Publishers Adopt AI Tools as Google AI Overviews Siphon News Traffic

AI + Journalism: Collaboration or Collision? — Ai4

Ai4

Newsrooms face an existential revenue threat from zero-click search engines, forcing publishers to monetize archives through tech licensing deals while deploying generative models to scale investigative reporting under strict human oversight.

AI-assisted summary of AI + Journalism: Collaboration or Collision? — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 27, 2025, this presentation features veteran technology reporter Nancy Scola exploring the intersection of artificial intelligence and the news industry. Drawing on her extensive reporting background, Scola analyzes whether generative technology represents an existential collision or a productive collaboration for working journalists.

## TL;DR Scola frames AI's arrival against an already devastated American newspaper landscape that has lost 80% of its employment over recent decades. She outlines tangible collision points—including low-quality 'AI slop,' hallucinated financial explainers, flawed algorithmic bias tools, and traffic-depleting zero-click search overviews. However, she demonstrates how journalists can successfully collaborate with generative tools to overcome idea block, achieve rapid domain literacy on deadline, audit massive document troves, and streamline multi-format publishing. Ultimately, using case studies like The New York Times, she argues that newsrooms must secure fair commercial compensation while ensuring core editorial responsibility remains strictly human.

## Atmosphere & style Delivered with the clear-eyed pragmatism of a seasoned beat reporter, the talk sidesteps sensational doom-and-gloom in favor of candid, newsroom-tested reality. Scola’s delivery is conversational, analytical, and fast-paced, balancing sobering statistics on local news deserts with humorous anecdotes about AI blunders and genuine excitement for workflow efficiencies. The presentation feels deeply grounded and constructive, transitioning from an urgent examination of institutional vulnerability into an empowering, tactical guide for modern media professionals.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=4s) AI enters journalism during a severe structural crisis, with U.S. newspaper employment dropping 80% between 1990 and 2025, over 3,000 papers closing, and 200 counties becoming news deserts. - [02:18](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=138s) Cost-cutting attempts to deploy AI for automated reporting have resulted in low-quality 'AI slop' and damaging factual errors, such as CNET's botched financial explainers and Sports Illustrated's fake bylines. - [02:18](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=138s) Automated attempts to measure or balance bias can backfire catastrophically, exemplified by the Los Angeles Times 'Insights' tool whitewashing the Ku Klux Klan. - [04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) Zero-click AI search summaries, like Google AI Overviews, directly threaten digital journalism by siphoning away website traffic, ad revenue, and subscriptions. - [04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) Publisher licensing deals with AI companies have created internal friction, pitting corporate profitability against journalists who feel their work is being exploited without traffic reciprocation. - [04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) Audiences consistently demonstrate lower trust in journalism when they suspect artificial intelligence was involved in its creation. - [06:45](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=405s) Generative search tools help beat reporters overcome creative burnout and rapidly gain domain expertise on breaking news deadlines. - [08:40](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=520s) Machine learning and computer vision enable small investigative teams to analyze massive datasets, from reviewing 600 complex court dockets to identifying bomb craters in satellite imagery. - [08:40](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=520s) Civic AI tools like Michigan's 'Minutes' help small local newsrooms monitor hundreds of public municipal and school board meetings via automated transcription and keyword alerts. - [11:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=676s) Tools like NotebookLM serve as private 'external brains' for reporters, grounding answers strictly in verified uploaded documents to eliminate hallucinations. - [11:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=676s) Retrieval-Augmented Generation (RAG) over verified newsroom archives can automate the drafting of repetitive background boilerplate in breaking news coverage. - [14:02](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=842s) AI assists with peripheral publishing tasks, including generating headline options, tailoring social media copy for specific demographics, and curating automated topical newsletters. - [15:40](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=940s) Machine translation and synthetic voice narration enable news organizations to cost-effectively expand into multilingual global markets and improve article accessibility. - [16:41](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1001s) Major news organizations are adopting a dual commercial strategy by suing uncompensated data scrapers while entering paid licensing deals with tech companies. - [16:41](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1001s) Leading newsrooms maintain that while AI can augment research, translation, and distribution, ultimate editorial responsibility and truth-telling must remain with human journalists.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=4s) Introduction and the Baseline: An Industry Already on the Brink

Nancy Scola, a veteran Washington, D.C.-based technology reporter, opens her talk by addressing the core question of whether artificial intelligence is destined to collide with journalism or collaborate with it. A contributing writer for *Politico Magazine* and *Washingtonian*, and a former staff writer at *Politico* and *The Washington Post*, Scola regularly covers tech policy, antitrust, and AI—including profiling figures like Bruce Reed, the Biden administration's White House point person on AI. Setting aside the conventional "doom and gloom" tone, she previews a two-part agenda: first detailing the tangible collisions threatening the press, followed by the practical collaborations and upside potential that working reporters are already discovering.

