diff --git a/instrumentation/opentelemetry-instrumentation-genai-anthropic/src/opentelemetry/instrumentation/genai/anthropic/messages_extractors.py b/instrumentation/opentelemetry-instrumentation-genai-anthropic/src/opentelemetry/instrumentation/genai/anthropic/messages_extractors.py
index 05c1769f2..e6cabdf36 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-anthropic/src/opentelemetry/instrumentation/genai/anthropic/messages_extractors.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-anthropic/src/opentelemetry/instrumentation/genai/anthropic/messages_extractors.py
@@ -25,8 +25,9 @@
 from opentelemetry.util.genai.invocation import InferenceInvocation
 from opentelemetry.util.genai.types import (
     InputMessage,
-    MessagePart,
     OutputMessage,
+    SystemInstructionPart,
+    TextPart,
 )
 from opentelemetry.util.types import AttributeValue
 
@@ -125,10 +126,16 @@ def get_input_messages(
 
 def get_system_instruction(
     system: str | Iterable[TextBlockParam] | None,
-) -> list[MessagePart]:
+) -> list[SystemInstructionPart]:
     if system is None:
         return []
-    return convert_content_to_parts(system)
+    if isinstance(system, str):
+        return [TextPart(content=system)] if system else []
+    return [
+        TextPart(content=block["text"])
+        for block in system
+        if block.get("text")
+    ]
 
 
 def get_output_messages_from_message(
diff --git a/instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/extractors.py b/instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/extractors.py
index 88cdaffe3..064fdbca9 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/extractors.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/extractors.py
@@ -4,6 +4,7 @@
 from __future__ import annotations
 
 import json
+from collections.abc import Mapping, Sequence
 from typing import Any, TypeGuard
 from urllib.parse import urlparse
 
@@ -19,6 +20,7 @@
     OutputMessage,
     ReasoningPart,
     Role,
+    SystemInstructionPart,
     TextPart,
     ToolCallRequestPart,
     ToolCallResponsePart,
@@ -235,7 +237,7 @@ def extract_content_block(block: dict[str, Any]) -> MessagePart | None:
         "toolRemoval",
     ):
         if key in block:
-            return GenericPart(type=key, value=None)
+            return GenericPart(type=key)
 
     return None
 
@@ -256,6 +258,30 @@ def _extract_parts(content: Any) -> list[MessagePart]:
     return parts
 
 
+def _extract_system_parts(
+    content: str | Sequence[Mapping[str, Any] | str] | None,
+) -> list[SystemInstructionPart]:
+    if not content:
+        return []
+    if isinstance(content, str):
+        return [TextPart(content=content)] if content else []
+    parts: list[SystemInstructionPart] = []
+    for item in content:
+        if isinstance(item, str):
+            if item:
+                parts.append(TextPart(content=item))
+        elif _is_dict(item):
+            text = item.get("text")
+            if isinstance(text, str) and text:
+                parts.append(TextPart(content=text))
+            else:
+                for key in item:
+                    if key != "text":
+                        parts.append(GenericPart(type=key))
+                        break
+    return parts
+
+
 def _extract_guardrail_id(
     params: dict[str, Any], invocation: InferenceInvocation
 ) -> None:
@@ -324,7 +350,7 @@ def extract_converse_request(
     # system instruction
     raw_system = kwargs.get("system")
     if capture_content and raw_system:
-        system_parts = _extract_parts(raw_system)
+        system_parts = _extract_system_parts(raw_system)
         if system_parts:
             invocation.system_instruction = system_parts
 
@@ -533,7 +559,7 @@ def extract_invoke_model_request(
     # System instruction (e.g. Anthropic / Nova)
     raw_system = body.get("system")
     if raw_system:
-        system_parts = _extract_parts(raw_system)
+        system_parts = _extract_system_parts(raw_system)
         if system_parts:
             invocation.system_instruction = system_parts
 
diff --git a/instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/test_converse.py b/instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/test_converse.py
index 36745dd81..0450fa270 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/test_converse.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/test_converse.py
@@ -27,6 +27,7 @@
 )
 from opentelemetry.trace import StatusCode
 from opentelemetry.util.genai.handler import TelemetryHandler
+from opentelemetry.util.genai.types import GenericPart, TextPart
 
 
 @pytest.mark.vcr
@@ -596,3 +597,47 @@ def test_extract_converse_request_prompt_variables_no_content(
         == "sgi5gkybzqak"
     )
     assert "gen_ai.prompt.variable.user_name" not in invocation.attributes
+
+
+def test_extract_converse_request_system_instruction(tracer_provider) -> None:
+    handler = TelemetryHandler(tracer_provider=tracer_provider)
+    invocation = handler.inference(provider="aws.bedrock")
+
+    extract_converse_request(
+        {
+            "system": [
+                {"text": "Be concise"},
+                {"text": "Answer politely"},
+            ],
+        },
+        invocation,
+    )
+
+    assert invocation.system_instruction == [
+        TextPart(content="Be concise"),
+        TextPart(content="Answer politely"),
+    ]
+
+
+def test_extract_converse_request_system_instruction_generic(
+    tracer_provider,
+) -> None:
+    handler = TelemetryHandler(tracer_provider=tracer_provider)
+    invocation = handler.inference(provider="aws.bedrock")
+
+    extract_converse_request(
+        {
+            "system": [
+                {"text": "Be concise"},
+                {"guardContent": {"guardrailIdentifier": "gr-123"}},
+                {"cachePoint": {"type": "default"}},
+            ],
+        },
+        invocation,
+    )
+
+    assert invocation.system_instruction == [
+        TextPart(content="Be concise"),
+        GenericPart(type="guardContent"),
+        GenericPart(type="cachePoint"),
+    ]
diff --git a/instrumentation/opentelemetry-instrumentation-genai-langchain/.changelog/511.fixed b/instrumentation/opentelemetry-instrumentation-genai-langchain/.changelog/511.fixed
deleted file mode 100644
index ca6b58e10..000000000
--- a/instrumentation/opentelemetry-instrumentation-genai-langchain/.changelog/511.fixed
+++ /dev/null
@@ -1 +0,0 @@
-Divert `SystemMessage` inputs to `gen_ai.system_instructions` instead of emitting them as `role="system"` inside `gen_ai.input.messages`.
diff --git a/instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/callback_handler.py b/instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/callback_handler.py
index 493460592..054e4874c 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/callback_handler.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/callback_handler.py
@@ -40,7 +40,7 @@
     prepare_tool_definitions,
     resolve_response_model_and_id,
     response_fields_from_generation,
-    split_system_and_input_messages,
+    to_input_messages,
 )
 from opentelemetry.util.genai.handler import TelemetryHandler
 from opentelemetry.util.genai.invocation import (
@@ -327,15 +327,14 @@ def on_chat_model_start(
             if "ls_max_tokens" in metadata:
                 max_tokens = metadata.get("ls_max_tokens")
 
