Guidelines for using LLMs

In the context of research, writing, and editing, it’s important to remember what LLMs actually are: models trained on vast data sets to predict and generate text that sounds like it was written by a human. LLMs are powerful tools for increasing efficiency on certain kinds of content tasks. As we all now know from many highly publicized cases of AI gone wrong, LLMs also generate errors and raise concerns about confidentiality. They predict, but they don’t actually understand. That’s a key distinction.

We also know that the path to adoption of AI capabilities may not be a smooth one. It’s possible that teams will continue to adopt these tools at a rapid pace. It’s also possible that regulatory and energy-related issues will interrupt access to the most popular platforms in unpredictable ways. Even as we build our skills in using them, becoming over-reliant poses risks.

While each firm has its own practices and policies for these tools, most follow sensible guidelines. Best practices include:

• All content must be reviewed by at least one human to ensure accuracy.

• Confidential information should never be shared with an AI platform.

• All members of a project team must be transparent about how AI is and is not being used in the work.

• Smart people can disagree about the best role for AI in content creation. These tools are always changing, and effective policies come from discussions that carefully weigh the risks and benefits.

• Focus on value. AI tools should add value by creating efficiency and consistency, generating new ideas, and generally lightening the load on the team. Quickly creating a poor-quality finished product does not actually save anyone time or money.