AI can help shape a narrative for a report, a grant, or an outreach post. It drafts, tightens, and reworks the words. The facts, the consent, and the dignity of the people in the story stay with you. A tool can polish a story; it should not invent one.
Where the line is
Polish the words, protect the story
AI can help
Tighten a grant narrative, reshape a story you gathered with consent for a new audience, or draft an outreach post from real, de-identified facts.
Keep it human
Do not let AI invent quotes, details, or statistics, use a person’s story without their consent, or paste any identifying detail into a tool.
The wording is where AI helps. The truth, the consent, and the dignity are yours to hold.
Good for the craft: turning your notes and real, consented outcomes into a clear draft, tightening a grant narrative, and adapting one story for different audiences.
It will invent details. Language models fill a gap with fluent, unsupported content, so an AI draft can add a quote, a statistic, or a scene that never happened.
Consent is a conversation you keep having, not a signature. The person should know how their story will be used, be able to review and approve it, and be able to withdraw it.
Center dignity and agency. Tell strengths and complexity, and avoid a pity narrative that trades a person’s hardship for attention.
Do not use an AI image tool to depict a typical client or refugee, since these tools reproduce stereotypes even when the prompt never mentions identity.
Never paste identifying details, and disclose when a story or image is AI-assisted.
How it works, and where it breaks
1A model fills a gap with something that never happened
When a model writes up a story from thin notes, it completes the gaps by predicting fluent text, and that text often asserts specifics the source never contained. In a careful annotation of model-written summaries, more than seventy percent contained hallucinated content, and most of it was extrinsic, meaning it could not be verified from the source at all, and the large majority of those additions were simply wrong (Maynez and colleagues, 2020). Researchers separate a hallucination that contradicts the source from one the source cannot verify; a fabricated quote or an invented scene is the second kind (Ji and colleagues, 2023). This is not an abstract risk. Lawyers have filed briefs containing court cases a chatbot invented, complete with fabricated quotations, and drawn sanctions for it (Mata v. Avianca, 2023).
2The invented detail tends to follow a stereotype
What a model invents to fill a gap is not neutral. Left to complete a sentence that began with a group name, an earlier model produced violent content for Muslims about two thirds of the time and mapped the word Muslim to terrorist in roughly a quarter of tests, far more than for other groups, and priming it with positive words only partly reduced this (Abid and colleagues, 2021). Image tools show the same pattern, reproducing demographic stereotypes even when the prompt never mentions identity, in a way users cannot reliably prompt their way out of (Bianchi and colleagues, 2023). For a story about an immigrant or refugee client, that means an AI-added detail can quietly default to a demeaning stereotype, which is why you do not let it invent and do not use it to depict a typical client.
3Consent, accuracy, and dignity are obligations a tool cannot hold
The parts of storytelling that protect the person are professional judgments, not drafting tasks. The social work code holds accuracy in how clients are represented, and the disclosure of only the least information necessary, as obligations of the worker (National Association of Social Workers, 2021). Sector guidance on ethical storytelling treats consent as an ongoing, revocable conversation and gives the storyteller the right to review and approve how their account is used, the applied form of the principle that nothing about a person is decided without them (Voice of Witness, n.d.; Commons Social Change Library, n.d.). A tool can arrange your words; it cannot obtain consent, verify a fact, or weigh a person's dignity, and those stay with you.
Your notes and a funder's frameworkA grant vignette to fact-check and approve
Draft a grant vignette from consented, de-identified facts
Here are true, de-identified facts about an outcome and the framework this funder uses. Draft a short vignette that fits the framework. Use only the facts I give you, do not add any quote, number, or detail I did not provide, and mark anything you were unsure about. [paste de-identified facts, no names or identifying details]
Guardrail. Check every sentence against your records, keep identifiers out, and have the person review and approve the vignette before it goes out. See Module 4.
An approved testimonial and your outcome dataA report narrative with the quote intact
Build a report narrative around an approved quote
Draft the narrative around this approved testimonial and these outcome figures. Keep the quote word for word, do not rewrite or paraphrase it, and do not add any achievement or number I did not list. [paste the approved quote and your real figures]
Guardrail. Never let the tool reword a quote into words the person did not say, and check every figure against your data. See Module 2.
A story told in the person's languageA plain draft for a bilingual and storyteller review
Draft a plain-language version of a story for review
Put this consented story into plain language at a simple reading level for a newsletter, keeping the meaning and the person's own emphasis. Do not add details, and flag any part where the wording might change the meaning. [paste the consented story, no identifying details]
Guardrail. A qualified bilingual person and the storyteller both review it, and consent covers this version specifically, before it is shared. See Module 3.
Worked example
A quote nobody said
You give the model a few true facts about a client’s outcome and ask for a short story for a newsletter.
The draft is moving and specific, with a heartfelt quote in the client’s voice and a round statistic.
Neither the quote nor the number came from you. The model invented both to make the story land.
You cut the invented quote and number, keep only what is true and consented, and let the person review it. AI arranged your facts; it should not manufacture new ones.
Guardrail. You are the conduit for someone’s story, not its author. Keep consent, accuracy, and dignity with you, and let AI help only with the wording. This connects to AI bias and fairness, since an invented detail often follows a stereotype. See also Module 4, and Analyze and visualize your program data for the numbers behind the story.
Sources
Maynez, J., Narayan, S., Bohnet, B., & McDonald, R. (2020). On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 1906-1919). Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.173
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38. https://doi.org/10.1145/3571730
Abid, A., Farooqi, M., & Zou, J. (2021). Persistent anti-Muslim bias in large language models. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (pp. 298-306). https://doi.org/10.1145/3461702.3462624
Bianchi, F., Kalluri, P., Durmus, E., Ladhak, F., Cheng, M., Nozza, D., Hashimoto, T., Jurafsky, D., Zou, J., & Caliskan, A. (2023). Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (pp. 1493-1504). https://doi.org/10.1145/3593013.3594095
Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023).