¿Solo quieres estudiar? Entra sin inscribirte
El curso IA generativa y pensamiento dialéctico tiene una edición de lectura abierta con sus 22 lecciones. No necesitas una cuenta para consultarlas. Solo debes registrarte y matricularte en Aula Clara si quieres presentar los cuestionarios, registrar tu progreso y obtener certificado.
6.2 · Privacy, Bias and Responsible Organizational AI Use
Goal: establish practical governance before sending information to a model or allowing it to take action.
A reasonable question can involve unreasonable disclosure
A commercial manager wants to discover why customers stop ordering. It may be useful to analyze retention; it is not necessarily appropriate to paste customer names, phone numbers, histories and confidential notes into a third-party AI service. First determine the minimum information needed, where it will be processed, who can access it and what contractual or technical controls apply.
Seven basic safeguards
- State the task’s legitimate purpose and responsible owner.
- Minimize data; use fictional, aggregate or suitably anonymized records where appropriate.
- Review provider terms, storage, training, retention, access and deletion controls.
- Separate drafting permissions from permission to query or change external systems.
- Require human approval for significant economic, legal, health or personal decisions.
- Record source versions, computations, decisions and known limitations.
- Establish error reporting, monitoring and withdrawal procedures.
Bias is often in the sample, not just in the wording
A model trained or prompted only with feedback from satisfied customers may characterize the entire market incorrectly. Customers who silently leave are missing from the survey. Before debating whether the generated summary sounds fair, examine coverage, response rates, definitions and who is absent.
Use this fictional business workflow. Identify sensitive information, sample bias, high-impact decisions and proportionate controls. Do not request real names or credentials. Separate suggestions, calculations, retrieved evidence and external actions. Require explicit approval where a tool could modify records.
Exercise: draft a one-page AI policy for a five-person team: permitted uses, prohibited uses, minimum data, access permissions, review responsibility, incident reporting and approval requirements. Compare it with the NIST Generative AI Profile. A tool that processes a file locally does not grant the same privacy protections to external chat services you might subsequently use.
Para registrar avances, presentar los cuestionarios y recibir un certificado debes inscribirte en el curso.
Inscribirme para guardar mi progreso