The stakes could hardly be higher. Public trust in journalism has been declining for years, and news avoidance is rising across democratic societies. Surveys consistently show that audiences are sceptical about AI in news, distrust fully automated content, and want publishers to be open about their use of the technology. Yet the same research reveals a paradox that has come to be known as the transparency backfire: disclosing AI involvement can itself reduce trust in news, making readers question content that was, in fact, produced or checked by human journalists. Publishers are therefore caught between an ethical obligation to disclose and a commercial imperative not to undermine the very credibility that makes people pay for journalism. As one senior FT manager put it bluntly in the study: mistrust is the top threat, because once you lose that, you forget the rest.

The study finds that the FT operationalises transparency through a hybrid of policy, process, and practice, underwritten by explicit endorsement from the very top of the organisation. Editor in Chief Roula Khalaf has publicly committed the newsroom to transparency, promising that all newsroom experimentation will be recorded in an internal register. The FT’s internal AI Principles state that the organisation will not allow AI to compromise the integrity of its journalism, which will be reported, written, and created by journalists and editors, while permitting experimentation with secondary content such as summaries, alt-text, captions, and translations, provided there is appropriate disclosure. New AI applications are vetted by cross-functional AI Use Case Panels, which decide whether a use case is approved and advise on the appropriate level of transparency. Some red lines are absolute: the FT does not generate photographic imagery with AI, except for clearly flagged illustrations in stories about AI itself.

Internally, transparency functions as a trust-building and risk-management mechanism rather than a public relations exercise. The FT’s internal AI Policy places accountability squarely on staff, requiring them to review AI output as if they had created it themselves. If an AI-generated summary contains an error, the journalist whose name appears on the story bears the responsibility just as before. AI functionality inside the content management system is signposted with the now-familiar sparkle emoji, accompanied by reminders that generated content may contain hallucinations. Technical systems even highlight portions of AI output that are statistically more likely to contain errors, directing human attention where it matters most. Regular sharing sessions across editorial, product, and technology teams, an intranet serving as a single source of truth, and an AI Transformation group designed to prevent information silos all reinforce the message. Managers described a culture in which even senior leaders are willing to admit they do not know, backed by what one participant called a lot of data and checks, and kill switches the whole way along.

Externally, the FT calibrates the prominence of disclosure to the degree of autonomy in the system. The clearest illustration is the contrast between two reader-facing features. Ask FT, a generative chatbot that lets subscribers interrogate nearly two decades of archival reporting, operates with no human in the loop, and therefore carries a strong disclaimer informing users they are interacting with an AI system provided by Anthropic rather than a human. Bullet-point summaries, by contrast, are generated by a language model but sit within articles subject to editorial oversight, and receive lighter labelling. The FT’s internal Responsible AI Guidelines codify this logic: the more automatic or impactful a process, or the larger the proportion of output generated by AI, the more prominent the disclosure should be. When in doubt, disclosure is the default option.

The researchers identified nine factors that systematically shape what, when, and how the FT discloses AI use to readers. Legal and provider requirements come first: some AI vendors contractually mandate disclosure, down to Google Translate’s requirement for a specific logo and a paragraph of capitalised text. Industry benchmarking follows, with the FT studying how publishers such as the New York Times, the Washington Post, and Aftonbladet, along with the BBC’s published AI transparency guidelines, approach the same problems. The nature of the task matters too: behind-the-scenes AI used by the search engine optimisation team, whose output never reaches readers directly, goes undisclosed. The degree of human oversight, the novelty and uncertainty of the system, audience expectations gathered through usability testing and pre- and post-deployment research, the perceived risk of harm or error, commercial sensitivities such as the cost of documenting AI methods in investigative work, and finally design constraints, including the limited screen real estate of mobile devices, all weigh on every disclosure decision.

That last point deserves emphasis, because it reveals how mundane the mechanics of transparency can be. Readers, the FT found, find AI disclosures far more obvious on desktop than on mobile, where there is simply less space to call out specific elements. Designers debated the background colour of AI-generated elements to signal that this is not the journalism itself but something layered on top. The sparkle emoji was adopted partly because readers already understood it and it was becoming an industry-wide standard. Yet cluttering pages with tiny explanatory text, one manager argued, becomes less and less helpful as AI use becomes intuitive to users. Transparency, in other words, is a design problem as much as an ethical one, and the optimal solution shifts as audiences grow accustomed to the technology.

The study also documents five cross-cutting challenges that persist despite the FT’s considerable resources. Achieving consistent, site-wide labelling remains elusive, with participants acknowledging the organisation still lacks a unified agreement on design components and transparency guidelines. Internal communication struggles against silos and fatigue: one manager estimated that no more than a couple hundred of the FT’s three and a half thousand staff had read the responsible AI framework. Standards themselves are moving targets, as models improve and public familiarity grows, prompting the question of when AI becomes just another editorial tool that no longer needs labelling, much as streaming platforms run recommendation algorithms without pop-up warnings. Human over-reliance on confident tools is an anticipated future risk. And audience misinterpretation is real: in usability testing, some participants assumed that the mere presence of an AI tool in an article meant the article itself had been machine-generated.

Beneath these practical difficulties, the researchers identified deeper tensions rooted in what institutional theorists call competing logics. Editorial participants framed transparency primarily in terms of reputation, accuracy, and accountability; product and analytics staff emphasised whether disclosure was intelligible at the point of use; legal and policy voices foregrounded provider requirements and liability exposure. The study interprets these dynamics through the lens of isomorphic pressure, the well-documented tendency of organisations to model themselves on peers under uncertainty. Regulation such as the EU AI Act exerts coercive pressure, professional norms and charters like the Paris Charter on AI and Journalism exert normative pressure, and the absence of any settled best practice drives mimetic copying of other leading publishers. Transparency at the FT is thus not owned by any single department; it is a site of ongoing cooperation and occasional conflict between professional, managerial, and commercial priorities.

The authors’ central argument is that AI transparency in news is best understood as a spectrum rather than a binary, a lived governance practice that evolves with technological capability, commercial reality, professional ethics, and shifting audience attitudes. Optimising for one factor, such as maximal disclosure demanded by regulators, can come at the expense of audience trust or revenue when transparency itself breeds distrust. The study offers several no-regret principles for other organisations: scale disclosure with a system’s autonomy and impact, pair AI labels with concise and comprehensible explanations, continuously measure how users actually interpret them, keep internal transparency and accountability high even when external disclosure is light, and favour consistent, reusable design components across products and platforms. The research carries limitations, being a single-organisation case study reliant on interviews and documents rather than ethnographic observation, with an insider researcher whose access was both asset and potential bias, mitigated by external ethical review and independent co-analysis. Still, its message resonates far beyond one newsroom: in the age of generative AI, telling audiences the truth about how the news is made is simple in principle, fiendishly complex in execution, and never a decision you make just once.

Subject of Research: AI transparency practices and disclosure decision-making in a major news organisation

Article Title: Tricky trade-offs on a transparency spectrum: how the Financial Times approaches transparency about AI use in news

Article References: Lohn, L., & Simon, F. M. (2026). Tricky trade-offs on a transparency spectrum: how the Financial Times approaches transparency about AI use in news. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03369-z

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03369-z

Keywords: artificial intelligence, journalism, transparency, Financial Times, generative AI, newsroom, trust in news, media ethics, disclosure, EU AI Act, audience trust, institutional logics