-        # Flatten ``list[list[BaseMessage]]`` (one inner list per generation
-        # request) before splitting into system / input.
+        # ``messages`` from on_chat_model_start is ``list[list[BaseMessage]]``
+        # (one inner list per generation request). Flatten and let
+        # :func:`to_input_messages` produce spec-conformant ``InputMessage`` s
+        # with proper roles, tool-call requests, tool results, and reasoning.
         flattened: list[BaseMessage] = [msg for sub in messages for msg in sub]
-        system_instruction: list[MessagePart] = []
         input_messages: list[InputMessage] = []
         if self._telemetry_handler.should_capture_content():
-            system_instruction, input_messages = (
-                split_system_and_input_messages(flattened)
-            )
+            input_messages = to_input_messages(flattened)
 
         llm_invocation = self._telemetry_handler.inference(
             provider,
@@ -343,8 +342,6 @@ def on_chat_model_start(
         )
         llm_invocation.conversation_id = _conversation_id(metadata)
         llm_invocation.input_messages = input_messages
-        if system_instruction:
-            llm_invocation.system_instruction = system_instruction
         llm_invocation.top_p = top_p
         llm_invocation.frequency_penalty = frequency_penalty
         llm_invocation.presence_penalty = presence_penalty
diff --git a/instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/utils.py b/instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/utils.py
index 2a1bfbe06..192fd9a24 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/utils.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/utils.py
@@ -367,41 +367,6 @@ def to_input_messages(
     return result
 
 
-def split_system_and_input_messages(
-    messages: Iterable[Any],
-) -> tuple[list[MessagePart], list[InputMessage]]:
-    """Split ``messages`` into ``system_instruction`` parts and ``InputMessage`` s.
-
-    Called only when content capture is enabled
-    (``TelemetryHandler.should_capture_content()``).
-    """
-    materialized = list(messages)
-    try:
-        normalized: Iterable[BaseMessage] = convert_to_messages(materialized)
-    except Exception:  # pylint: disable=broad-except
-        normalized = [m for m in materialized if isinstance(m, BaseMessage)]
-
-    system_parts: list[MessagePart] = []
-    input_messages: list[InputMessage] = []
-
-    for message in normalized:
-        if isinstance(message, SystemMessage):
-            system_parts.extend(_content_to_parts(message.content))
-        else:
-            parts = _message_parts(message)
-            if not parts and not _has_content(message):
-                continue
-            input_messages.append(
-                InputMessage(
-                    role=_normalize_role(message) or Role.USER.value,
-                    parts=parts,
-                    name=_message_name(message),
-                )
-            )
-
-    return system_parts, input_messages
-
-
 def to_output_messages(
     messages: Iterable[BaseMessage],
     *,
diff --git a/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/_shared.py b/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/_shared.py
deleted file mode 100644
index 1564bf9d2..000000000
--- a/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/_shared.py
+++ /dev/null
@@ -1,21 +0,0 @@
-# Copyright The OpenTelemetry Authors
-# SPDX-License-Identifier: Apache-2.0
-
-"""Helpers shared across langchain conformance scenarios."""
-
-from __future__ import annotations
-
-from typing import Any
-
-from opentelemetry.test.weaver_live_check import LiveCheckReport
-
-
-def span_attribute_values(report: LiveCheckReport, name: str) -> list[Any]:
-    """Return every value of ``name`` across all span samples in ``report``."""
-    return [
-        attr["value"]
-        for entry in report["samples"]
-        if "span" in entry
-        for attr in entry["span"]["attributes"]
-        if attr["name"] == name
-    ]
diff --git a/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/inference.py b/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/inference.py
index 16e7976e8..90c0c109e 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/inference.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/inference.py
@@ -16,15 +16,12 @@
 from opentelemetry.sdk._logs import LoggerProvider
 from opentelemetry.sdk.metrics import MeterProvider
 from opentelemetry.sdk.trace import TracerProvider
-from opentelemetry.test.weaver_live_check import LiveCheckReport
 from opentelemetry.test_util_genai.conformance import (
     ExpectedViolation,
     Scenario,
 )
 from opentelemetry.test_util_genai.instrumentor import instrument
 
-from ._shared import span_attribute_values
-
 
 class InferenceScenario(Scenario):
     expected_spans = {"chat": 1}
@@ -40,20 +37,6 @@ class InferenceScenario(Scenario):
         ),
     )
 
