diff --git a/data/category/neural/v3/image_embeddings_model_thresholds.json.gz b/data/category/neural/v3/image_embeddings_model_thresholds.json.gz
deleted file mode 100644
index 567b4959cb..0000000000
--- a/data/category/neural/v3/image_embeddings_model_thresholds.json.gz
+++ /dev/null
@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a24c0f17bb63f4e1f1b047d10f18fc2228ce22c0cfe52f655f6e07334c05be49
-size 18783
diff --git a/robotoff/insights/importer.py b/robotoff/insights/importer.py
index 81f8aa2002..3170802762 100644
--- a/robotoff/insights/importer.py
+++ b/robotoff/insights/importer.py
@@ -34,7 +34,6 @@
     InsightImportResult,
     InsightType,
     JSONType,
-    NeuralCategoryClassifierModel,
     ObjectDetectionModel,
     PackagingElementProperty,
     Prediction,
@@ -826,17 +825,9 @@ def generate_candidates(
         ]
         taxonomy = get_taxonomy(InsightType.category.name)
 
-        # Make sure we yield candidates with `above_threshold=True` even if
-        # there are candidates that are deepest in the category taxonomy
         yield from (
             ProductInsight(**candidate.to_dict())
             for candidate in select_deepest_taxonomized_candidates(candidates, taxonomy)
-            if candidate.data.get("above_threshold") is True
-        )
-        yield from (
-            ProductInsight(**candidate.to_dict())
-            for candidate in select_deepest_taxonomized_candidates(candidates, taxonomy)
-            if candidate.data.get("above_threshold", False) is False
         )
 
     @staticmethod
@@ -869,17 +860,6 @@ def add_optional_fields(cls, insight: ProductInsight, product: Optional[Product]
             # campaign tag
             campaigns.append("agribalyse-category")
 
-        if (
-            insight.predictor == "neural"
-            and insight.data.get("model_version")
-            == NeuralCategoryClassifierModel.keras_image_embeddings_3_0.value
-            and insight.data.get("above_threshold", False)
-        ):
-            # Add `v3-categorizer-automatic-processing` campaign to category
-            # insights that will be applied automatically soon
-            # (experimental phase)
-            campaigns.append("v3-categorizer-automatic-processing")
-
         if product and not product.categories_tags:
             # Add a campaign to track products with no categories filled in
             campaigns.append("missing-category")
diff --git a/robotoff/prediction/category/neural/category_classifier.py b/robotoff/prediction/category/neural/category_classifier.py
index 013462be44..63d2073aeb 100644
--- a/robotoff/prediction/category/neural/category_classifier.py
+++ b/robotoff/prediction/category/neural/category_classifier.py
@@ -171,27 +171,12 @@ def predict(
             image_embeddings=image_embeddings,
             category_taxonomy=self.taxonomy,
         )
-
-        # Threshold for automatic detection, only available for
-        # `keras_image_embeddings_3_0` model.
-        # Currently we don't apply yet the category automatically, we only add
-        # a flag to add a specific annotation campaign during the insight
-        # import
-        thresholds = (
-            (keras_category_classifier_3_0.get_automatic_processing_thresholds())
-            if model_name.keras_image_embeddings_3_0
-            else {}
-        )
-
         predictions = []
 
         for category_id, score, neighbor_predictions in raw_predictions:
             if category_id not in self.taxonomy:
                 # If the category no longer exist in the taxonomy, ignore it
                 continue
-            # If the category is not in `thresholds` or if the score is
-            # below the threshold, set the above_threshold flag to False
-            above_threshold = score >= thresholds.get(category_id, 1.1)
             kwargs: dict[str, Any] = (
                 {}
                 if neighbor_predictions is None
@@ -203,13 +188,6 @@ def predict(
                     score,
                     model_name.value,
                     product_id=product_id,
-                    above_threshold=above_threshold,
-                    # We need to set a higher priority (=lower digit) if
-                    # above_threshold is True, as otherwise a deepest
-                    # predicted category with `above_threshold=False` will
-                    # take precedence, and we wouldn't generate any insight
-                    # for the prediction with `above_threshold=True`
-                    priority=0 if above_threshold else 1,
                     **kwargs,
                 )
             )
diff --git a/robotoff/prediction/category/neural/keras_category_classifier_3_0/__init__.py b/robotoff/prediction/category/neural/keras_category_classifier_3_0/__init__.py
index b60c34faed..01e74006f5 100644
--- a/robotoff/prediction/category/neural/keras_category_classifier_3_0/__init__.py
+++ b/robotoff/prediction/category/neural/keras_category_classifier_3_0/__init__.py
@@ -1,4 +1,3 @@
-import functools
 from typing import Literal, Optional
 
 import numpy as np
@@ -16,13 +15,12 @@
     serialize_byte_tensor,
 )
 from robotoff.types import JSONType, NeuralCategoryClassifierModel, ProductIdentifier
-from robotoff.utils import get_image_from_url, get_logger, http_session, load_json
+from robotoff.utils import get_image_from_url, get_logger, http_session
 
 from .preprocessing import (
     IMAGE_EMBEDDING_DIM,
     MAX_IMAGE_EMBEDDING,
     NUTRIMENT_NAMES,
-    V3_MODEL_DATA_DIR,
     generate_inputs_dict,
 )
 
@@ -225,20 +223,6 @@ def fetch_ocr_texts(product: JSONType, product_id: ProductIdentifier) -> list[st
     return ocr_texts
 
 
-@functools.cache
-def get_automatic_processing_thresholds() -> dict[str, float]:
-    """Return a dict mapping category ID to minimum detection threshold
-    required to be able to process the insight automatically.
-    Only available for the current default model,
-    `keras_image_embeddings_3_0`.
-
-    The threshold was selected category-wise as the lowest threshold for which
-    we have a precision >= 0.99 on validation + test dataset for this
-    category.
-    """
-    return load_json(V3_MODEL_DATA_DIR / "image_embeddings_model_thresholds.json.gz", compressed=True)  # type: ignore
-
-
 # In NeighborPredictionType objects, we stores the score of parents, children
 # and sibling categories (relative to the predicted categories). Under each
 # type (siblings, children, parents), we store scores as a dict mapping the