What you will be able to do
- Explain why deployers must tell users when they are interacting with an AI system
- Distinguish Claude's embedded text watermarks from Content Credentials (C2PA) on files
- Interpret what a detected mark, or a missing one, does and does not prove
- Apply the delegate-then-own pattern to ethical gray areas
1.Transparency starts with telling people AI is involved
Fairness asks whether a system treats people evenly. Transparency asks whether they know what they are dealing with. The first layer is plain disclosure. When Anthropic restructured its Acceptable Use Policy into the Usage Policy, it added disclosure requirements. Organizations that use its tools must help their own users understand that they are interacting with an AI system. The obligation does not stop with Anthropic. It passes to whoever builds on Claude and puts it in front of people.
Anthropic's builder course makes the same point as a pre-ship question. Under Responsibility the checklist asks: could this output be misread or misused, and have you been transparent about AI's role? It ties this to ownership: you own the outcome, not the output. Saying that AI produced something neither explains nor excuses what it does to the people who receive it.
Checkpoint 1 of 6· Check yourself
A company embeds Claude in a customer-facing chat widget. Under the disclosure requirements Anthropic added to its Usage Policy, what is expected of the company?
The disclosure requirement applies to the organizations using Anthropic's tools, which must help their own end users understand that they are talking to an AI system.
“organizations who use our tools also help their own users understand they are interacting with an AI system”Source: support.claude.com
Checkpoint 2 of 6· Exam question
During a red-team exercise on a Claude-based loan-eligibility assistant, testers find that the assistant's explanations for declining an application vary in specificity depending on how the applicant phrases their question, even when the underlying facts are identical. Which response best addresses the ethical concern this raises?
Correct answer: A — Standardize the explanation prompt so decline rationales cite the same categories of decision factors regardless of wording, then re-test across phrasing variants.
- A. Correct. Inconsistent explanation depth for identical facts is a transparency and fairness gap; standardizing what decision factors must be cited and verifying consistency across phrasing directly closes it.
- B. Incorrect. Removing explanations eliminates transparency for every applicant rather than fixing the inconsistency, and works against the fairness goal of the exercise.
- C. Incorrect. Gating detailed explanations behind whether the applicant knows to ask for one creates a new access-based disparity instead of resolving the phrasing-based one that was found.
- D. Incorrect. A disclaimer discloses the problem without correcting it, leaving applicants with genuinely inconsistent information based on identical facts.
2.Machine-readable marks: text watermarks and Content Credentials
Disclosure covers the live conversation. Content, though, travels on long after the conversation ends. Anthropic's view is that as AI-generated content becomes commonplace, greater transparency and signals about where content comes from can give people useful context about the information they consume. To deliver that, and to meet its commitments under the EU AI Act's Article 50(2) Code of Practice, Claude marks what it generates in two complementary ways.
The text watermark is imperceptible. It is woven into the generated text and does not change its meaning, quality or readability. Because it is part of the text, it travels with copy and paste and may persist through some editing. It applies at the model level, so it covers supported models on the Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, worldwide, and through AWS, Google Cloud and Microsoft Foundry.
Content Credentials cover files. When Claude generates a supported file type such as a PNG or JPEG, it attaches signed provenance metadata that follows the Coalition for Content Provenance and Authenticity (C2PA) open standard. Credentials are added only when Claude creates a file, so they appear only where a platform offers Claude's file-generation features.
Detection access differs as well. Anyone can check a file for a Claude-issued Content Credential with the free Claude Content Checker. Text watermark detection is in private preview, open to organizations that are eligible under EU law (regulators, media, fact-checkers, researchers and others) and to enterprises that must verify watermarking for their own compliance.
| Aspect | Embedded text watermark | Content Credentials (C2PA) |
|---|---|---|
| What is marked | Generated text | Generated files such as PNG or JPEG |
| Form | Imperceptible watermark woven into the text | Signed provenance metadata following the C2PA open standard |
| Where it applies | Model level: every Claude product or surface, including cloud partners | Only where a platform offers Claude's file generation features |
| How to check | Watermark detection, in private preview for eligible organizations | Free Claude Content Checker |
Checkpoint 3 of 6· Match them up
Match each marking concept to its description
Tap a term, then the definition that fits it.
Watermarks mark text and Content Credentials mark files. Each has its own detection route: a free checker for files and a private-preview detector for text.
“This metadata is called a Content Credential and follows the Coalition for Content Provenance and Authenticity (C2PA) open standard”Source: support.claude.com
Sources3
3.What a mark can and cannot prove
Anthropic is explicit that marks are signals, not proof, and that the limits apply in both directions.
A detected mark is not fully conclusive. It shows the content may have been processed by Claude, but it does not establish who wrote it. People use Claude to proofread, translate, summarize and convert files, so marked output can carry ideas, text or data that came from somewhere else. Marked content may also have been changed, excerpted or combined with other material afterwards.
A missing mark is not evidence of human authorship. Claude output may carry no detectable mark for several reasons. It may come from a model released before marking was supported for it. The text may have been heavily edited, paraphrased, translated or mixed into other writing. The passage may be too short to give a reliable signal. A file's metadata may have been stripped by format conversion, re-saving or a screenshot. Or the platform, feature or file type may not have supported that kind of marking.
The ethical consequence is a matter of proportion. Use marks as one input to a judgment about provenance, never as the sole basis for an accusation or a decision about a person.