Before dissecting artificial intelligence directly, Scola establishes why journalists are on edge: the American news industry has already suffered a historic collapse. Between 1990 and 2025, newspaper industry employment plunged by 80%. In the last two decades alone, more than 3,000 newspapers across the United States have shuttered—a casualty rate exceeding two publications every single week. This gutting of local media has left roughly 200 U.S. counties without any local news outlet whatsoever. AI enters the frame not into a healthy ecosystem, but as an existential shock to an industry already fighting for survival, capturing an anxiety neatly summarized by a *Guardian* headline: "We need to set the terms or we'll all be screwed."

- **Newspaper employment drop:** 80% decline in U.S. newspaper industry jobs between 1990 and 2025. - **Paper closures:** Over 3,000 U.S. newspapers closed in the past 20 years (averaging more than two per week). - **News deserts:** Approximately 200 counties in the United States currently operate with zero local news outlets.

---

[02:18](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=138s) The Collisions: AI Slop, Hallucinated Explainers, and Algorithmic "Neutrality"

The first collision point Scola identifies is the proliferation of low-quality automated content—often referred to as "AI slop." Publications attempting to cut labor costs have flooded the market with machine-generated articles passed off as genuine reporting. A notorious example occurred at *Sports Illustrated*, which published equipment reviews attributed to a fictional, AI-generated staff writer named "Drew Ortiz" evaluating volleyballs—copy that Scola notes was decidedly not Pulitzer Prize–winning material. The danger extends beyond poor prose to outright factual inaccuracy. At the struggling tech outlet *CNET*, management turned to AI to produce basic financial explainers to stay afloat, only to publish a widely ridiculed guide that fundamentally miscalculated and misexplained compound interest.

