Content moderation

LinkedIn Engineering describes its content moderation as three layers of protection that filter out policy-violating content before it affects members, whether in the feed or in private messages1. LinkedIn's Help Center says the second layer covers content that is likely to break the rules but where the algorithm is not confident enough to remove it automatically, so AI systems flag it for human review3. The third layer is human-led detection, which starts when members report content3. The Help Center adds that samples of content the automated system removes are sent to human review to measure how often its removals are correct, which lowers the chance that compliant content is removed3.

Classifiers are computer models that judge whether content breaks the rules, and LinkedIn builds them in-house2. LinkedIn's engineering team described developing models to detect viral spam in April 20238. Two Microsoft patents describe moderation methods, though they do not show what LinkedIn does today910. A 2018 patent describes building a language model using machine-translated content and preventing selected comments from being displayed10. A 2020 patent scores a reported message by its similarity to earlier reported messages and then picks a treatment based on that score9.

LinkedIn organizes and checks the work of its human reviewers closely311. LinkedIn's risk assessment under the EU's Digital Services Act (DSA) says its flagship moderation tool shows reviewers each queued item, such as a post, job, ad or message, and they label it and move to the next11. LinkedIn Engineering said in November 2023 that a system scores each queued post or comment by how likely it is to break the rules, so likely violations go to the front of the queue12. LinkedIn says this cut the average time to catch violating content by about 60%12. A Microsoft patent from 2022 describes a single dashboard where moderators see content matching rules, with predictions and recommended actions, and can restrict or block it13. LinkedIn says reviewers face strict time limits and receive training, regular audits of sampled work, group calibration sessions (reviewers meeting to align on common mistakes) and coaching for lower performers2511. Moderators can reach content policy managers through office hours and escalation paths, managers can consult in-house lawyers on complex cases, and LinkedIn is testing ways to improve review quality, including jury review and masking previous reviewer information511.

Members feed LinkedIn's content moderation through reports, which they can file on each part of LinkedIn, which LinkedIn calls a surface5. LinkedIn also uses an automated system, based in part on past human reviewer decisions, to resolve some reports, and the notice a reporter receives says whether the system or a human decided414. LinkedIn's DSA transparency reports say the system acts only above set confidence levels, does not act on some report types such as terrorist content, and is retrained over time414. LinkedIn's DSA auditor described service-level targets of two to 24 hours depending on content type, and a configuration that blocks a reporter who passes the permitted number of reports in 24 hours from submitting more6. Reporters and creators can both appeal, and LinkedIn grants roughly 40% of requested appeals, with those grants making up only 0.05% of content decisions5.

LinkedIn's DSA auditor reports that when violations reach predefined thresholds, LinkedIn's systems restrict the member account and notify the relevant audience6. Management reviews and approves the strike rules every year, and they consider the type and number of offenses6. When LinkedIn decides on content, an automated email gives the decision, a statement of reason and information on redress options7. The auditor's report refers to code configurations that classify and filter violations considered deceptive, high-volume commercial content to prevent notification of the author7.

The rules that LinkedIn's content moderation enforces are its Professional Community Policies, and LinkedIn publishes material on how it applies them15. LinkedIn's blog said in May 2022 that it updated these policies and added detailed examples of disallowed content and of how it handles account restrictions, including the types of comments and behaviors that violate the policies15. LinkedIn's Help Center says misinformation and inauthentic content are not allowed16.