Verification Under Pressure

Journalists and fact-checkers spend a large part of each verification task on manual groundwork before any real investigation begins, and that load is a recurring driver of burnout. This is the central conclusion of 20 interviews with journalists, fact-checkers, researchers and editorial leaders in Europe, Africa, South Asia and the United States, conducted in June and July 2026 and read against published industry research.

Five findings recurred across roles and regions:

1. The groundwork is long and repetitive. Finding the original source of a claim, checking that source and looking for other coverage are repeated by hand for almost every story.
2. Volume, not complexity, is the main pressure. Claims arrive through more platforms, formats and languages than small teams can triage.
3. AI-generated images and video are hard to identify. Visual cues no longer hold, and detection tools often disagree with each other.
4. Burnout follows from doing more with fewer people. Interviewees linked it less to difficult investigations than to the steady queue of routine checks, in a sector that has cut staff for two decades.
5. Journalists will not hand the verdict to a machine. They carry the risk of the byline and want traceable evidence, not a score that decides for them.

The evidence points to a clear need: tools that shorten the groundwork while leaving the judgment with the journalist, with every source linked, signals instead of verdicts, and uncertainty stated plainly.

This is independent qualitative research into how verification work is done in newsrooms and fact-checking organizations, and where it breaks down. It was carried out by Octavian Andrei Cincă, an independent researcher based in Romania who works on editorial verification systems.

The research began without a product in mind. Its findings later became the basis of Veruno, a verification triage platform the author is building. We disclose this interest so readers can weigh the conclusions accordingly. The findings sections report what interviewees and published sources said; the product is described only in its own section near the end.

The report is based on 20 in-depth interviews conducted in June and July 2026, compared with published industry research.

Participants. Journalists, fact-checkers, researchers and editorial leaders, from reporters to the heads of teams and organizations. They work at international news agencies, national newsrooms, independent fact-checking organizations and research groups.

Regions. Europe, North and East Africa, South Asia and the United States.

Recruitment. Participants were identified through LinkedIn. The author searched for news, research and fact-checking organizations, then for the people responsible for the stages of verification this research examines. Those who accepted a connection request received a message explaining the research; those who agreed scheduled an online interview at a time that suited them.

Format. Online conversational interviews in English, typically about 25 minutes long and sometimes longer, built around six themes adapted to each participant's role:

1. Finding a story worth pursuing
2. Identifying and verifying the original source
3. Finding other outlets and sources covering the same claim
4. Detecting AI-generated images and video
5. Coping with the volume of information online
6. Workload, time pressure and burnout

Anonymity. Participants spoke on condition of anonymity. They are described only by role and region, and their organizations are not named, since a role and an outlet together could identify a person. No direct quotes are attributed.

Analysis. Notes from all 20 interviews were collected in a single file and compared theme by theme. A finding is reported only when it recurred across several participants and roles. Each finding is then set against published research, listed under References.

Saturation. An initial round of three interviews, with people working in fact-checking, investigative journalism and open-source research, produced a first synthesis. Its two central findings, the problem of volume and distrust of automated verdicts, were tested in the remaining interviews and held across roles and regions. As the interviews continued, answers converged: participants in different countries described the same workflow, the same pressures and the same expectations of tools. In qualitative research this is known as thematic saturation, the point at which further interviews stop producing new themes.

The verification workload described here falls on an industry that has lost staff and revenue for twenty years. Several interviewees raised layoffs without being asked.

The causes came in waves. In the 2000s, classified and print advertising moved online. In the 2010s, digital advertising concentrated on a few large platforms. Throughout, consolidation by large chains reduced local staff.

The latest pressure is generative AI. AI answers in search reduce traffic to news sites: Google search referrals to more than 2,500 news sites tracked by Chartbeat fell 33% in the year to November 2025, and media leaders expect search referrals to fall by a further 43% over the next three years (Newman, 2026) [7]. Some publishers now cite AI directly. Business Insider cut 21% of its staff in 2025, pointing to "extreme traffic drops" from AI-driven search (Nieman Lab) [8].

