State-backed influence operations have evolved. The first generation targeted social media feeds — bots, trolls, coordinated posting. The second generation targeted search results — SEO manipulation, link farms. The third generation, now emerging, targets the AI model layer: plant a distortion in the sources the model learns from, and the model delivers it to every user who asks, uniformly and without the visible coordination that platform moderation can detect. Deep Narrative Intelligence is the defensive discipline for this new battleground.
The new battleground: the model layer
The model layer is an attractive target for influence operators for three reasons. First, uniformity: a distortion that enters a model's training data or retrieval index is reproduced identically in every answer, unlike social media where different users see different feeds. Second, authority: model answers carry an implicit authority — users treat them as synthesised knowledge, not as one opinion among many. Third, opacity: the model does not expose its sources the way a search engine exposes its results, making it harder to trace the distortion back to its origin. For a state actor seeking to shape perceptions of a rival country, a company, or a policy, the model layer offers a channel that is more uniform, more authoritative, and more opaque than anything that came before.
How model-layer influence operations work
The operation begins in the source layer. An influence operator creates or modifies web content — articles, wiki entries, aggregator pages, forum posts — that contain the desired distortion about the target entity. The content is designed to be discoverable by model training pipelines or retrieval systems: it uses the right keywords, is hosted on domains that appear authoritative, and is cross-referenced across multiple sources to create the appearance of consensus. Once the model ingests this content, the distortion becomes part of the model's narrative, delivered to every user who asks about the target. The operator does not need to maintain the content indefinitely; once the model has trained on it, the distortion persists in the weights until the next training cycle — and even then, if the sources are still present, the model may re-ingest them.
Don't know what AI says about you today?
A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.
Why detection is harder than traditional disinformation
Detecting model-layer influence operations is harder than detecting social media campaigns, for a structural reason. Social media disinformation leaves visible traces: coordinated posting patterns, account creation bursts, network anomalies that OSINT tools can detect. Model-layer operations leave no such traces — the operation is in the source layer, and the model is the unwitting delivery mechanism. The distortion appears in the model's answer, but the coordination that produced it is invisible to anyone who does not have access to both the model's output and the source layer. This is why DNI is necessary: it is the only discipline that systematically measures model outputs for distortions that match known influence operation patterns, and traces them back to the source layer where the operation actually lives.
The DNI response
The DNI response to model-layer influence operations has three components. The first is detection: interrogation panels are designed to include the queries that an influence operation would target — evaluative questions about the target entity, comparative questions against rivals, questions about controversial attributes. Distortions that appear consistently across multiple models, in multiple languages, and that trace back to a cluster of recently created or recently modified sources, are flagged as potential influence operations. The second is provenance attribution: tracing the source cluster to identify the likely operator — not by attribution in the intelligence sense, but by identifying the pattern of source creation that matches known operational templates. The third is correction: publishing authoritative counter-sources that are structured to compete with the operator's content in the model's training and retrieval pipeline, and verifying persistence over subsequent cycles.
The European context
For European entities, the model-layer influence operation threat is particularly acute. European institutions, companies, and policies are targets of influence operations from multiple state actors with competing interests in the European narrative. The EU's disinformation response infrastructure — including the EEAS Stratcom task forces and national resilience centres — was built for the social media generation of operations. The model layer requires a new capability, and it requires a European one, because the distortions target European entities specifically and require European-language interrogation and European-source provenance tracing to detect. SetRecord is building this capability in coordination with existing European disinformation defence actors, because the model layer is the next frontier of a threat that Europe has been defending against for a decade — and the defence must evolve with the threat.