DRAFT proposal for a code of conduct for AI literacy, policy and other practitioners whose main job it is translate AI information from the expert domain to the general public (this may include AI journalists, popularisers and even AI ethicists):

Giving examples of AI output

• Always list the model, tool, and prompt when sharing outputs (date if using historical examples)
• Do not report on any examples of AI performance based on a single use, always report out of how many tries the result obtains
• Before sharing examples of LLM limitations or capabilities confirm whether there is recent research that confirms the issue still exists or the capabilities have been shown to replicate
• If there is variability in how the model responds to different prompts; always try multiple models if you are giving an example as evidence of general behavior

Sharing AI research

• Never reference media reports about research, always read the original paper (this includes science tabloids like SciAm, Quanta Magazine, or University press releases)
• When sharing research results always include when empirical work was done (months matter in AI)
• Never simply repeat points from abstract or discussion section, always check that the methods warrant the conclusions
• Correct documents, slides, post when new evidence appears

Making general statements about AI

• Always be clear what kind of AI, you're talking about when it is not LLMs (for example when you cite papers published before 2022)
• Always mention how representative you think an issue is when sharing anecdotes
• Avoid journalistic headline phrases like "research shows AI could" if you cannot substantiate the quality of the research and its representativeness
• Admit uncertainty when appropriate (for instance, accompany any statements about what AI will never do or will definitely do with a disclaimer about the poor track record of such predictions)

Expertise and professional integrity

• Be clear about the limits of your own expertise; if you do not follow developments in detail, always check the latest state of play
• Do not use metaphors of AI, if you cannot articulate the exact nature of the target domain