A parallel threat emerges from media executives attempting to replace human editorial judgment with supposedly unbiased bots. Facing widespread public distrust, some outlet owners have turned to algorithmic scoring to prove their reporting is impartial. At the *Los Angeles Times*, owner Patrick Soon-Shiong floated the concept of attaching an automated "bias meter" to news articles to label whether a piece leaned left or right. Following pushback from newsroom staff, the initiative was dialed back to opinion pieces through a tool called "Insights," designed to offer contrary viewpoints. That experiment imploded when the tool analyzed a column discussing the Ku Klux Klan and generated a rebuttal asserting that the Klan was not necessarily an organization founded on racial hatred, but merely an expression of "white Protestant culture." The tool was swiftly taken down.

```mermaid flowchart TD accTitle: AI and Journalism Collision Modes accDescription: Diagram showing the three primary collision points between AI implementation and journalistic integrity: AI slop, publisher appeasement, and automated bias tools.

subgraph ContentDegradation["Content Degradation"] A["Low-Cost AI Generation"] --> B["'AI Slop' & Pseudonyms<br/>(e.g., Sports Illustrated)"] A --> C["Factual Hallucinations<br/>(e.g., CNET Compound Interest)"] end

subgraph AlgorithmicJudgment["Replacing Judgment"] D["Automated Bias Meters"] --> E["'Insights' Counter-arguments"] E --> F["Whitewashing Extremism<br/>(e.g., LA Times KKK incident)"] end

subgraph EconomicExtraction["Audience & Revenue Loss"] G["Zero-Click Search Engines"] --> H["Google AI Overviews"] H --> I["Traffic & Ad Revenue Evaporation"] end ```

---

[04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) Zero-Click Search, Publisher Appeasement, and the Loss of Human Trust

Beyond editorial misfires, AI platforms are actively siphoning away the audience that sustains digital journalism. Google’s rollout of AI Overviews exemplifies the threat of "zero-click search": rather than directing users to primary source links, the search engine synthesizes an answer directly on the results page. Pew Research Center findings confirm that users presented with these summaries become far less likely to click through to publishers' websites. Scola points to an *NPR* headline calling the shift an "extinction-level event" for online newsrooms, which depend on site traffic to generate advertising revenue and subscriptions. While Google has insisted publishers are misinterpreting the data, the economic fallout remains severe.

Simultaneously, an internal rift has emerged over media organizations appeasing AI developers. High-profile publishers such as *The Atlantic* entered commercial licensing agreements with OpenAI to feed journalistic archives into generative models as premium training and retrieval data. While staff writers protested that these arrangements simply hollow out their work without redirecting meaningful traffic back to their reporting, executive leadership countered bluntly: the licensing revenue made the publication profitable for the first time in years, leaving journalists told to simply get on board.

Underlying all these economic strains is a fundamental philosophical problem: the erosion of humanity. Journalism is, at its core, a human transaction—an exchange of information, emotional nuance, and lived experience between people. When algorithms replace that human relationship, public faith plummets; studies already demonstrate that readers trust reporting even less once they suspect AI is involved in its production.

---

[06:45](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=405s) Collaborations: Defeating "Idea Block" and Instant Subject Expertise

Pivoting from the "parade of horribles," Scola outlines how generative tools can serve as potent editorial assistants for individual beat reporters. One common challenge is burnout-induced "idea block." Citing writer Jennifer Peters—who described feeling as though she had run out of all original thoughts—Scola demonstrates how query engines like Perplexity provide an ideation sounding board. When an editor asks for a fresh angle on AI adoption in government, for example, a targeted prompt might suggest exploring the shifting procurement dynamics between federal agencies and defense contractors—raw material a reporter can refine into a viable pitch.

AI search also addresses a fundamental demand of daily journalism: acquiring deep domain literacy on an impossible deadline. Scola recalls covering an unexpected shake-up in the Department of Justice’s antitrust division where two officials were abruptly fired. While interviewing a source, the contact asked if she was familiar with the Tunney Act of 1974. Admitting she was not, Scola kept the source on the line, quickly queried Perplexity to understand the law's relationship to DOJ transparency, and mastered the context in real time. Within ninety seconds, she was dialing her next source, fluently asking how the Tunney Act applied to the firings.

---

[08:40](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=520s) Scaling Investigations: High-Volume Document Audits and Public Meetings

AI tools have proven equally transformative for processing massive data troves that would otherwise overwhelm small investigative teams. Scola highlights two major newsroom investigations enabled by model-driven analysis:

* **Bloomberg Law (2021 Voting Rights Analysis):** To gauge the fallout of a pivotal 2021 Supreme Court ruling on voting rights litigation across the country, reporters needed to review 600 complex court dockets. By combining large language models with human editorial spot-checks, the team successfully analyzed all 600 cases and demonstrated a sharp drop in subsequent voting rights filings—a project that would have taken months of dedicated reporter hours manually. * **The New York Times (Munitions in Southern Gaza):** Visual investigators used computer vision and AI tools to analyze satellite imagery of Southern Gaza, identifying craters wider than 40 feet—the telltale footprint of 2,000-pound munitions. The system identified more than 200 distinct craters, which human journalists subsequently verified to prove that heavy munitions were deployed in designated civilian safe zones.

```mermaid flowchart LR accTitle: Investigative Data Processing with AI accDescription: Workflow showing how large data pools are ingested by AI models and validated by human journalists.

subgraph DataIngestion["Unstructured Data"] A["600 Court Dockets<br/>(Bloomberg Law)"] B["Satellite Imagery<br/>(The New York Times)"] C["Public Meeting Video<br/>(Minutes / Michigan)"] end

subgraph AIProcessing["AI Processing Layer"] D["LLM Pattern Extraction"] E["Computer Vision (>40ft Craters)"] F["Speech-to-Text & Keyword Tagging"] end

subgraph HumanVerification["Journalistic Output"] G["Reporter Fact-Check & Context"] H["Investigative Publishing & Alerts"] end

A --> D --> G B --> E --> G C --> F --> H G --> H ```

In day-to-day beat reporting, automated audio transcription has evolved from simple time-savers into searchable databases. Beyond Otter’s basic live transcription and audio-sync playback, Otter AI Chat allows reporters to run natural-language queries across years of archived interview transcripts (e.g., retrieving a half-remembered conversation comparing U.S. and French regulatory approaches without remembering the speaker's name or date).

