Skip to content
AI Knowledge MapSearch / Text directory

How do I distinguish permission to propose ideas from permission to invent facts?

C015 · C Controlling changes · Output scope and expression

Key answer

Allow creative proposals, stories and hypotheses while separating them from real-world claims. A new campaign idea permits invention; a company success story needs real events, data and sources. Novel ideas do not make accompanying figures and cases true.

Full explanation

Doshi and Hauser randomly assigned 293 participants to write eight-sentence stories, some with GPT-4 ideas, then had independent readers evaluate them. AI-assisted works improved on some creativity/writing measures while becoming more similar to one another. This measured creative performance, not factual accuracy. Authors' manuscript, design and results

Factual writing needs a different check. FActScore splits biographies into small verifiable facts and checks reliable-source support, finding supported and unsupported information mixed in one passage. Sections 2–3

Distinguish three content types:

· Facts: supplied or verified, with supporting sources.

· Suggestions: unimplemented proposals labeled as recommendations or possibilities.

· Hypotheses/fiction: demonstration assumptions labeled near the example.

Do not label every sentence; distinguish cases readers might mistake for actual achievements. For example:

Suggest three campaign ideas. Product functions and existing achievements must come from the attachment; ideas may be creative. Mark missing sales, testimonials and test results “to be supplied.” Label any illustrative case “fictional example” in a separate paragraph.

“Assume 100 users” can demonstrate a calculation; “we surveyed 100 users” claims a survey happened and needs records. Plausibility does not replace a source.

Before delivery, check new names, institutions, amounts, percentages, customer quotations and performance promises. Ensure fictional labels survive copying. Assess ideas for utility/feasibility and verify facts individually to preserve exploration without contaminating records.

References

1. Doshi and Hauser (2024), Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances; authors' manuscript

2. Min et al. (2023), FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation, EMNLP, Sections 2–3