-    def validate(self, report: LiveCheckReport) -> None:
-        super().validate(report)
-        system_instructions = span_attribute_values(
-            report, "gen_ai.system_instructions"
-        )
-        assert len(system_instructions) == 1, (
-            "chat span with a SystemMessage input should set "
-            f"gen_ai.system_instructions once; saw {system_instructions}"
-        )
-        assert "You are a helpful assistant!" in system_instructions[0], (
-            "gen_ai.system_instructions should carry the SystemMessage "
-            f"content; got {system_instructions[0]}"
-        )
-
     def run(
         self,
         *,
diff --git a/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/inference_streaming.py b/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/inference_streaming.py
index 068ec6b1b..65bf0c609 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/inference_streaming.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/conformance/inference_streaming.py
@@ -28,8 +28,6 @@
 )
 from opentelemetry.test_util_genai.instrumentor import instrument
 
-from ._shared import span_attribute_values
-
 
 class InferenceStreamingScenario(Scenario):
     expected_spans = {"chat": 1}
@@ -55,22 +53,17 @@ class InferenceStreamingScenario(Scenario):
 
     def validate(self, report: LiveCheckReport) -> None:
         super().validate(report)
-        stream_values = span_attribute_values(report, "gen_ai.request.stream")
+        stream_values = [
+            attr["value"]
+            for entry in report["samples"]
+            if "span" in entry
+            for attr in entry["span"]["attributes"]
+            if attr["name"] == "gen_ai.request.stream"
+        ]
         assert stream_values == [True], (
             "streaming chat should set gen_ai.request.stream=true on the chat "
             f"span; saw {stream_values}"
         )
-        system_instructions = span_attribute_values(
-            report, "gen_ai.system_instructions"
-        )
-        assert len(system_instructions) == 1, (
-            "streaming chat span with a SystemMessage input should set "
-            f"gen_ai.system_instructions once; saw {system_instructions}"
-        )
-        assert "You are a helpful assistant!" in system_instructions[0], (
-            "gen_ai.system_instructions should carry the SystemMessage "
-            f"content; got {system_instructions[0]}"
-        )
 
     def run(
         self,
diff --git a/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/test_llm_call.py b/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/test_llm_call.py
index d60c1be9a..2be21d95b 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/test_llm_call.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-langchain/tests/test_llm_call.py
@@ -28,7 +28,6 @@
     LangChainInstrumentor,
 )
 from opentelemetry.instrumentation.genai.langchain.utils import (
-    split_system_and_input_messages,
     to_input_messages,
 )
 from opentelemetry.sdk.trace import ReadableSpan
@@ -39,6 +38,7 @@
 from opentelemetry.semconv._incubating.metrics import gen_ai_metrics
 from opentelemetry.semconv.attributes import error_attributes
 from opentelemetry.test_util_genai.instrumentor import instrument
+from opentelemetry.util.genai.types import TextPart
 
 
 def _openai_cassette_name(model, base: str) -> str:
@@ -726,87 +726,39 @@ def test_function_message_role_maps_to_tool():
     assert result[0].role == "tool"
 
 
-def test_split_system_and_input_messages_diverts_system_instructions():
-    system, inputs = split_system_and_input_messages(
+def test_system_message_role_maps_to_system():
+    result = to_input_messages(
         [
             SystemMessage(content="You are helpful."),
             HumanMessage(content="Hi"),
         ]
     )
-    assert len(system) == 1
-    assert system[0].content == "You are helpful."
-    assert system[0].type == "text"
-    assert len(inputs) == 1
-    assert inputs[0].role == "user"
-
-
-def test_split_system_and_input_messages_position_independent():
-    # A ``SystemMessage`` interleaved between non-system messages still lands
-    # in ``system_instruction`` — semconv treats it as a top-level list rather
-    # than an interleaved role, and combining multiple system chunks in order
-    # matches how providers deliver them.
-    system, inputs = split_system_and_input_messages(
-        [
-            SystemMessage(content="First guidance."),
-            HumanMessage(content="Hi"),
-            AIMessage(content="Hello"),
-            SystemMessage(content="Second guidance."),
-            HumanMessage(content="Follow-up"),
-        ]
-    )
-    assert [part.content for part in system] == [
-        "First guidance.",
-        "Second guidance.",
-    ]
-    assert [msg.role for msg in inputs] == ["user", "assistant", "user"]
+    assert len(result) == 2
+    assert result[0].role == "system"
+    assert result[0].parts == [TextPart(content="You are helpful.")]
+    assert result[1].role == "user"
 
 
-def test_split_system_and_input_messages_handles_system_message_chunk():
-    # ``SystemMessageChunk`` (streaming) subclasses ``SystemMessage`` — the
-    # helper matches by ``isinstance`` so streamed system content is diverted
-    # too instead of leaking into ``gen_ai.input.messages``.
-    system, inputs = split_system_and_input_messages(
+def test_system_message_handles_system_message_chunk():
+    result = to_input_messages(
         [
             SystemMessageChunk(content="Streamed system prompt."),
             HumanMessage(content="Hi"),
         ]
     )
-    assert len(system) == 1
-    assert system[0].content == "Streamed system prompt."
-    assert [msg.role for msg in inputs] == ["user"]
+    assert len(result) == 2
+    assert result[0].role == "system"
+    assert result[0].parts == [TextPart(content="Streamed system prompt.")]
+    assert result[1].role == "user"
 
