AI models give materially wrong answers to US election questions roughly 90% of the time, according to a study cited in a Bloomberg Law opinion piece by Cesca Antonelli, published via The National Law Review. The same study found that AI models cited state-owned media in foreign policy answers about 35% of the time overall, with ChatGPT pulling from such sources in over half its responses. These are not edge-case failures. They are structural ones, and they point directly at a problem the AI industry has spent considerable effort not discussing publicly.
The framing Antonelli uses is worth taking seriously: journalism is becoming critical infrastructure that AI answers are invisibly built on, in the same way that data centres depend on power grids and water supplies. When the underlying infrastructure degrades, everything built on top of it degrades too. That analogy is accurate, and it is also the part the official story leaves out.
How we got here
The business logic is straightforward, if not flattering. AI model providers have, in most cases, either negotiated weak licensing deals with news publishers or avoided licensing arrangements entirely. Every dollar not paid for quality reported content is a dollar saved on training data costs. The models get built, the publishers get little or nothing, and the downstream accuracy problem lands somewhere else entirely — on the businesses and individuals relying on AI-generated answers.
Original reporting meanwhile operates on economics that were already under severe pressure before large language models arrived. Shrinking newsrooms produce less original reporting. Less original reporting means fewer high-quality, factually grounded sources for AI systems to train on and cite. As Antonelli argues, the result is a race to the bottom: AI reliability degrades as the journalism beneath it shrinks, and the journalism shrinks partly because AI companies have not funded its continuation at scale.
The 35% state-owned media citation figure is particularly instructive here. That is not a bug a product update will fix. It reflects what happens when models are trained and retrieval systems are built without disciplined source hierarchies — when the distinction between independent reporting and state-directed content is not treated as a hard constraint. ChatGPT citing state-owned media in over half its foreign policy answers is not a neutral technical outcome. It is the consequence of specific decisions about what counts as acceptable training data and what does not.
The strongest case for the other side
The counterargument from the AI industry runs roughly as follows: models are improving rapidly, retrieval-augmented generation now lets systems pull from live and licensed sources, and accuracy benchmarks are rising quarter on quarter. Publishers, on this view, benefit from the traffic and visibility that AI-surfaced citations generate. The infrastructure problem is real but temporary, and market pressure will push model providers toward better sourcing as enterprise customers demand it.
That argument is not without merit on the trajectory point. But it sidesteps the compounding nature of the current situation. If quality journalism continues contracting during the period when AI companies are not paying adequately for it, the future models those same companies build will have a diminished pool of high-quality material to work from, regardless of how sophisticated the retrieval layer becomes. Retrieval-augmented generation pulls from what exists. If what exists has narrowed and degraded, cleaner retrieval still returns worse answers. The infrastructure argument does not go away because the technology improves.
What this means for content and marketing teams
For marketing directors and content leads, the implication splits into two distinct problems, and conflating them is a mistake most teams currently make.
The first problem is reliability. If your team uses AI tools to research markets, check facts, summarise competitive landscapes, or draft content grounded in current events, the 90% error rate on election-related questions and the state-owned media sourcing data should recalibrate your verification assumptions immediately. AI outputs on any topic touching regulation, policy, public figures, or geopolitics carry structural accuracy risk that generic disclaimers about hallucinations do not fully capture. Build source verification into your workflow as a standing step, not an occasional audit.
The second problem is strategic. Many content teams are currently optimising for AI answer visibility — restructuring content so that AI systems surface it in response to user queries. That is a rational short-term move. But if the foundation those AI systems draw from continues degrading, the answers they generate become less trustworthy, users lose confidence in AI-sourced information, and the channel itself loses authority. Optimising aggressively for a channel whose reliability is structurally at risk is a position worth stress-testing before you commit significant budget to it.
The practical response is not to abandon AI-assisted content work or AI-answer optimisation. It is to treat original, well-sourced reporting — whether your own or from credible outlets — as an asset with increasing relative value, precisely because it is becoming scarcer. Content built on documented primary sources, named experts, and verifiable data is not just better editorially. It is the kind of material that, if AI sourcing practices improve, gets cited. And if they do not improve, it remains accurate and defensible regardless.
The AI industry’s underinvestment in journalism is not an oversight waiting to be corrected. It is the current business model. Marketing and content teams are not bystanders to that arrangement. They are, right now, building strategies on its outputs.