Checkpoint 4 of 6· Check yourself
A teacher runs a short student paragraph through watermark detection and finds no Claude mark. What can they conclude?
Anthropic lists several reasons Claude output can lack a detectable mark, including very short passages and heavy editing. So a missing mark does not show that the text was human-written.
“The passage is very short, leaving too little text for a reliable signal”Source: support.claude.com
Checkpoint 5 of 6· Exam question
An education technology company embeds Claude to give students feedback on persuasive essays. A parent complains that the assistant seems to nudge students toward the assistant's own opinion on debate topics rather than helping students develop their own reasoning. Which principle from Claude's constitution is most relevant to correcting this behavior?
Correct answer: A — Epistemic autonomy: Claude should respect the student's right to reach their own conclusions through their own reasoning rather than steering toward an answer.
- A. Correct. The constitution's fairness and autonomy framing includes epistemic autonomy, respecting the user's right to reach their own conclusions through their own reasoning process, which directly addresses nudging students toward the assistant's opinion.
- B. Incorrect. Counterfactual impact concerns whether Claude's involvement is decisive in a harm calculus; it is not the principle governing whether Claude should steer users toward its own opinions.
- C. Incorrect. Operator compliance concerns following platform-specific operational guidance, not the specific concern about the assistant imposing its own views on students' independent reasoning.
- D. Incorrect. Maximizing revision throughput does not address the parent's concern about opinion-steering and could worsen it by prioritizing output volume over independent reasoning.
Sources3
4.Ethical judgment where policy runs out
Disclosure rules and marking standards cover the clear cases. Most ethical decisions sit outside them. Anthropic's academy says that using AI responsibly isn't just about following rules. It is about ethical judgment in the gray areas where policy gives no clear answer.
The academy frames that judgment as a loop around every AI interaction. Before the work, you decide what is right to hand off: the right task, the right tool, the right data. Afterward you take responsibility for what comes back. Its exercise asks you to write down which decisions you will *not* delegate, for example final judgment, relational feedback or high-stakes evaluation. It also asks what you will do when unsure: pause, consult a colleague, review guidance or pilot cautiously.
The commitments are not fixed. The academy says ethical AI use is continuous, and commitments should be revisited as context and tools change. That fits the rest of this lesson. Marking capabilities are still being rolled out, and a deployment's users change over time, so a transparency stance written once at launch will go out of date.
Checkpoint 6 of 6· Check yourself
In the academy's Delegation–Diligence loop, what happens after the AI produces its output?
The loop puts a delegation decision first and the human's responsibility for the result second. Choosing the tool well does not transfer ownership of what it produces.
“Afterward you take responsibility for what comes back.”Source: academy.claude.com
Sources4
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.If watermark detection finds no Claude mark, the content was written by a human.Why is that wrong?
Claude output can lack a detectable mark for several reasons, for example when it has been edited, paraphrased, translated, shortened or produced on an unsupported surface.
Covered in What a mark can and cannot prove
2.A detected Claude mark proves that Claude authored the content.Why is that wrong?
A mark only shows the content may have been processed by Claude. Claude may have proofread, translated or summarized material that someone else wrote.
Covered in What a mark can and cannot prove
3.Content Credentials (C2PA) are attached to every Claude output on every platform, just like text watermarks.Why is that wrong?
Text watermarks apply at the model level. Content Credentials are added only when Claude creates a file, so they appear only where a platform offers Claude's file-generation features.
Covered in Machine-readable marks: text watermarks and Content Credentials
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“organizations who use our tools also help their own users understand they are interacting with an AI system”
↩︎ Transparency starts with telling people AI is involved - 2.
“You own the outcome, not the output.”
↩︎ Transparency starts with telling people AI is involved - 3.
“greater transparency and signals about where content comes from can give people useful context about the information they consume”
↩︎ Machine-readable marks: text watermarks and Content Credentials“Generated text will carry embedded watermarks, and generated files will include Content Credentials (C2PA) where supported.”
↩︎ Machine-readable marks: text watermarks and Content Credentials“To check whether a file contains a Claude-issued Content Credential, use the free Claude Content Checker.”
↩︎ Machine-readable marks: text watermarks and Content Credentials“A detected mark provides a signal that content was processed by Claude, but is not fully conclusive.”
↩︎ What a mark can and cannot prove“The text has been heavily edited, paraphrased, translated, or mixed into other writing”
↩︎ Exam trap 1“Claude may not be the original author.”
↩︎ Exam trap 2“Content Credentials (C2PA) are added when Claude creates a file”
↩︎ Exam trap 3“Watermarking will be applied at the model level”
↩︎ Prediction“This metadata is called a Content Credential and follows the Coalition for Content Provenance and Authenticity (C2PA) open standard”
↩︎ Checkpoint“The passage is very short, leaving too little text for a reliable signal”
↩︎ Checkpoint - 4.https://academy.claude.com/courses/ai-fluency-for-k-12-educators/ethics-responsible-useOfficial docs
“Using AI responsibly isn't just about following rules; it's about ethical judgment in the gray areas where policy gives no clear answer.”
↩︎ Ethical judgment where policy runs out“Ethical AI use is continuous. Revisit your commitments as students, context, and tools change.”
↩︎ Ethical judgment where policy runs out“Afterward you take responsibility for what comes back.”
↩︎ Checkpoint