Most of these figures come from the United States, where tracking is most complete. Interviewees described similar pressures elsewhere.

Verification teams now face more claims, more formats and a more doubtful public, with fewer staff and less money.

News increasingly reaches people through platforms, creators and AI chatbots, which audiences trust less than news overall. Concern about what is real or fake online rose in many markets, by as much as nine points in Belgium. Fact-checkers are reaching larger audiences with smaller teams, so each verifier carries more of the work.

Interviewees described a first phase of verification that repeats with almost every story and can take hours before in-depth work begins. Across roles and regions, the same four steps came up:

1. Finding a topic worth pursuing, by scanning feeds, wires, social platforms and messaging apps.
2. Identifying the original source of a claim, often several reposts away from where it started.
3. Checking that source: who runs it, how long it has existed, whether it has a track record.
4. Finding other coverage of the same claim, in other outlets and languages.

Checking for earlier fact-checks and documenting the trail are also standard practice, but interviewees did not single them out.

None of these steps is difficult on its own. The problem interviewees described is that they are done by hand, tab by tab and search by search, and repeated for each new claim. That time is taken from the investigation itself.

Published research describes the same pattern. In a 2025 study of eight experienced investigative journalists in Germany, participants relied on automation for repetitive work such as monitoring, scraping and summarizing. One described geolocating a single house in a photo as taking hours; another said journalists lose significant time copying data between files. The same study stresses that human judgment is still needed to validate any automated result (Cifliku and Heuer, 2025) [3].

Interviewees consistently named the amount of information online as one of the defining pressures of the job. Claims arrive from more places, in more formats and languages, and faster than a small team can triage.

The difficulty is fragmentation as much as volume. A single claim may appear on a video platform, in a messaging group and on a niche website at once, each copy slightly different. Deciding which claims deserve attention is itself a time-consuming task, before any verification starts.

Interviewees were clear that the problem is volume, not complexity. They described themselves as skilled and rigorous, but working with tools that have not kept pace with bot networks, AI-generated media and coordinated campaigns. During breaking news, hundreds of videos, images and posts can arrive at once, and each has to be opened, reverse-searched and cross-checked separately. The bottleneck is rarely the verification decision. It is deciding what to look at first.

Audience data reflect the same shift. In 2026, social media and video networks became the most used source of news worldwide for the first time, at 54% (Reuters Institute, Digital News Report 2026) [4]. Fact-checkers are adapting: 60.6% now publish in more than one language, up from 52.5% a year earlier (IFCN, via Poynter) [5]. Each new platform and language widens the ground a verifier has to cover.

Interviewees described AI-generated images and video as one of the newest and most difficult parts of the job. The visual cues they once relied on no longer hold, and a single convincing image can reach a large audience before anyone checks it.

Detection tools help, but they do not settle the question. A NewsGuard audit published in May 2026 ran 45 images through five leading AI detectors: 15 authentic news photos, 15 lightly edited with AI and 15 significantly edited with AI to change their meaning (Blachez et al., 2026) [2]. It found that:

• Taken together, the tools labelled authentic photos as AI-generated 13.33% of the time; the worst tool did so for 6 of the 15 real images.
• The tools least likely to flag real photos wrongly were also the worst at catching significantly edited ones.
• On 35 of the 45 images, at least one tool disagreed with the others.

Earlier testing by WITNESS, published by the Reuters Institute, found that cropping, compression or screenshots can lead detectors to classify AI content as real (anlen and Vázquez Llorente, 2024) [1]. Platforms routinely compress and strip files on upload, so this is the everyday case rather than the exception.

The risk runs both ways. A missed fake spreads; a real photo wrongly flagged as fake can be used to dismiss authentic evidence. Both sources conclude that a detector's result should be one signal among several, alongside reverse image search, provenance and context.