 
-def test_split_system_and_input_messages_empty_when_no_system_message():
-    system, inputs = split_system_and_input_messages(
-        [HumanMessage(content="Hi")]
-    )
-    assert system == []
-    assert [msg.role for msg in inputs] == ["user"]
-
-
-def test_split_system_and_input_messages_normalizes_shorthand_inputs():
-    # Short-hand tuple / dict forms are normalized via ``convert_to_messages``
-    # before the ``isinstance`` partition so that a short-hand system entry is
-    # diverted to ``system_instruction`` instead of leaking into
-    # ``gen_ai.input.messages`` as ``role: "system"``.
-    system, inputs = split_system_and_input_messages(
-        [
-            ("system", "You are helpful."),
-            {"role": "user", "content": "Hi"},
-        ]
-    )
-    assert [part.content for part in system] == ["You are helpful."]
-    assert [msg.role for msg in inputs] == ["user"]
-
-
-def test_split_system_and_input_messages_preserves_name():
-    system, inputs = split_system_and_input_messages(
+def test_to_input_messages_preserves_name():
+    inputs = to_input_messages(
         [
             HumanMessage(content="Hi", name="Alice"),
             AIMessage(content="Hello", name="Bob"),
         ]
     )
-    assert len(inputs) == 2
     assert inputs[0].name == "Alice"
     assert inputs[1].name == "Bob"
 
@@ -916,18 +868,11 @@ def assert_openai_completion_attributes(
     if verify_content:
         input_message = attributes[gen_ai_attributes.GEN_AI_INPUT_MESSAGES]
         assert input_message is not None
-        assert '"role":"system"' not in input_message
+        assert '"role":"system"' in input_message
+        assert '"content":"You are a helpful assistant!"' in input_message
         assert '"role":"user"' in input_message
         assert '"content":"What is the capital of France?"' in input_message
 
-        system_instructions = attributes[
-            gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS
-        ]
-        assert system_instructions is not None
-        assert (
-            '"content":"You are a helpful assistant!"' in system_instructions
-        )
-
         # Assert output message
         output_message = attributes[gen_ai_attributes.GEN_AI_OUTPUT_MESSAGES]
         assert output_message is not None
@@ -937,7 +882,7 @@ def assert_openai_completion_attributes(
     else:
         assert gen_ai_attributes.GEN_AI_INPUT_MESSAGES not in attributes
         assert gen_ai_attributes.GEN_AI_OUTPUT_MESSAGES not in attributes
-        assert gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS not in attributes
+    assert gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS not in attributes
 
 
 def assert_openai_completion_attributes_with_error(
@@ -974,24 +919,17 @@ def assert_openai_completion_attributes_with_error(
     if verify_content:
         input_message = attributes[gen_ai_attributes.GEN_AI_INPUT_MESSAGES]
         assert input_message is not None
-        assert '"role":"system"' not in input_message
+        assert '"role":"system"' in input_message
+        assert '"content":"You are a helpful assistant!"' in input_message
         assert '"role":"user"' in input_message
         assert '"content":"What is the capital of France?"' in input_message
 
-        system_instructions = attributes[
-            gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS
-        ]
-        assert system_instructions is not None
-        assert (
-            '"content":"You are a helpful assistant!"' in system_instructions
-        )
-
         # Assert output message
         assert gen_ai_attributes.GEN_AI_OUTPUT_MESSAGES not in attributes
     else:
         assert gen_ai_attributes.GEN_AI_INPUT_MESSAGES not in attributes
         assert gen_ai_attributes.GEN_AI_OUTPUT_MESSAGES not in attributes
-        assert gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS not in attributes
+    assert gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS not in attributes
 
 
 def assert_bedrock_completion_attributes(
@@ -1231,6 +1169,12 @@ def assert_log_record(log_record, parent_span, response=None):
         attrs.get(gen_ai_attributes.GEN_AI_INPUT_MESSAGES, [])
     )
     expected_input = [
+        {
+            "parts": [
+                {"content": "You are a helpful assistant!", "type": "text"}
+            ],
+            "role": "system",
+        },
         {
             "parts": [
                 {"content": "What is the capital of France?", "type": "text"}
@@ -1238,18 +1182,13 @@ def assert_log_record(log_record, parent_span, response=None):
             "role": "user",
         },
     ]
-    assert len(input_msgs) == 1
+    assert len(input_msgs) == 2
     for i, exp in enumerate(expected_input):
         got = _normalize_to_dict(input_msgs[i])
         assert got["role"] == exp["role"]
         assert _normalize_to_list(got["parts"]) == exp["parts"]
 
-    system_instructions = _normalize_to_list(
-        attrs.get(gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS, [])
-    )
-    assert system_instructions == [
-        {"content": "You are a helpful assistant!", "type": "text"}
-    ]
+    assert gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS not in attrs
 
     output_msgs = _normalize_to_list(
         attrs.get(gen_ai_attributes.GEN_AI_OUTPUT_MESSAGES, [])
@@ -1296,6 +1235,12 @@ def assert_log_record_when_error(log_record, parent_span):
         attrs.get(gen_ai_attributes.GEN_AI_INPUT_MESSAGES, [])
     )
     expected_input = [
+        {
+            "parts": [
+                {"content": "You are a helpful assistant!", "type": "text"}
+            ],
+            "role": "system",
+        },
         {
             "parts": [
                 {"content": "What is the capital of France?", "type": "text"}
@@ -1303,19 +1248,13 @@ def assert_log_record_when_error(log_record, parent_span):
             "role": "user",
         },
     ]
-    assert len(input_msgs) == 1
+    assert len(input_msgs) == 2
     for i, exp in enumerate(expected_input):
         got = _normalize_to_dict(input_msgs[i])
         assert got["role"] == exp["role"]
         assert _normalize_to_list(got["parts"]) == exp["parts"]
 
