Publishers do not need a perfect AI detector before they can explain how a piece was made. They need a better byline.

That became clear in the dispute involving the Financial Times and Ricardo Hausmann. According to Semafor’s reporting, the FT said Hausmann broke its rules by using AI to “condense” his work. One modest verb carried an unreasonable amount of weight. Did the model cut repetition, reorganize the argument, rewrite passages, select evidence, or alter the author’s meaning? Who reviewed the result? What did the publication finally stand behind?

The controversy invites a binary verdict: human or AI. “Condense” refuses to provide one.

My read is that publishers are treating this as an authorship-policing problem when it has become a reading-experience problem. AI assistance is entering ordinary editorial production. The buildable opportunity is a concise provenance signal, with deeper context available on demand, that tells readers what the system did, what human judgment remained, and who owns the finished work.

“AI-assisted” collapses the useful distinctions

Frontier models can transcribe interviews, translate notes, identify repetition, test structures, shorten drafts, suggest headlines, and rewrite complete passages. Used well, these capabilities can help domain experts publish faster and give editors more room to focus on argument, evidence, and audience.

That capability is worth using. Treating every model interaction as contamination would punish legitimate experimentation and encourage quieter, less accountable use.

The trouble begins when materially different actions receive the same label. “AI-assisted” can describe light compression of an original draft or the generation of its central argument. “Human-written” can obscure a workflow filled with machine transcription, research support, translation, and editing. Neither phrase tells the reader which decisions mattered.

We have already argued that an AI label can tell the truth and still reduce trust when it collapses distinct forms of assistance into one warning. The next design question is more useful: what should replace that blunt verdict?

Researchers are now examining how AI disclosures affect credibility and engagement across topics. That is the right level of inquiry because disclosure is interpreted against what the audience thought it was receiving. Assistance that feels appropriate for translation may feel very different when applied to analysis, testimony, or reported facts.

A reader does not need an exhaustive machine log. The reader needs the distinctions that define the editorial promise.

Put progressive disclosure on the page

Progressive disclosure is a familiar product pattern. Show the essential information first, then let interested users inspect the detail. Applied to publishing, it could turn provenance from a policy footnote into part of the reading experience.

A short signal near the byline might read:

AI assistance: A language model was used to shorten and reorganize the author’s original draft. The author supplied the evidence, analysis, and conclusions. An editor reviewed and approved the final text.

An expanded view could explain which stages involved AI, whether generated language remained in the published version, how source material was handled, and which editorial role approved the result. The concise layer protects the flow of the article. The deeper layer serves readers for whom process is material.

The production workflow can support this without becoming paperwork theater. A publishing system could record a few editorially meaningful categories as assistance occurs, such as transcription, translation, structural editing, condensation, generative drafting, or image creation. An agent could assemble the reader-facing note from that record. An editor would approve the description alongside the article.

This is an AI-native product primitive: let the system handle recurring synthesis while keeping the publication decision with an accountable person. The same structured provenance could travel with newsletters, syndication, and other versions of the piece.

The new hypothesis is specific. Transparency alone is not the product. A layered interface that explains material assistance is.

A detector cannot explain an editorial relationship

Substack’s Pangram-powered Scan is an instructive audience-facing example. It shows that platforms recognize the demand for visible authorship signals. Detection can be useful for triage or as one clue, but a detector score cannot explain why a model was used, whether that use was authorized, what changed, or who accepted responsibility for the result.

It identifies a textual pattern. The publisher knows the editorial relationship.

The Associated Press’s updated newsroom standards point to the operating side of the same design. Standards can define acceptable assistance, preserve human review, and keep responsibility with the newsroom. A policy governs behavior inside the production process. The interface explains the published result to the reader.

That distinction also keeps the promise honest. A provenance note can describe how an article was made. It cannot certify that every claim is true or that the argument is sound. The publisher can explain its process and stand behind its decision. The reader still has to judge the work.

This matters because AI can reproduce the surface cues of a trusted genre with increasing ease. A column can look like a column. A testimonial can sound intimate. A reported feature can carry the expected rhythm. Process legibility gives publishers another way to state what those familiar forms no longer prove on their own.

The commercial test is the second click

Chartbeat reported that overall web traffic fell 6% in 2025 while internal navigation accounted for 41% of pageviews. Those figures do not establish any causal relationship between provenance and retention. They do clarify where publishers have room to create value: after a reader arrives.

I keep coming back to the second click. A visitor may land on an article through search, social media, a newsletter, or an AI-generated answer. The publisher does not control all of that arrival context. It can shape what the reader encounters next.

Provenance could contribute if it communicates editorial confidence rather than risk aversion. A concise explanation says that the publication understands its own production process, can distinguish assistance from delegation, and is willing to name where responsibility sits. That turns process legibility into an expression of the publisher’s brand.

The differentiator does not have to be machine purity. It can be a recognizable standard for how capable tools are used and where human editorial judgment remains decisive.

Whether that supports recirculation, direct return, or subscription intent remains a product hypothesis. The thing to watch is what readers do after encountering the signal. Do they open the detail? Do they continue to another article? Does the standard become familiar enough to inform a later choice?

Provenance will not create demand for a weak story. It could reduce uncertainty at the moment a reader decides whether this publication deserves more attention. That is narrower than a promise of trust, but commercially more useful than a generic badge.

Explain the work, then stand behind it

A useful provenance interface has to be concise enough to scan, specific enough to distinguish material forms of assistance, and restrained enough not to claim more than the process can show.

The tool can provide a signal. The publisher can explain its choices and name the accountable role. The reader judges the finished work. Keeping those responsibilities separate avoids both false confidence and reflexive suspicion.

It also creates room for more ambitious AI use. Models can accelerate editing, adaptation, translation, and production without forcing every article into a fictional choice between untouched human authorship and machine generation. Clear role boundaries make experimentation easier to defend because the publication can describe what actually happened.

AI may increasingly mediate the first arrival. The second click is still a choice. The valuable signal is not that no machine touched the work. It is that the publisher can explain the work clearly enough to earn that next choice.

Sources and further reading