what is detection routing?

the idea is older than AI-content detection by decades. in 1991, Jacobs, Jordan, Nowlan, and Hinton described a system “composed of several different ‘expert’ networks plus a gating network that decides which of the experts should be used for each” case. that paper, “Adaptive Mixtures of Local Experts,” is the root of what the field now calls mixture-of-experts: pair a set of specialist sub-models with a small router, then let the router pick who handles each input. only the chosen experts run, which is what keeps the computation sparse rather than dense.

the idea scaled. Shazeer et al. (2017) built a trainable gating network that selects “a sparse combination of these experts to use for each example,” reporting greater than 1000x gains in model capacity at minor cost in efficiency. that is a capacity-and-compute claim, not a detection-accuracy one. Switch Transformer (Fedus, Zoph & Shazeer, 2021) cut the routing down further, sending each input to a single expert and reporting up to roughly 7x pre-training speedup at constant compute per token. again: a speed-and-scale figure from the authors, about language models, not about catching fakes.

routing is not an ensemble

an ensemble runs many models and combines their outputs. routing decides which models to run in the first place. the difference is selection versus averaging, and the two can be stacked: route to a subset, then combine what comes back. a cascade is the cheapest routing variant, running a fast model first and escalating to a stronger one only when confidence is low. FrugalGPT (Chen, Zaharia & Zou, 2023) is the canonical example, an “LLM cascade which learns which combinations of LLMs to use for different queries,” with the authors reporting up to 98% cost reduction against GPT-4. that headline is a self-reported cost figure, not a measure of detection accuracy.

why route a detection scan

because no single model is best at everything. system-level routing, the same idea moved up from internal layers to whole models, aims to “route each query to the model that is most likely to produce a correct answer,” as RouterBench (Hu et al., 2024) frames it, since “no single model can optimally address all tasks.” detection has the same shape. a 2025 review of AI-image detection (Mahara & Rishe) notes that generative models leave distinct, sometimes generator-specific fingerprints, yet detectors “struggle with generalization, often misclassifying images from unseen generative models as real.” a detector strong on one generator family can be weak on the next. routing leans toward the detectors that tend to be strongest for an input's likely generator instead of diluting that signal by averaging everything equally.

routing does not solve the generalization problem, and it would be dishonest to say it does. the mapping from mixture-of-experts and LLM routing onto “send this scan to the best detector for the likely generator” is an explanatory analogy, not a benchmarked technique with a peer-reviewed accuracy number attached. cross-generator generalization stays genuinely hard, and a new generator can degrade older detectors on contact. that is exactly why any verdict here reads as a probabilistic estimate, a best guess, never proof.

this is how amige. reasons about a scan. it treats each available detector as an expert and steers the input toward the ones that tend to perform best for it, runs a panel of independent detectors built by different teams, names the likely maker as a best guess (“looks like” rather than “is”), and returns “uncertain” when the reads conflict instead of forcing a call. as generators change, the routing gets recalibrated. you can follow the whole path end to end in the machine.