A six-person fact-checking team at India Today spends its days sorting through false and misleading images: AI-generated archival photos of politicians, flood footage taken out of context and deepfakes of prominent figures.

Now the team is testing Backstory, an experimental tool from Google DeepMind that brings several parts of that verification process into one place, according to Nieman Lab’s Andrew Deck.

A reporter can upload an image and ask Backstory a question. The tool then decides which checks to run. It can make a SynthID tool call to look for watermarks embedded in images created with Nano Banana, Google’s image-generation model. It can also check content credentials tied to the Coalition for Content Provenance and Authenticity standard and trace where an image has appeared online.

The result is an AI-generated report with citations and a log of the steps Backstory took.

The potential time savings are a major attraction. Mike Caulfield, the digital literacy researcher who created the SIFT method, told Nieman Lab that Backstory can shrink one initial verification process that might take 50 minutes to about three.

Bal Krishna, who leads India Today’s fact-checking team, described the benefit another way: “It is doing all the work you might have done with five different tools, five different logins, in the same place.”

Backstory is free for now but available only through Google’s Trusted Testers Program. DeepMind has given access to thousands of journalists, open-source intelligence researchers, librarians and researchers.

That newsroom input has not always been part of the development process for authentication tools. A July report from Princeton University’s Center for Information Technology Policy found that many such tools are built without enough consultation with journalists. Some produce confidence scores about whether an image is synthetic while offering users little clarity about how those scores were reached or how they should be explained.

Backstory also makes mistakes.

When Deck uploaded his headshot, the tool confused it with a photo of another journalist and generated a report about the wrong image. Caulfield described the mistake as “conflation,” a problem that can occur when large language models encounter visually similar images.

There are other limits. Backstory currently works only with still images, and its context-tracing feature can find only material indexed by search engines, leaving gaps on platforms such as TikTok and closed messaging services.

India Today treats the tool as a starting point rather than a source. Krishna said reporters manually confirm information they get from Backstory, and the organization does not cite its AI-generated reports directly in published fact-checks.

The tool also comes from a company producing the kinds of synthetic images fact-checkers increasingly encounter. Google added Nano Banana image generation to Google Earth in July, then said it would roll back the feature while developing stronger guardrails after users shared misleading generated satellite-style images.

Zoe Darmé, a DeepMind product manager, told Nieman Lab that Backstory was not developed as a direct response to criticism of Google’s image tools.

For Krishna, the appeal is simpler. Asked whether Backstory has sped up his team’s daily work, he said: “There is no doubt about it.”

The post Google’s Backstory aims to speed up image verification for fact-checkers appeared first on The Media Copilot.

Anthropic is offering new details about how it plans to watermark text written by Claude — including what happens when someone edits the output.

The company uses a simple example to explain the technology. If Claude can choose between words such as “overcast” and “grey” to describe the weather, it can make those small choices in a pattern. A reader wouldn’t notice. But someone with the right detection tool could identify the pattern as a watermark.

Anthropic described the system in a Friday TechCrunch blog post answering questions about the plan, which it announced earlier in the week to comply with the European Union’s AI transparency rules. We covered the original announcement here.

The company plans to use SynthID Text, a watermarking technology developed by Google DeepMind. Anthropic also says it will release an API to detect the watermark.

The system has limits, particularly once people start changing Claude’s words.

“Light editing probably won’t remove the watermark completely,” Anthropic said. But a complete rewrite in which every word is replaced would remove it. Anthropic said in such a case, it becomes harder to describe the finished product as AI-generated.

Whether a watermark can be detected also depends on how much Claude actually wrote. If someone gives Claude a human-written document and asks it to make only minor edits, Anthropic said that leaves “very little (if anything) for the watermark to attach to.”

Code presents a similar problem. Code gives a model fewer stylistic choices than prose. Anthropic said it can still watermark choices, including some comments and names within code, but the technology should have “a negligible effect on the actual code produced.”

The company is also drawing a distinction between watermarking and statistical AI detectors.

Those tools look for patterns associated with AI writing and estimate whether a model produced a piece of text. A watermark works differently, using a pattern deliberately inserted when the text was generated.

An AI detector makes a statistical judgment. A watermark provides evidence that a particular system marked the text when it was created.

Not everyone is happy about the change. TechCrunch reported that dozens of Claude users on X said they had canceled subscriptions over the watermarking plan, citing Business Insider.

For news organizations, the technology could make questions about AI use easier to investigate.