-    system_instructions = _normalize_to_list(
-        attrs.get(gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS, [])
-    )
-    assert system_instructions == [
-        {"content": "You are a helpful assistant!", "type": "text"}
-    ]
-
+    assert gen_ai_attributes.GEN_AI_SYSTEM_INSTRUCTIONS not in attrs
     assert gen_ai_attributes.GEN_AI_OUTPUT_MESSAGES not in attrs
     assert_log_parent(log_record, parent_span)
 
diff --git a/instrumentation/opentelemetry-instrumentation-genai-llama-index/src/opentelemetry/instrumentation/genai/llama_index/_handler.py b/instrumentation/opentelemetry-instrumentation-genai-llama-index/src/opentelemetry/instrumentation/genai/llama_index/_handler.py
index 6bc34b0de..87801c306 100644
--- a/instrumentation/opentelemetry-instrumentation-genai-llama-index/src/opentelemetry/instrumentation/genai/llama_index/_handler.py
+++ b/instrumentation/opentelemetry-instrumentation-genai-llama-index/src/opentelemetry/instrumentation/genai/llama_index/_handler.py
@@ -45,6 +45,7 @@
     OutputMessage,
     ReasoningPart,
     Role,
+    SystemInstructionPart,
     TextPart,
     ToolCallRequestPart,
     ToolDefinition,
@@ -428,7 +429,7 @@ def new_span(
             )
             tool_definitions = _tool_definitions(instance)
             system_prompt = instance.system_prompt
-            system_instruction: list[MessagePart] = (
+            system_instruction: list[SystemInstructionPart] = (
                 [TextPart(content=system_prompt)]
                 if capture_content and system_prompt
                 else []
diff --git a/instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/interactions.py b/instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/interactions.py
index b6f232cdc..2656c31e4 100644
--- a/instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/interactions.py
+++ b/instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/interactions.py
@@ -213,7 +213,7 @@ def _interactions_input_to_messages(
             )
             parts.append(part)
         elif item_type is not None:
-            part = GenericPart(type=item_type, value=type(item).__name__)
+            part = GenericPart(type=item_type)
             parts.append(part)
 
     return [InputMessage(role=Role.USER.value, parts=parts)]
diff --git a/instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/message.py b/instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/message.py
index cd7f3dd02..aabdf2336 100644
--- a/instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/message.py
+++ b/instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/message.py
@@ -10,10 +10,12 @@
 from opentelemetry.util.genai.types import (
     BlobPart,
     FinishReason,
+    GenericPart,
     InputMessage,
     MessagePart,
     OutputMessage,
     Role,
+    SystemInstructionPart,
     TextPart,
     ToolCallRequestPart,
     ToolCallResponsePart,
@@ -56,11 +58,15 @@ def content_to_output_message(
 def to_system_instructions(
     *,
     content: genai_types.Content,
-) -> list[MessagePart]:
-    parts = (
-        _to_part(part, idx) for idx, part in enumerate(content.parts or [])
-    )
-    return [part for part in parts if part is not None]
+) -> list[SystemInstructionPart]:
+    instructions: list[SystemInstructionPart] = []
+    for idx, part in enumerate(content.parts or []):
+        msg_part = _to_part(part, idx)
+        if isinstance(msg_part, TextPart):
+            instructions.append(msg_part)
+        elif msg_part is not None:
+            instructions.append(GenericPart(type=msg_part.type))
+    return instructions
 
 
 def _to_input_message(
diff --git a/instrumentation/opentelemetry-instrumentation-google-genai/tests/generate_content/test_message.py b/instrumentation/opentelemetry-instrumentation-google-genai/tests/generate_content/test_message.py
new file mode 100644
index 000000000..577335959
--- /dev/null
+++ b/instrumentation/opentelemetry-instrumentation-google-genai/tests/generate_content/test_message.py
@@ -0,0 +1,48 @@
+# Copyright The OpenTelemetry Authors
+# SPDX-License-Identifier: Apache-2.0
+
+import unittest
+
+from google.genai import types as genai_types
+
+from opentelemetry.instrumentation.google_genai.message import (
+    to_system_instructions,
+)
+from opentelemetry.util.genai.types import GenericPart, TextPart
+
+
+class TestGoogleGenAiMessage(unittest.TestCase):
+    def test_to_system_instructions_text(self):
+        content = genai_types.Content(
+            parts=[genai_types.Part.from_text(text="Be concise")]
+        )
+        instructions = to_system_instructions(content=content)
+        self.assertEqual(instructions, [TextPart(content="Be concise")])
+
+    def test_to_system_instructions_non_text_parts(self):
+        content = genai_types.Content(
+            parts=[
+                genai_types.Part.from_text(text="Be concise"),
+                genai_types.Part.from_bytes(
+                    data=b"image_bytes", mime_type="image/png"
+                ),
+                genai_types.Part.from_uri(
+                    file_uri="gs://bucket/doc.pdf",
+                    mime_type="application/pdf",
+                ),
+            ]
+        )
+        instructions = to_system_instructions(content=content)
+        self.assertEqual(
+            instructions,
+            [
+                TextPart(content="Be concise"),
+                GenericPart(type="blob"),
+                GenericPart(type="uri"),
+            ],
+        )
+
+    def test_to_system_instructions_empty_or_unknown_part(self):
+        content = genai_types.Content(parts=[genai_types.Part()])
+        instructions = to_system_instructions(content=content)
+        self.assertEqual(instructions, [])
diff --git a/instrumentation/opentelemetry-instrumentation-google-genai/tests/interactions/test_parser.py b/instrumentation/opentelemetry-instrumentation-google-genai/tests/interactions/test_parser.py
index 0587fc290..cc4a2c539 100644
--- a/instrumentation/opentelemetry-instrumentation-google-genai/tests/interactions/test_parser.py
+++ b/instrumentation/opentelemetry-instrumentation-google-genai/tests/interactions/test_parser.py
@@ -103,7 +103,6 @@ def test_input_to_messages_generic_fallback(self) -> None:
         self.assertEqual(len(messages[0].parts), 1)
         self.assertIsInstance(messages[0].parts[0], GenericPart)
         self.assertEqual(messages[0].parts[0].type, "some_unsupported_type")
-        self.assertEqual(messages[0].parts[0].value, "dict")
 
