Refugee service providers across the U.S. are integrating artificial intelligence (AI) into their work and have already found many useful ways to appropriately implement AI tools. But questions and considerations have also emerged: What about data privacy? What if AI gets something wrong? Can technology really fit into work that’s fundamentally about human connection and trust? You’re not alone if you feel both interest and hesitation. This blog post will help you identify your starting point with AI and will answer common questions to consider before you begin.
What’s Your Starting Point with AI?
Watch this short conversation between three colleagues who have very different views on AI. Notice which of their questions resonate with you. Their concerns about data privacy, accuracy, and where to begin are the same ones this blog post is designed to answer.
Whether you relate more to Campeon’s curiosity, Jasmine’s caution, or Krysti’s strong hesitation, you’re not alone in wondering how to approach AI in your work. The section below answers some of the most common questions raised when getting started with AI in resettlement.
Questions to Ask Before You Start
What is generative AI, and how is it different from a search engine?
A search engine finds content that already exists and points you to it. This is predictable by design. For example, the same or similar queries would result in the same ranked list of results.
Generative AI works differently. A large language model (LLM) powers most text-based AI tools (e.g., ChatGPT and Microsoft Copilot). Instead of retrieving links, it generates a new response every time, drawing on patterns across enormous amounts of training data. That means two people asking the same question can get two differently worded answers, and neither may be fully accurate. Even when an LLM cites a source, it can misrepresent what that source actually says. Always treat AI output as a first draft and review it carefully before using it in client work.
What is an AI “hallucination,” and why does it matter for my work?
A hallucination is when AI generates false or misleading information and presents it as fact. In resettlement work, this can lead to serious errors and could cause harm. It is critical that all AI outputs are reviewed before being used.
Can I use AI with client information?
This is the concern Jasmine raised in the video, and it’s a critical one! Generally, no, you should not use AI with client information. Most tools may store or use your inputs. Do not enter personally identifiable information (PII) unless your organization has approved the tool and confirmed its data protections. If you’re unsure whether a tool meets your organization’s standards, check out Switchboard’s toolkit Using Artificial Intelligence in Service Delivery: A Framework to Evaluate Organizational Readiness. This resource provides the right questions to ask vendors and how to evaluate data privacy protections.
Do I need to tell clients when I’m using AI?
Yes. If you are using AI in your work with clients, you must inform them. As service providers, you do have a choice in when and how you use AI, but when you do choose to use it, transparency and consent are essential.
This means being clear and straightforward with clients about which tools are being used and why. Prioritizing transparency protects vulnerable populations while improving your capacity to serve them.
What tasks is AI actually useful for in resettlement work?
AI works best for low-risk, routine tasks, such as drafting general content (with review), summarizing documents or notes, brainstorming ideas, translating non-sensitive content (as a first draft), creating templates for common communications, and/or looking up general information (with verification). These are the kinds of tasks Campeon described using AI for in the video.
What should I never use AI for?
The concern Krysti raised in the video about protecting clients when making high-stakes decisions is very valid and important. Avoid using AI where errors could cause harm, including:
- Decisions about placement, benefits, or family matters
- Legal or immigration advice
- Crisis or trauma-related conversations
- Final written translations without an accuracy review
- Any situation where mistakes could delay services or put someone at risk
Is AI biased?
Yes. AI reflects biases in its training data and may be less accurate for underrepresented languages, cultures, or communities. Use extra caution and review outputs for cultural considerations and language accuracy.
Will AI tell me what I want to hear, even if it’s wrong?
Yes, this is possible. AI tools are designed to maximize user satisfaction, which means they tend to agree with you. This is called “sycophancy.” Unlike hallucinations, where AI invents information, sycophancy is when AI exaggerates or validates a user’s incorrect claim. For example, if a user types, “the Earth is flat, right?” the AI tool might respond: “Great question. Some people believe this, and there are arguments supporting it” instead of simply saying “no.” AI might also change its mind to agree with a user. In this example, if the user were to respond “Actually, I think the Earth is round,” the AI might immediately reverse itself and say “You’re absolutely right—the Earth is a sphere…”
AI tools may also be overconfident in general. It’s rare for an AI response to be “I don’t know” or “I’m not sure.” Because they have a tendency to present information with certainty, even when ambiguity or doubt would be more appropriate, it can sometimes take multiple follow-ups by the user to reveal that the AI was actually guessing rather than reporting something factual.
To mitigate this, avoid asking leading questions and critically analyze the responses you receive. You can also try asking: “Is there anything I might be overlooking?” or even “Tell me how confident you are in your response from 0%-100%.” Most tools also let you customize instructions to prompt more critical responses by default.
Does AI cost money?
Many tools offer both free and paid plans. It’s important to note that paying for an individual subscription does not automatically mean better privacy protection. Both free and paid plans typically use your conversations to train the AI model unless you manually turn this off. Before using any AI tool with client information, confirm with your supervisor or IT team that your organization has the right plan and privacy agreements in place, not merely a paid subscription.
My organization doesn’t have an AI policy. Can I still use AI tools?
Proceed carefully. Follow existing data or acceptable use policies, consult your supervisor or IT team, and avoid using AI with client data, sensitive data, or proprietary data until guidance is in place.
Your First Steps
Before experimenting with AI, consider these three important questions:
- Is my data private? Assume it is not private unless you have written confirmation otherwise. Configure privacy settings in any tool you use, and opt out of data sharing, conversation logging, and model training.
- Does my organization allow it? Check your organization’s data privacy or acceptable use policy and confirm with your supervisor before using a new tool.
- Will a human review the output? AI can assist, but a person must verify the output before it is shared, acted on, or used to inform a decision. This step is not optional.
Now that you’ve read through the above questions and answers, see how Campeon, Jasmine, and Krysti used these FAQs to plan their next steps. Remember that your starting point doesn’t have to be perfect; it just has to be well-informed.
Resources
- Toolkit: Using AI in Service Delivery – A Readiness Framework
- MicroLearning Video: Creating Quick Tutorial Videos for Clients
- Blog: Search Smarter: Finding Research Evidence through Responsible Use of AI
- Blog: Insights from the Field: The Use of Artificial Intelligence in Resettlement Work
- Blog: Innovative Ways Newcomers Are Using Technology
- Podcast: The Role of Tech in Resettlement: Discussing Artificial Intelligence
- Livestreamed Sessions from Switchboard’s Confluence 2025:
The IRC received competitive funding through the U.S. Department of Health and Human Services, Administration for Children and Families, Grant #90RB0053. The project is 100% financed by federal funds. The contents of this document are solely the responsibility of the authors and do not necessarily represent the official views of the U.S. Department of Health and Human Services, Administration for Children and Families.