An editor could check whether copy submitted by a freelancer carries Claude’s watermark. But competitors, critics and others with access to Anthropic’s detection system could check published work as well.

That makes newsroom policies around AI use more consequential. Publishers may need clear rules about when it is permitted, how much AI-generated material can appear in published work and when that use should be disclosed.

Claude won’t necessarily be the only model leaving marks behind. Anthropic says other major AI developers that signed the same European code of practice are also working on watermarking systems.

What remains unclear is whether those systems will work together. Publishers find themselves checking text against a different system for every AI company.

The post Anthropic says Claude’s text watermark survives light editing but not rewrites appeared first on The Media Copilot.

A PR agency representative offered guaranteed placement on a website impersonating Forbes for $10,000, according to a recorded sales call obtained by Press Gazette. Forbes later confirmed that forbesla.com has no connection to the company and said it is taking action against it and another site posing as a Forbes publication.

The two fake Forbes sites publish dozens of stories a day under bylines with little or no apparent presence elsewhere online, Press Gazette first reported. Reporters Nathan Pine and Liam Redmond appear to have no social media profiles. AI detection tool Pangram classified one of Redmond’s articles as entirely AI-generated.

Press Gazette also reported that Identifai flagged the byline photo of Harriet Caldwell, described by Forbes Liechtenstein as an experienced news editor, as AI-generated. Her articles were also flagged as likely to have been generated by AI.

The sales pitch extended beyond the fake Forbes sites. Baden Bower, a public relations agency that promises clients guaranteed media coverage, was the company whose representatives offered the placements documented by Press Gazette. One representative, Zac, offered placement in the Los Angeles Times, said The Guardian and The New York Times were the only publications where he could not secure coverage, and quoted $9,900 for a Bloomberg TV interview with an Emmy-winning journalist. He also suggested paid articles could help applicants seeking U.S. skilled worker visas.

A second Baden Bower representative, Camilo, said he was unaware of Forbes LA and that the agency works only with licensed regional Forbes editions. He quoted $25,000 for Forbes US, $2,000 for Forbes Georgia and $100,000 for The Wall Street Journal.

Asked whether paid placements are labeled as promotional, Camilo said, “Some are, some are not.” He also claimed Baden Bower has commercial relationships with more than 1,980 publications, including Press Gazette. Press Gazette said it has never published a sponsored article from the company.

Much of the content on the fake sites promotes companies, including Baden Bower itself. One Forbes Liechtenstein article, which Pangram classified as entirely AI-generated, projects that the agency will reach 7,600 clients and 90,000 published stories by the end of 2026 and quotes company leader AJ Ignacio touting its commercial results.

Baden Bower then cites that article on its own website as evidence of its standing, alongside pieces from an AI-generated Philippines edition of Fast Company and an apparent fake German edition of Vanity Fair.

The sites are not new. The Internet Archive shows Forbes LA online since January 2025 and Forbes Liechtenstein since March 2025. The London office listed by Baden Bower has a single contact number, a mobile number that appears to be disconnected.

For publishers, the risk is to the masthead itself. Fake sites can cheaply trade on the authority of established news brands, turning their names and reputations into promotional inventory. The model echoes fake AI news networks built to game search results and fabricated AI expert profiles reaching real reporters, but extends the problem into the PR industry.

Sarah Waddington, CEO of the Public Relations and Communications Association, said the practice violates industry standards on transparency, honesty and earned credibility. Former BBC journalist Omar Hamdi, who first noticed founders sharing fake Forbes coverage on LinkedIn, said the promotional tone of the articles raised his suspicions.

For newsrooms, the simplest defense may be basic verification: check that the publication behind a citation actually exists before trusting it.

The post Fake Forbes sites sold guaranteed coverage for $10,000, PR pitch shows appeared first on The Media Copilot.

James Allen of Billpin was quoted as a finance expert by outlets including CBS News, the New York Post and Business Insider. Qwoted later barred his account after flagging repeated AI-generated submissions and finding that he had failed its verification process, according to Press Gazette’s investigation.

The case is a stark warning about the source pipeline behind everyday newsroom copy. Qwoted, a U.S. platform that connects journalists with sources, says it has banned hundreds of agencies, executives and experts for conduct ranging from fabricated identities to AI-generated pitches and false claims of expertise.

Billpin’s website is now blank, and Allen’s LinkedIn profile has disappeared, Press Gazette reported. The publication said Allen did not respond to requests to verify his identity. Press Gazette concluded the operation appeared designed to win media mentions that could boost search rankings, or backlinks.

