diff --git a/README.md b/README.md
index c51a1b1..f7d9095 100644
--- a/README.md
+++ b/README.md
@@ -21,6 +21,7 @@ The app uses a vertical scrolling learning experience inspired by modern short-f
 - Track progress, bookmarks, and study activity
 - Navigate with keyboard shortcuts
 - Use offline sandbox fallback when backend is unavailable
+- Show user-safe fallback messages when AI generation is unavailable
 
 ## Stellar/Soroban Exploration Roadmap
 
diff --git a/backend/main.py b/backend/main.py
index b0ab609..c2bf486 100644
--- a/backend/main.py
+++ b/backend/main.py
@@ -24,6 +24,10 @@
 APP_VERSION = "1.0.0"
 MAX_PDF_BYTES = 10 * 1024 * 1024
 MAX_NOTES_CHARS = 15000
+AI_FALLBACK_MESSAGE = (
+    "AI generation is temporarily unavailable, so Study Copilot is showing offline practice cards instead."
+)
+AI_FALLBACK_GUIDANCE = "You can keep studying now and try live AI generation again later."
 
 GOOGLE_API_KEY = os.getenv("GOOGLE_API_KEY")
 MODEL_CANDIDATES = [
@@ -121,8 +125,7 @@ def get_fallback_cards(topic: str) -> list[dict[str, Any]]:
             "type": "concept",
             "title": f"Introduction to {topic}",
             "content": (
-                f"Welcome to {topic}. This fallback card appears when Gemini is unavailable. "
-                "Set GOOGLE_API_KEY to enable live AI-generated study cards."
+                f"Welcome to {topic}. {AI_FALLBACK_MESSAGE} {AI_FALLBACK_GUIDANCE}"
             ),
             "timestamp": now_iso(),
         },
diff --git a/docs/ai-error-message-policy.md b/docs/ai-error-message-policy.md
new file mode 100644
index 0000000..9f0f737
--- /dev/null
+++ b/docs/ai-error-message-policy.md
@@ -0,0 +1,32 @@
+# AI Error Message Policy
+
+Study Copilot should keep learning flows usable when AI generation is unavailable while avoiding technical or sensitive details in user-facing responses.
+
+## Current User-Facing Behavior
+
+- The backend starts in offline sandbox mode when no AI provider is configured or provider setup fails.
+- Topic, notes, and PDF study-card generation return offline fallback cards when live AI generation cannot be used.
+- Fallback cards explain that live AI generation is temporarily unavailable and that offline practice cards are being shown instead.
+- Responses do not expose API keys, environment variable values, provider stack traces, raw model errors, or internal exception text.
+
+## Safe Message Pattern
+
+Use concise messages that tell the learner what happened and what they can do next:
+
+> AI generation is temporarily unavailable, so Study Copilot is showing offline practice cards instead. You can keep studying now and try live AI generation again later.
+
+This wording is intentionally provider-neutral. It avoids naming secrets, configuration keys, internal model IDs, request payloads, or exception details.
+
+## Implementation Notes
+
+- Keep provider setup and generation exceptions out of API response bodies.
+- Prefer stable fallback content over passing raw exception messages to the frontend.
+- If structured error metadata is added later, use stable reason codes such as `ai_generation_unavailable` instead of raw exception strings.
+- Backend logs may include operational details for maintainers, but logs should still avoid API key values and full prompt payloads.
+
+## Review Checklist
+
+- Does the message help the user continue or retry?
+- Does it avoid stack traces, raw exception text, and secret names or values?
+- Does it avoid exposing full prompts, notes, uploaded document text, or generated provider responses?
+- Does offline fallback still work without a real AI API key?