why. When colleagues discover that a shared task was automated without notice, that discovery damages the trust that disclosure protects. So the director tells the senior recruiter before the agency sends Claude's first job summary. Using Claude to produce a colleague's contribution without telling them is an ethical question of honesty between colleagues, not a question of output quality. Examples are appraisal comments that the colleague was meant to write and a section that the colleague was meant to own.
In the exam
Questions ask who owns an output, who should have known, who is described, or whose work changed. The right answer keeps accountability with the person named on the work. Wrong options spread accountability onto the tool or the person who configured the Project, or offer a general rule to declare everything, which is over-caution.
What to notice in the exam
An option whose main content is checking, cross-referencing or validating belongs to a Domain 2 objective. The right answer for objective 6.4 is always about ownership, disclosure, fairness or colleagues.
Where These Rules Come From
The two fixed disclosure rules and the accountability rule come from the Usage Policy. What a disclosure says comes from the Academy's diligence statement tutorial. Telling colleagues which parts of shared work were AI-assisted follows the Academy's AI Fluency course. The rest follows the exam guide's objective, read through ordinary professional practice. That covers applying the accountability rule to everyday workplace outputs, the reliance test for disclosure, and telling the colleagues whose work changes.
Exam Traps
Transferring accountability to the tool or its configurer
Explaining that Claude produced the figure, or that someone else set up the Project, describes how the error was made, not who owes the correction. Responsibility attaches to the name on the work, so the author corrects the figure and fixes the check.
Using disclosure to avoid blame or as a rule for every output
An AI-assistance note added to soften ownership of an error misuses disclosure, and declaring assistance on every routine output is over-caution that drowns the cases where it genuinely matters. Disclosure is about what the audience is relying on, not a ritual or a deflection.
Signing off work about people once the figures are confirmed as correct
Fairness failures survive accuracy checks by definition. The characterisation review is a separate pass by someone who knows the people described, and an option that stops at verified figures has answered the accuracy question while the ethical one is still open.
Practice Scenario
The operations manager of a software company publishes the annual service report to customers and investors, largely drafted by Claude from the company's own support and uptime records. A customer spots that the report overstates how quickly support requests were resolved last year, and asks the manager who wrote it. What is the professional response?
Build Exercise
Audit One Published Piece
What you'll learn
- How to write the ownership line for a piece of AI-assisted work
- How to decide whether disclosure was owed and on what ground
- How to run a fairness pass separately from an accuracy pass
- How to identify the colleagues a piece of work affected
- Pick one piece of AI-assisted work you have already sent or published, ideally with your name on it and an audience beyond your own team
Why: The audit is only useful for work with real ownership and real readers. Drafts and private notes have nothing at stake to examine.
You should see: One named piece: what it was, who received it, whose name was on it.
- Write its ownership line: whose name is on it, and exactly what you would say if a reader challenged a claim in it tomorrow
Why: The anchor only holds if you can say the words; drafting them now is what prevents the accountability transfer under pressure.
You should see: Two sentences: the owner, and the response to a challenge that accepts ownership before explaining anything.
- Decide whether disclosure was owed: what was the reader relying on, the information itself or the fact that you did the thinking, and did any policy or client agreement require a declaration
Why: Deciding when disclosure is owed needs practice on real cases, because both extremes feel safe and neither is.
You should see: A yes or no with the ground named: reliance, authorship, or a policy requirement.
- Re-read every passage that describes a group of people and mark anything accurate but unfair; name who knows those people well enough to review it, and whether they did
Why: The fairness pass is a separate skill from fact-checking, and it only becomes visible when run separately.
You should see: Marked passages, or a note that none describe people; plus the name of the right reviewer.
- Note whose work the piece changed or replaced, and whether they knew; if they did not, that conversation is the output of the exercise
Why: The colleague stakeholder is the one this objective says candidates always miss, and the audit is incomplete without them.
You should see: Either a note that colleagues knew, or the name of the person who deserves the conversation and a line on what you will say.
Sources
- Usage Policy — Anthropic
- AI Fluency: Framework and Foundations: A Closer Look at Diligence — Anthropic
- Writing an AI diligence statement — Anthropic
- Claude Certified Associate – Foundations Exam Guide — Anthropic. The two disclosure rules and the high-risk responsibility sentence come from the Usage Policy. What a disclosure says comes from the Academy's diligence statement tutorial. The sections on fairness and on colleagues draw on the exam guide's objective and standard professional practice.