A quote in a respected publication can generate a link that helps an otherwise obscure website gain visibility in search results. That creates an incentive to manufacture expert identities that appear credible enough to earn media coverage.

Qwoted CEO Dan Simon said generative AI has dramatically accelerated an already difficult problem. “AI has just poured petrol on the disinformation fire,” he told Press Gazette.

The platform says it now handles more than 15 violations a day and blocks about 50 prospective clients from registering each day. Shelby Bridges, Qwoted’s head of user experience, said the volume of low-quality pitches began rising rapidly after ChatGPT’s public release in November 2022.

Simon described a source ecosystem that increasingly blurs the line between legitimate expertise and fabrication. “You’ve got real agencies representing fake experts with fake expertise. You’ve got real agencies with fake experts and real agencies with real experts with fake expertise, and fake agencies with fake experts with fake expertise.”

One recurring pattern is a purported expert responding to queries on subjects with no obvious connection to their background. Press Gazette identified examples of sources offering commentary on finance, cleaning, human resources, pet sitting and mental health.

Qwoted combines human review with signals such as unusually high pitching volumes. It also uses Pangram, an AI-content detection service, to flag potential abuse. But AI detection is only one signal. A model score cannot establish that a person is fictional, that a quote is false or that an expert lacks relevant experience. Qwoted’s reported checks therefore matter more than a detector’s output alone.

Simon said some reporters have told him they do not object to real experts using ChatGPT to polish their responses. The bigger risk comes when AI helps conceal a fabricated identity or invented credentials, allowing a bogus source to slip into print before anyone questions its legitimacy. Platforms can suspend accounts after the fact, but the final responsibility still rests with newsrooms.

The post Qwoted bans AI-generated expert profiles as PR spam floods reporters’ inboxes appeared first on The Media Copilot.

Substack readers can now run a check on any post longer than 100 words to see how much of it an AI model likely wrote. The company rolled out the feature on July 21 in partnership with the detection firm Pangram, with the scan available on web and iOS and Android support promised later.

The tool works only on content published starting on July 21. Alongside it, Substack is giving writers a dedicated space to state whether and how they used AI in a given piece. The idea, per the company, is to set reader expectations rather than police anyone.

No AI detector is reliable enough to call a piece machine-written with certainty. Pangram uses a deep learning classifier trained on around a million documents rather than the older perplexity-and-burstiness heuristics, and it reports strong internal accuracy numbers. But The Atlantic examined how well Pangram actually performs and found the picture more complicated than the marketing suggests. False positives are a known problem across the category, and lightly edited or paraphrased AI text is far harder to catch than raw model output.

Substack did acknowledge those limits in its announcement post. It also hinted at further features tied to AI content and reader preferences without saying what they are.

Its acknowlegdment matters because a wrong flag carries real cost. A human writer who gets tagged as an AI user has little recourse in the moment, and the accusation sticks. Non-native English writers have historically drawn more false positives from detection systems, which raises fairness questions for a platform with a global base of newsletter writers. Pairing the automated score with a voluntary disclosure field is Substack’s hedge against that, letting writers put their own account on the record next to the machine’s.

For newsrooms and independent publishers, the move is a preview of how disclosure norms may harden across platforms. If readers start expecting an AI provenance label on every post, editors will face pressure to document their own use of the technology and to decide where the line sits between an AI writing assistant and AI ghostwriting. Publishers who have built AI into their workflows, a shift The Media Copilot has tracked closely, will want a clear internal policy before a third-party score answers the question for them.

The real test isn’t whether the detector works. It’s whether readers trust a score that Substack itself acknowledges can be wrong. The company is betting users will treat AI detection as one signal among many, not a definitive verdict.

The post Substack adds AI detection to its posts through a new Pangram partnership appeared first on The Media Copilot.

Media consumption recently passed a big milestone: people now turn to social media and video networks like YouTube for news more than any other source. The Reuters Institute’s Digital News Report, now in its 15th year, found 54% of audiences now rely on social and video platforms to get their news, putting them ahead of publisher websites at 51% and TV at 52%.

And the trust side of the ledger keeps getting worse. Just 37% of people say they trust most news most of the time, the lowest reading since Reuters started tracking it in 2015. In the United States the figure sinks to 25%. Gallup’s October 2025 poll landed in the same place, with U.S. trust in mass media at 28%, down from 31% the year before and 40% five years ago.

