“Use AI responsibly” sounds sensible until a marketer needs to make a decision. Can they paste a customer's complaint into a writing tool? Can a generated avatar deliver a testimonial? Can they publish a statistic that appears in a draft without a source?
A useful policy answers those questions in plain language and gives someone responsibility for difficult cases. Start with the decisions your team faces regularly, then check the laws and platform rules that apply to your actual activity.
Keep claims tied to evidence
AI-generated text can invent details that make a message persuasive: customer names, performance numbers, quotations, and product capabilities. Require a verifiable source for material claims and a person who checks it.
Label fictional examples as hypothetical. Do not present them as customers, studies, or outcomes. Avoid fake first-person experience: an assistant's sentence beginning “we tested this” needs a real test behind it.
The FTC's guidance on consumer reviews and testimonials explains rules addressing deceptive reviews, including fake experiences. AI assistance does not make an invented endorsement truthful. The precise legal assessment depends on the activity and jurisdiction; your policy should route uncertain cases to the responsible specialist.
Decide what information may enter a tool
Create practical data categories rather than asking every employee to make a privacy judgment from scratch. Approved public product facts are different from account records, private messages, sensitive information, or confidential client material.
For each approved use, document the vendor, access controls, retention, and the purpose of processing. Keep unnecessary details out of prompts. Where GDPR applies, the European Commission's data protection principles include purpose limitation and data minimisation. A public comment can still contain personal data.
Do not describe a vendor as universally compliant or private. Assess the product, terms, settings, and data flow you actually use. Our AI personalization guide shows how relevance can often be achieved with less information.
Make disclosure fit the reader's decision
Ask what a reader could reasonably misunderstand. A person may need to know that a support response is automated, that an image depicts a fictional scene, or that a presenter is synthetic rather than a customer.
Check the relevant platform and legal requirements. Disclosure is not one universal sentence that satisfies every situation. Place necessary information where people can notice it before relying on the content.
Never imply that a fictional testimonial, generated audience, or automated action represents a real customer's experience. Truthfulness comes before the choice of label.
Review fairness where it affects people
Test whether the wording makes assumptions about age, gender, language, location, or ability. Check whether personalization changes who sees an offer or receives assistance in a way you cannot justify.
You do not need to claim that a few prompt tests prove the system is unbiased. Record the cases you checked, the problems you found, and the scope you have not evaluated. If the use affects important opportunities or sensitive decisions, it requires more than a marketing copy review.
Define permission to act separately
Drafting, publishing, contacting people, and changing records have different consequences. Decide who approves each action and whether the software is permitted on the platform.
LinkedIn's automated activity policy includes restrictions on certain third-party software and extensions. Manual approval does not by itself establish compliance. For broader workflow boundaries, see the ethics of social media copilots.
Write five rules and a correction path
- No invented proof: Publish claims, quotes, and experiences only with support.
- Approved data only: Use the minimum permitted information in each tool.
- Clear representation: Do not let readers mistake fiction or automation for a real person or experience.
- Named review: Assign someone to check facts, fairness, and action permissions.
- Visible correction: Report errors, correct affected content, and repair the process that caused them.
For a hypothetical correction, suppose a published post says “our customers save two hours a day,” but the reviewer cannot find supporting research. Pause reuse of the claim, identify the copies you control, and have the responsible owner decide the correction. A plain explanation might be: “We removed the time-saving figure because we could not substantiate it. The product description has been corrected.” Do not quietly replace it with a different unsupported number. Internally, record which source and approval were missing so the next draft meets the rule.
Give employees a contact for exceptions and examples of safe and unsafe use. Review the policy when the tool, data, or task changes. A short policy people can apply is more useful than a long statement of values that leaves every decision unresolved.