Related: sources · notes · metadata · Published Pieces

News Is an Adversarial Game

In an adversarial game, the answer is not better summaries. The answer is verification infrastructure.

News is not a neutral flow of facts waiting to be summarized.

News is an adversarial game.

Something happens. Immediately, actors with different incentives attempt to define what happened, why it happened, who is responsible, what it means, what should be done, and what should be ignored. Some actors lie. Some tell partial truths. Some are sincere but wrong. Some are early. Some are late. Some are paid. Some are captured. Some are afraid. Some optimize for clicks. Some protect institutions. Some prosecute enemies. Some preserve access. Some try to understand.

AI intensifies all of this.

If it becomes cheap to generate persuasive claims, claims multiply. If it becomes cheap to generate plausible analysis, analysis multiplies. If it becomes cheap to make fake images, fake clips, fake experts, fake consensus, fake outrage, and fake balance, the fog thickens. The old media problem was scarcity of trustworthy reporting. The new media problem is abundance of plausible interpretation.

In an adversarial game, the answer is not merely better summaries. A summary can be captured by the frame it summarizes. If the inputs are poisoned, compressed poison remains poison.

The answer is verification infrastructure.

Verification does not mean a single authority declares truth. That is too brittle and politically naïve. Verification means preserving the structure of claims and evidence so disagreement becomes inspectable.

Who made the claim? When? What source supports it? What source contradicts it? What was known at the time? What changed later? Who predicted this? Who dismissed it? Who corrected themselves? Who doubled down? Which sources have a strong record? Which sources are laundering uncertainty? Which institutions have conflicts? Which frames are being repeated without evidence?

This is the kind of work AI should do for news: not replace journalists with generic summaries, not generate fake anchors, not produce artificial podcasts with cheerful hosts. The real work is building and maintaining the claim graph.

A claim graph is adversarial memory. It does not assume consensus. It records conflict. It lets users traverse competing explanations while preserving provenance. It lets agents cite prior work automatically. It lets track records accumulate. It turns public speech into something closer to a statted field.

Every serious news system will need agentic search. Search cannot merely return links. It has to retrieve relevant prior artifacts, source trails, contradictions, forecasts, corrections, and people who saw the issue earlier. The user should be able to ask: who was early on this? what is the strongest counterframe? what did this person say before? what changed since last week? which claim is least supported? where is the original source?

In the old media environment, most people could not ask these questions at scale. They relied on journalists, editors, experts, or social feeds. In the AI media environment, every serious reader will expect the information system to maintain the adversarial record.

News is adversarial because incentives are adversarial. States want legitimacy. Companies want valuation. Politicians want power. Founders want narrative. Journalists want audience and access. Activists want mobilization. Investors want edge. Influencers want attention. Platforms want retention. Ordinary people want dignity, identity, and some chance of not being fooled.

The automatic newspaper does not abolish these incentives. It makes them visible enough to reason about.

That is the best we can do. It is also a lot.