The Frontier Fractures: Why AI Needs Something to Stand On

What does it mean for a system to “know” something, if five versions of that system, reasoning from the same question, arrive at five different answers?

A new study from Lenz Research (2026) asked five frontier models to fact-check a thousand real-world user claims, and found that on 67%, they did not concur. On a third of all claims, two models landed at least two verdict categories apart, from systems that we broadly consider to be at the cutting edge. While clear-cut True and False verdicts were much more consistent, more nuanced claims fractured almost entirely in the middle; and let’s be honest, this is where most of the consequential topics actually live.

Why does this happen?

Large language models do not retrieve facts so much as reconstruct plausible accounts of them, weighted by patterns in the information they have absorbed or retrieved. And that information environment is not a neutral sample of human knowledge. It is a web-scale corpus where high-engagement content is structurally overrepresented, with information that spreads virally on social media shaping the probabilistic landscape from which AI systems subsequently reason. A rumour shared ten thousand times carries more training signals than a correction shared two hundred times. So when you ask five frontier models a question with an ambiguous, contested claim; the kind that has been noisily debated across the internet, you get five slightly different reconstructions of the surrounding noise, each internally coherent, none obviously wrong, and no clear way to adjudicate between them. More model capability does not resolve this, because a more confident reconstruction is still a reconstruction.

An example of how problematic this issue is, was reported in a 2024 NewsGuard audit, which tested ten leading AI chatbots against Russian disinformation narratives that had been laundered through a network of fake local-news sites. The chatbots repeated the false narratives around a third of the time, often treating the fabricated local outlets, or associated YouTube claims, as credible sources. In that sense, the problem here is not just that AI can generate falsehoods, but that it can inherit and reorganise falsehoods that have already been made to look like reliable knowledge.

This is the context in which I find building infrastructure for verifiable communication genuinely interesting; the idea being that organisations can cryptographically prove what they have published, and that AI systems can verify claims against that original source directly, rather than against a model’s reconstruction of what others have said.

In the example above, disinformation did not need to persuade the whole public directly, it only needed to become available as something the machine could retrieve, summarise, and cite. And while the models are getting better at using preferential sources, a model grounded in a timestamped, authenticated record of what an institution actually published however, is operating in a different epistemic register to one that is reasoning from training data saturated with the social media aftermath of the original claim.

What should one take away from this?

The narrative that bigger models converge on better answers is doing significant work in boardrooms and policy circles right now, and the Lenz findings suggest that convergence is considerably shallower than this narrative implies. As AI takes on a larger role in how information is mediated, the gap between what a model thinks is true and what is verifiably true becomes harder to ignore. Grounding model reasoning in verifiable sources does not resolve that gap by making AI smarter, but by changing the nature of the problem; from inference over engagement-weighted noise, to verification against a known source.

And that distinction matters more than it might initially seem.

For organisations, it is not just an abstract epistemological issue, but fundamentally changes the conditions under which communication has to operate. A public statement, press release, product claim, or crisis update no longer sits only on the organisation’s website, waiting to be read in context. It may be scraped, summarised, translated, quoted, challenged, and reassembled into answers by systems that users increasingly treat as interfaces to their reality. And in that environment, the question is not only whether the organisation has communicated accurately, but whether its communication can still be recognised, verified, and distinguished from the noise that gathers around it.