     def test_input_to_messages_single_non_sequence_step(self) -> None:
         step = {"type": "text", "text": "Hello single step"}
diff --git a/util/opentelemetry-util-genai/.changelog/612.added b/util/opentelemetry-util-genai/.changelog/612.added
new file mode 100644
index 000000000..27efd0908
--- /dev/null
+++ b/util/opentelemetry-util-genai/.changelog/612.added
@@ -0,0 +1 @@
+Add SystemInstructionPart model, align GenericPart with semantic conventions by removing value, and update invocation types.
diff --git a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_agent_invocation.py b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_agent_invocation.py
index 68583bf4f..9c80d2883 100644
--- a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_agent_invocation.py
+++ b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_agent_invocation.py
@@ -20,6 +20,7 @@
     InputMessage,
     MessagePart,
     OutputMessage,
+    SystemInstructionPart,
     ToolDefinition,
 )
 from opentelemetry.util.genai.utils import ContentCapturingMode
@@ -98,7 +99,10 @@ def __init__(
 
         self.input_messages: list[InputMessage] = []
         self.output_messages: list[OutputMessage] = []
-        self.system_instruction: list[MessagePart] = []
+        self.system_instruction: (
+            list[SystemInstructionPart] | list[MessagePart]
+        ) = []
+        """System instructions for the agent. Passing ``MessagePart`` is deprecated; use ``SystemInstructionPart``."""
         self.tool_definitions: list[ToolDefinition] | None = None
 
         self._start(self._get_start_attributes())
diff --git a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_fetch_response_invocation.py b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_fetch_response_invocation.py
index f9f8c49cd..318b3007c 100644
--- a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_fetch_response_invocation.py
+++ b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_fetch_response_invocation.py
@@ -22,6 +22,7 @@
     ErrorTypeResolver,
     MessagePart,
     OutputMessage,
+    SystemInstructionPart,
     ToolDefinition,
 )
 from opentelemetry.util.genai.utils import ContentCapturingMode
@@ -114,7 +115,10 @@ def __init__(
         self.finish_reasons: list[str] | None = None
         self.stream_cursor: str | None = None
         self.output_messages: list[OutputMessage] = []
-        self.system_instruction: list[MessagePart] = []
+        self.system_instruction: (
+            list[SystemInstructionPart] | list[MessagePart]
+        ) = []
+        """System instructions for the model. Passing ``MessagePart`` is deprecated; use ``SystemInstructionPart``."""
         self.tool_definitions: list[ToolDefinition] | None = None
         self._start(self._get_start_attributes())
 
diff --git a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_inference_invocation.py b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_inference_invocation.py
index fdacd54a3..493dca910 100644
--- a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_inference_invocation.py
+++ b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_inference_invocation.py
@@ -23,6 +23,7 @@
     InputMessage,
     MessagePart,
     OutputMessage,
+    SystemInstructionPart,
     ToolDefinition,
 )
 from opentelemetry.util.genai.utils import (
@@ -78,7 +79,10 @@ def __init__(
 
         self.input_messages: list[InputMessage] = []
         self.output_messages: list[OutputMessage] = []
-        self.system_instruction: list[MessagePart] = []
+        self.system_instruction: (
+            list[SystemInstructionPart] | list[MessagePart]
+        ) = []
+        """System instructions for the model. Passing ``MessagePart`` is deprecated; use ``SystemInstructionPart``."""
         self._response_model_name: str | None = None
         self.response_id: str | None = None
         self.finish_reasons: list[str] | None = None
@@ -273,7 +277,9 @@ class LLMInvocation:
     request_model: str | None = None
     input_messages: list[InputMessage] = field(default_factory=list)  # pyright: ignore[reportUnknownVariableType]
     output_messages: list[OutputMessage] = field(default_factory=list)  # pyright: ignore[reportUnknownVariableType]
-    system_instruction: list[MessagePart] = field(default_factory=list)  # pyright: ignore[reportUnknownVariableType]
+    system_instruction: (  # pyright: ignore[reportUnknownVariableType]
+        list[SystemInstructionPart] | list[MessagePart]
+    ) = field(default_factory=list)
     provider: str | None = None
     response_model_name: str | None = None
     response_id: str | None = None
diff --git a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_invocation.py b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_invocation.py
index ee0b242ae..380a6e33b 100644
--- a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_invocation.py
+++ b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_invocation.py
@@ -10,7 +10,7 @@
 from contextvars import Token
 from dataclasses import asdict
 from types import TracebackType
-from typing import TYPE_CHECKING, Any, TypeAlias
+from typing import TYPE_CHECKING, Any, TypeAlias, cast
 