Burnout came up in every role, from reporters to editorial leaders. Interviewees linked it less to the hardest investigations than to the steady load of routine checks, the constant flow of new claims and the sense that the queue never empties.

Several placed this in the context of layoffs. They described teams cut for economic reasons while the number of claims kept rising, so the same standards had to be met by fewer people (see Background).

Large surveys point the same way. In an industry-wide survey of US journalists conducted in 2023 by the Reynolds Journalism Institute and SmithGeiger, 80% reported burnout and 91% had seen colleagues struggle with it (Raškauskaitė, 2025) [10]. Respondents named being forced to do more with less as the biggest negative factor, and 20% of active journalists were planning to leave the profession, or had already left, because of burnout.

For fact-checkers, shrinking teams add to the pressure: the share of organizations cutting staff rose from 14.9% to 38.3% in a single year, while their audiences grew (IFCN, via Poynter) [5].

Tools alone will not solve burnout. But every hour of manual groundwork removed is an hour returned to judgment, reporting or rest.

Interviewees were consistent: they want help with the groundwork, not a tool that decides what is true. A journalist's name is on the byline, and the journalist carries the professional and reputational risk of publishing something false. A tool that tries to replace that judgment would be rejected, however accurate it claims to be. What they asked for instead was traceability: every signal tied to its source, so they can check the tool's own work.

Published evidence supports their caution:

• Automated verdicts are unreliable. AI detectors disagree with each other and fail in both directions (Finding 3). A tool that produced a single truth score would inherit those errors and hide them.
• Human validation remains necessary. Research on newsroom automation stresses that journalists must still check and interpret any automated result (Cifliku and Heuer, 2025) [3].
• Trust in AI answers is low. Only 20% of people worldwide trust news answers from AI chatbots, against 37% for news overall (Reuters Institute, Digital News Report 2026) [4].

Adoption is already under way: more than half of fact-checking organizations used AI tools in 2025 (IFCN, via Poynter) [5]. The question is no longer whether verifiers use automation, but whether it shows its sources and leaves the decision to them.

Four design principles follow from this: show every source with a link; present signals as leads to investigate, not verdicts; state uncertainty plainly; and leave the conclusion with the journalist.

Veruno was designed from these findings: a triage platform that gives verifiers a faster, documented starting point for investigating a claim, not a verdict.

After the interviews ended in July 2026, the notes were brought together, compared and turned into a first synthesis of the problems and of how each could be made faster and simpler to handle. Development ran from July to the end of October 2026, starting as the last interviews were completed. During that period the design was adjusted using the interview findings and feedback from practitioners and public-sector contacts in the author's professional network. The platform is now ready for its first pilot phase, with a date still to be set.

Veruno differs from enterprise platforms that detect coordinated campaigns across whole networks. Those tools are built to map influence operations at scale. Veruno is built for claim-by-claim triage: by newsrooms and fact-checkers of any size, and by public institutions, such as media regulators, that need to check specific claims, for example during an election. Its signals are transparent, and its pricing is set for small and mid-sized teams. After two decades of cuts, the most valuable thing a tool can return to a newsroom is working hours, and the most costly thing it can help prevent is a published error.

This is qualitative research: it identifies patterns rather than measuring them.

• Sample. Twenty interviews cannot represent the whole industry. Latin America, West and Southern Africa and East Asia were not covered.
• Self-report. Time spent on tasks is based on interviewees' accounts, not on observation or time tracking.
• Author interest. The author is building a product based on these findings. To limit bias, each finding is checked against independent published research.

Next steps:

1. Pilot the platform with regulatory and journalism institutions in Romania before public launch, and incorporate their feedback.
2. Measure time saved during the pilot by timing the same tasks with and without the platform.
3. Extend the interviews to regions not yet covered.
4. Repeat the study after a year of use to see whether workload and burnout indicators change.