The natural read is that media brands matter less every year, drifting toward irrelevance as audiences scatter into feeds. AI chatbots seem to accelerate the slide. The Reuters report puts news consumption via AI chatbots at 10%, up from 7% a year ago. If brand erosion plus AI summarization is the trajectory, the long-term picture suggests publishers will eventually be reduced to information wholesalers, supplying the raw facts and quotes that someone else interprets, packages, and presents back to the reader.

That story has been the dominant framing for about two years now. But the data underneath doesn’t actually back it.

What the click pattern tells you

Most news consumption on social platforms is incidental. Posts and clips arrive between the workout tips and the gadget ads, and the Reuters report identifies a growing slice—now 12% of people and double the 2020 figure—who run into news only while they’re online for something else. That’s not really audience; it’s adjacency.

Behavior inside AI products reads very differently. Among people who click out of an AI answer, 44% do it to verify the news is correct, against 36% on search and 33% on social. Another 43% click to find out more about the source, versus 35% and 34%. Only 51% click for more detail, well below 59% on search and 60% on social.

That’s a behavioral signal worth paying attention to. Inside an interface designed to strip out bylines and erase visual brand cues, audiences are reaching back through the answer to get to the publisher who supplied it. The dominant reason isn’t curiosity. It’s verification. Readers don’t fully trust the summary, so they reach for the name they recognize to check it.

That breaks the simple “trust in news is collapsing” story. The aggregate trend is real, we don’t live in the aggregate. People can hold low trust in “the media” while continuing to rely on the specific publications they’ve read for years. The Reuters data confirms it: Overall trust fell in 29 of the 48 markets surveyed, but trust in the most widely used individual brands held its ground, with several major names sitting above the broader decline. Behavior and stated preference point at the same answer. Audiences are funneling toward names they already know.

The brand still matters. Arguably more than at any point in the last decade, because the brand is the only fixed object as the surrounding interface keeps changing.

Trust converts but not on impact

We should be realistic about the size of the audience that gets news via AI—it’s still only 10%, and just 1% call AI their main news source. But the slice is growing faster than any other channel, and it skews toward the most engaged readers. Among the biggest news lovers, 18% already use AI for news. That is the cohort every publication has been trying to win for the last decade.

The catch is that trust is not directly convertible. A reader who treats your name as a stamp of credibility inside a chatbot summary may never click. A reader who does click to verify a fact on your site likely arrives, scans, and bounces. Brand reliance at the moment of consumption often produces no measurable lift.

The conversion, however, can happen somewhere else. The reader who keeps reaching for your name to check the machine is the reader who eventually subscribes, who shares your work to a contact, who recommends the publication when a friend asks where they get their information. Reuters found that 46% of paying news consumers now cite values-based reasons for paying, rather than the specific content they’re buying. Those reasons accrue. The brand-reliance behavior happening inside AI interfaces is the leading indicator of the durable reader relationship that eventually shows up in revenue.

The practitioner work for the next 18 months is operational. To make the most of AI audiences, publishers need to build instrumentation that captures the moments when readers reach for the brand, even when the click numbers look thin. Build persuasion strategy that converts those signals into something countable.

Stop playing defense

The headline finding of the Reuters report implies a strategy for media: get on more surfaces, get on them harder, push more short-form video, and lean into the platforms that audiences actually use. Most publishers are following that script. On AI, the script has been the opposite, with many media sites blocking crawlers completely.

All of that is defense. While defense is important, if it’s your entire strategy, you will lose. The offensive posture is to fight to be the default name in your lane, the publication readers reach for when they doubt whatever the surface is showing them.

That means using social, but treating it as funnel rather than destination. Casual readers get a taste; the strategy is to convert a fraction of them into a brand relationship that survives outside the platform. It means blocking crawlers that take without permission, but pairing the block with clean, machine-readable paths for partners and licensees. It means producing the clip, but anchoring the clip to deep, comprehensive coverage that earns the reader’s return visit and, eventually, their subscription.

The creator economy points the same direction. About 27% of people now get news from creators who explicitly focus on news, and 46% from creators of any kind. Those creators score better than legacy outlets on relatability and entertainment value. They also rate lower on trust and impartiality. And the audience that watches them consumes more traditional media than the average reader, not less. Only 3% rely on creators alone. Creators introduce audiences to topics. The brands pick up the verification.