 from typing_extensions import Self
 
@@ -33,6 +33,7 @@
     InputMessage,
     MessagePart,
     OutputMessage,
+    SystemInstructionPart,
     ToolDefinition,
 )
 from opentelemetry.util.genai.utils import (
@@ -198,7 +199,9 @@ def _call_completion_hook(
         *,
         inputs: list[InputMessage] | None = None,
         outputs: list[OutputMessage] | None = None,
-        system_instruction: list[MessagePart] | None = None,
+        system_instruction: list[SystemInstructionPart]
+        | list[MessagePart]
+        | None = None,
         tool_definitions: list[ToolDefinition] | None = None,
         log_record: LogRecord | None = None,
     ) -> None:
@@ -210,7 +213,9 @@ def _call_completion_hook(
         self._completion_hook.on_completion(
             inputs=inputs or [],
             outputs=outputs or [],
-            system_instruction=system_instruction or [],
+            system_instruction=cast(
+                "list[MessagePart]", system_instruction or []
+            ),
             tool_definitions=tool_definitions,
             span=self.span,
             log_record=log_record,
@@ -265,7 +270,7 @@ def get_content_attributes(
     *,
     input_messages: Sequence[InputMessage],
     output_messages: Sequence[OutputMessage],
-    system_instruction: Sequence[MessagePart],
+    system_instruction: Sequence[SystemInstructionPart | MessagePart],
     tool_definitions: Sequence[ToolDefinition] | None,
     for_span: bool,
     content_capturing_mode: ContentCapturingMode | None = None,
@@ -275,7 +280,8 @@ def get_content_attributes(
     Args:
         input_messages: Input messages to serialize.
         output_messages: Output messages to serialize.
-        system_instruction: System instructions to serialize.
+        system_instruction: System instructions to serialize. Passing ``MessagePart``
+            is deprecated; use ``SystemInstructionPart``.
         tool_definitions: Tool definitions to serialize (may be None).
         for_span: If True, serialize for span attributes (JSON string);
                   if False, serialize for event attributes (list of dicts).
diff --git a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_upload/completion_hook.py b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_upload/completion_hook.py
index 9900cf13d..c7df68678 100644
--- a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_upload/completion_hook.py
+++ b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/_upload/completion_hook.py
@@ -11,7 +11,7 @@
 import posixpath
 import threading
 from collections import OrderedDict
-from collections.abc import Callable
+from collections.abc import Callable, Sequence
 from concurrent.futures import (
     Future,
     ThreadPoolExecutor,
@@ -28,8 +28,14 @@
 from opentelemetry._logs import LogRecord
 from opentelemetry.semconv._incubating.attributes import gen_ai_attributes
 from opentelemetry.trace import Span
-from opentelemetry.util.genai import types
 from opentelemetry.util.genai.completion_hook import CompletionHook
+from opentelemetry.util.genai.types import (
+    InputMessage,
+    MessagePart,
+    OutputMessage,
+    TextPart,
+    ToolDefinition,
+)
 from opentelemetry.util.genai.utils import gen_ai_json_dump
 
 GEN_AI_INPUT_MESSAGES_REF: Final = (
@@ -59,10 +65,10 @@
 
 @dataclass
 class Completion:
-    inputs: list[types.InputMessage] | None
-    outputs: list[types.OutputMessage] | None
-    system_instruction: list[types.MessagePart] | None
-    tool_definitions: list[types.ToolDefinition] | None
+    inputs: list[InputMessage] | None
+    outputs: list[OutputMessage] | None
+    system_instruction: list[MessagePart] | None
+    tool_definitions: list[ToolDefinition] | None
 
 
 @dataclass
@@ -80,15 +86,15 @@ class CompletionRefs:
 
 
 def is_message_part_list_hashable(
-    message_parts: list[types.MessagePart] | None,
+    message_parts: Sequence[MessagePart] | None,
 ) -> bool:
     return bool(message_parts) and all(
-        isinstance(x, types.TextPart) for x in message_parts
+        isinstance(x, TextPart) for x in message_parts
     )
 
 
 def hash_tool_definitions(
-    tool_definitions: list[types.ToolDefinition] | None,
+    tool_definitions: list[ToolDefinition] | None,
 ) -> str | None:
     if not tool_definitions:
         return None
@@ -218,8 +224,8 @@ def done(future: Future[None]) -> None:
 
     def _calculate_ref_path(
         self,
-        system_instruction: list[types.MessagePart],
-        tool_definitions: list[types.ToolDefinition] | None = None,
+        system_instruction: Sequence[MessagePart],
+        tool_definitions: list[ToolDefinition] | None = None,
     ) -> CompletionRefs:
         # TODO: experimental with using the trace_id and span_id, or fetching
         # gen_ai.response.id from the active span.
@@ -227,9 +233,11 @@ def _calculate_ref_path(
         if is_message_part_list_hashable(system_instruction):
             # Get a hash of the text.
             system_instruction_hash = hashlib.sha256(
-                "\n".join(x.content for x in system_instruction).encode(  # pyright: ignore[reportUnknownMemberType, reportAttributeAccessIssue, reportUnknownArgumentType, reportCallIssue, reportArgumentType]
-                    "utf-8"
-                ),
+                "\n".join(
+                    x.content
+                    for x in system_instruction
+                    if isinstance(x, TextPart)
+                ).encode("utf-8"),
                 usedforsecurity=False,
             ).hexdigest()
 
@@ -298,10 +306,10 @@ def _do_upload(
     def on_completion(
         self,
         *,
-        inputs: list[types.InputMessage],
-        outputs: list[types.OutputMessage],
-        system_instruction: list[types.MessagePart],
-        tool_definitions: list[types.ToolDefinition] | None = None,
+        inputs: list[InputMessage],
+        outputs: list[OutputMessage],
+        system_instruction: list[MessagePart],
+        tool_definitions: list[ToolDefinition] | None = None,
         span: Span | None = None,
         log_record: LogRecord | None = None,
         **kwargs: Any,
@@ -321,10 +329,10 @@ def on_completion(
         )
 
         def to_dict(
-            dataclass_list: list[types.InputMessage]
-            | list[types.OutputMessage]
-            | list[types.MessagePart]
-            | list[types.ToolDefinition],
+            dataclass_list: list[InputMessage]
+            | list[OutputMessage]
+            | list[MessagePart]
+            | list[ToolDefinition],
         ) -> JsonEncodeable:
             return [asdict(dc) for dc in dataclass_list]
 
diff --git a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/completion_hook.py b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/completion_hook.py
index 83c34164b..bec759d29 100644
--- a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/completion_hook.py
+++ b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/completion_hook.py
@@ -21,10 +21,15 @@
 from opentelemetry.util._importlib_metadata import (
     entry_points,  # pyright: ignore[reportUnknownVariableType]
 )
-from opentelemetry.util.genai import types
 from opentelemetry.util.genai.environment_variables import (
     OTEL_INSTRUMENTATION_GENAI_COMPLETION_HOOK,
 )
+from opentelemetry.util.genai.types import (
+    InputMessage,
+    MessagePart,
+    OutputMessage,
+    ToolDefinition,
+)
 
