The IP risks of LLMs at work are nontrivial (www.augmentedswe.com)

AI-written code can create real intellectual‑property headaches: LLMs trained on public repositories sometimes emit snippets that are verbatim or very close to existing open‑source code, potentially importing restrictive license obligations (e.g., GPL) or failing to preserve required attribution. Researchers found roughly 1% of code-model outputs are “strikingly similar” to public code, and there have been documented cases of GPL‑licensed code reproduced with no notice—exposing companies to accidental copyleft, attribution breaches, or even patent exposure. That’s why teams are layering technical and process controls: model‑level safeguards (filters and tuned prompting), provenance or training‑data tracing where available, automated license‑and‑similarity scanners integrated into CI, and legal/code‑review workflows that catch risky snippets before they land in production. Startups often rely on rapid heuristics and blocking tools, while mature firms add formal policies, audits, and IP reviews. The upshot for engineers and managers is clear—AI can boost productivity, but you need detection tooling, review gates, and legal processes to avoid slipping license or patent liabilities into your codebase.