The fragmentation story is real. Audiences are scattering across more surfaces, taking news in smaller pieces, and getting more of it from formats that didn’t exist a decade ago. But the behavior underneath that fragmentation runs the other way. The more the news gets sliced up, the harder readers lean on a name they trust to tell them what’s actually true. Audiences take their news in smaller bites now, but the chef still matters.

A version of this column appears in Fast Company.

The post The news brand is the only thing AI users still click for appeared first on The Media Copilot.

Within hours of Nicolás Maduro’s arrest by U.S. forces, fake AI-generated images of Venezuela’s ousted president spread across social media faster than newsrooms could verify them.

The incident marks one of the first times synthetic imagery has depicted a major figure during a rapidly unfolding news event, according to a New York Times report by Stuart A. Thompson and Tiffany Hsu.

“This was the first time I’d personally seen so many A.I.-generated images of what was supposed to be a real moment in time,” Roberta Braga, executive director of the Digital Democracy Institute of the Americas, told the Times.

Some fake images made it into Latin American news outlets before being quietly replaced with an official photo shared by President Trump. NewsGuard, which monitors the reliability of online information, tracked five fabricated images and two misrepresented videos that collectively drew more than 14.1 million views on X in under two days.

The Times tested a dozen AI generators and found most tools, including Google’s Gemini, OpenAI’s ChatGPT, and X’s Grok, quickly created fake arrest images despite stated policies against misleading content.

Jeanfreddy Gutiérrez, who runs a fact-checking operation covering Venezuela, said the fakes spread through “almost every Facebook and WhatsApp contact” he has before official images were available.

“It just took a lot of work, because we always lose the battle to convince people of the truth,” Gutiérrez told the Times.

A pattern of failures

That battle is getting harder. The Maduro deepfakes emerged just days after Grok became the center of a global regulatory firestorm. Governments in the EU, UK, France, India, Malaysia, and Australia have all launched investigations after the chatbot began generating non-consensual sexualized images of women and minors at scale.

Bloomberg reported that Grok was generating thousands of “undressed” images per hour earlier this week. The official Grok account posted an apology on X, writing that it “deeply regret[s]” generating sexualized images of girls “estimated ages 12-16.”

If a major image generator can’t prevent the creation of child sexual abuse material, its safeguards against political deepfakes are likely just as porous.

Gutiérrez said many people refused to believe the official image of Maduro posted by Trump was real.

“It’s funny, but very common,” he told the Times. “Doubt the truth and believe the lie.”

The post Journalists scramble as AI-generated images beat official photos to publication appeared first on The Media Copilot.

The Associated Press rolled out AP Verify on Monday, packaging AI-powered verification features alongside traditional authentication tools in a single web-based dashboard.

The platform combines reverse image search, frame-by-frame video analysis, and social listening with AI-driven geolocation, object detection, transcription, and a chatbot assistant. It also includes generative AI text detection and team collaboration features.

“In an era of rampant misinformation and digitally altered content, verification is more essential than ever,” said Gianluca D’Aniello, AP’s senior vice president and chief technology officer.

“AP Verify equips journalists with the essential tools they need to assess online content quickly and accurately all in one place – whether it’s identifying the source of a photo, analyzing video or vetting text.”

AP has used the tool internally for a year before offering it to other publishers. The newsroom used it to secure original Texas flood footage by tracing it to its source, verify a viral meteor sighting in South Carolina, and find eyewitness video from the Charlie Kirk assassination, according to Aimee Rinehart, senior product manager for AI strategy at AP.

Before AP Verify, journalists relied on a “patchwork of tools” including Google reverse image search, Rinehart told Press Gazette. The platform integrates third-party providers including Google’s Fact Check, Trint for transcription, Graylark’s geolocation, GPTZero’s AI text detection, Trendolizer, and identity solutions provider Pipl.

“None of the tools are 100 percent,” Rinehart said. “We would never recommend you go straight to publish just based on that tool’s information.”

The platform surfaced a need among under-resourced newsrooms. During market research, a local broadcaster told AP they sometimes run the wrong tornado video and apologize the next day.

Each newsroom subscription keeps content private. Competitors cannot see what others are verifying.

The launch will test whether publishers want centralized verification tools, Rinehart said. She pointed to last year’s Kate Middleton photo manipulation incident, when five agencies including AP pulled a palace-issued image, as evidence that “AP became a source of truth” for publishers.

“That’s what we really want to get to,” Rinehart said, “is everybody trying to discern what is real, what has been retouched, and what can we trust online?

The post AP launches verification dashboard combining AI and traditional tools appeared first on The Media Copilot.