 _logger = logging.getLogger(__name__)
 
@@ -68,10 +73,10 @@ class CompletionHook(Protocol):
     def on_completion(
         self,
         *,
-        inputs: list[types.InputMessage],
-        outputs: list[types.OutputMessage],
-        system_instruction: list[types.MessagePart],
-        tool_definitions: list[types.ToolDefinition] | None = None,
+        inputs: list[InputMessage],
+        outputs: list[OutputMessage],
+        system_instruction: list[MessagePart],
+        tool_definitions: list[ToolDefinition] | None = None,
         span: Span | None = None,
         log_record: LogRecord | None = None,
     ) -> None: ...
diff --git a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/types.py b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/types.py
index b14846b2c..d041923cd 100644
--- a/util/opentelemetry-util-genai/src/opentelemetry/util/genai/types.py
+++ b/util/opentelemetry-util-genai/src/opentelemetry/util/genai/types.py
@@ -36,17 +36,14 @@ class ContentCapturingMode(Enum):
 
 @dataclass()
 class GenericPart:
-    """Used for provider-specific message part types that don't match
-    the standard MessagePart types defined in semantic conventions. Set ``type``
-    to the provider-specific type discriminator and carry the payload in
-    ``value`` to explicitly opt-in to non-standard types.
-    This will be removed in a future version when all instrumentations use core types.
+    """Represents an arbitrary message part with any type and properties.
+    This allows for extensibility with custom message part types.
 
-    Per the semconv message schema, ``type`` is a free-form string (the
-    provider's own type name), not a fixed literal."""
+    This model is specified as part of semconv in `GenAI messages Python models - GenericPart
+    <https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/non-normative/models.py>`__.
+    """
 
     type: str
-    value: Any
 
 
 @dataclass()
@@ -308,6 +305,9 @@ class Role(str, Enum):
     TOOL = "tool"
 
 
+SystemInstructionPart = Union[TextPart, GenericPart]
+
+
 @dataclass()
 class InputMessage:
     role: str
diff --git a/util/opentelemetry-util-genai/tests/test_utils.py b/util/opentelemetry-util-genai/tests/test_utils.py
index 2670b0fb5..13b382f35 100644
--- a/util/opentelemetry-util-genai/tests/test_utils.py
+++ b/util/opentelemetry-util-genai/tests/test_utils.py
@@ -42,12 +42,14 @@
     ContentCapturingMode,
     File,
     FilePart,
+    GenericPart,
     InputMessage,
     MessagePart,
     OutputMessage,
     Reasoning,
     ReasoningPart,
     Role,
+    SystemInstructionPart,
     Text,
     TextPart,
     Uri,
@@ -403,6 +405,33 @@ def test_llm_start_and_stop_creates_span(self):  # pylint: disable=no-self-use
         )
         self.assertEqual(span_system[0]["type"], "text")
 
+    @patch.dict(
+        os.environ,
+        {
+            "OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT": "SPAN_ONLY",
+            "OTEL_INSTRUMENTATION_GENAI_EMIT_EVENT": "false",
+        },
+    )
+    def test_system_instruction_generic_part_on_span(self):
+        handler = TelemetryHandler(tracer_provider=self.tracer_provider)
+        with handler.inference("test-provider") as invocation:
+            invocation.system_instruction = [
+                TextPart(content="You are helpful"),
+                GenericPart(type="custom"),
+            ]
+
+        span = _get_single_span(self.span_exporter)
+        span_attrs = _get_span_attributes(span)
+        self.assertIn(GenAI.GEN_AI_SYSTEM_INSTRUCTIONS, span_attrs)
+        instructions = json.loads(span_attrs[GenAI.GEN_AI_SYSTEM_INSTRUCTIONS])
+        self.assertEqual(
+            instructions,
+            [
+                {"type": "text", "content": "You are helpful"},
+                {"type": "custom"},
+            ],
+        )
+
     @patch.dict(
         os.environ,
         {
@@ -1113,6 +1142,20 @@ def test_messages_with_name(self):
         serialized_out = json.loads(gen_ai_json_dumps(asdict(out_msg)))
         self.assertEqual(serialized_out["name"], "bot")
 
+    def test_system_instruction_part(self):
+        text_part: SystemInstructionPart = TextPart(content="You are helpful")
+        generic_part: SystemInstructionPart = GenericPart(type="custom")
+        self.assertEqual(text_part.content, "You are helpful")
+        self.assertEqual(generic_part.type, "custom")
+        self.assertEqual(
+            json.loads(gen_ai_json_dumps(asdict(text_part))),
+            {"type": "text", "content": "You are helpful"},
+        )
+        self.assertEqual(
+            json.loads(gen_ai_json_dumps(asdict(generic_part))),
+            {"type": "custom"},
+        )
+
 
 _REAL_PNG_BYTES = (
     b"\x89PNG\r\n\x1a\n\x00\x00\x00\rIHDR\x00\x00\x00\x01\x00\x00\x00"