At the local level, civic innovation projects are addressing severe staffing deficits. In Michigan, a programmer and a journalist built *Minutes*, a tool designed to track over 100 civic bodies—school boards, county commissions, and city councils—that a five-person newsroom could never attend in person. The system ingests public video streams, transcribes the testimony, and delivers automated alerts whenever specific topics like "land use development" are mentioned by public officials.

---

[11:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=676s) Editorial Workflows: NotebookLM, Perspective Testing, and the "Dreaded Third Third"

For synthesis and narrative structure, reporters are increasingly turning to specialized tools to enhance their daily workflow:

* **NotebookLM as an "External Brain":** Scola notes that reporters upload primary research, interview transcripts, and trusted source documents into Google's NotebookLM, creating a private reference engine that answers queries grounded exclusively in verified source material, eliminating public-web hallucinations. * **Style and Metaphor Generation:** Beyond basic grammar checking through tools like Grammarly, Scola uses models like Claude to brainstorm figurative framing. While writing about the uneven spread of a new social media network, she prompted Claude to suggest metaphors drawn from biological contagion, turning an otherwise dry technology story into an engaging narrative. * **Perspective Checking ("A Second Pair of Eyes"):** To stress-test stories against potential blind spots, Scola prompts Perplexity to simulate how different political demographics might interpret an issue. In an experimental query asking how Trump supporters, liberal Democrats, and libertarians would react to naming the hot dog the national food of the United States, the model reasoned that Democrats might embrace it due to its historical and cultural ties to the working class. * **Drafting the "Dreaded Third Third":** The bottom third of a standard news report typically consists of standard background boilerplate—historical context and past rulings that a reporter has already written many times. Recalling her coverage of the multi-year *Google v. Oracle* legal battle for *Politico*, Scola explains that reporters often dread rewriting identical case background for the seventh time. By deploying Retrieval-Augmented Generation (RAG) over an archive of previously vetted, editor-approved, and fact-checked stories, journalists can automatically synthesize accurate background context, freeing their time to focus on breaking news and original reporting.

[14:02](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=842s) Collaborative Workflows: Headlines, Social Copy, and Newsletters

Generative AI offers practical relief for journalists dealing with the exhausting peripheral tasks of publishing—especially "the horrors that are headlines and social media." When an exhausted writer finishes a story and their editor asks for headline options and promotional angles, language models can jump-start the brainstorming process. Prompting ChatGPT on a recent piece about the normalization of political betting yielded viable options such as *"Odds on the Oval Office."* Similarly, tools like Claude can adapt copy for diverse social platforms like X (formerly Twitter) and LinkedIn, matching specific nuances, adjusting humor, or tailoring the tone to target distinct demographics like millennials.

Beyond social media, AI streamlines the packaging and distribution of editorial content across formats. Newsrooms routinely use AI to condense lengthy reporting into engaging newsletter formats or curate topical roundups. For instance, *The Telegraph* produces an automated healthy living newsletter by using AI to scan their website daily, collect relevant health reporting, and package it for subscribers automatically.

```mermaid flowchart TD accTitle: AI Collaborative Workflow in Modern Newsrooms subgraph Editorial ["Editorial Core"] A["Finished Story / Article"] end

subgraph Repurposing ["AI Repurposing & Distribution"] B["Brainstorm Headlines"] C["Platform-Tailored Social Copy"] D["Automated Topical Newsletters"] end

A --> B A --> C A --> D ```

- **Headline Brainstorming**: AI helps unblock tired reporters by generating creative headline hooks and alternative angles. - **Social Tone Modulation**: Journalists can instruct models to shift tone from serious to humorous or customize voice for specific target audiences. - **Automated Curation**: Automated newsletter systems autonomously parse publishing archives to assemble themed roundups.

---

[15:40](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=940s) Expanding Global Reach: AI Translation and Synthetic Audio

AI-driven translation and audio narration provide newsrooms with unprecedented capabilities to expand their audience reach without proportional increases in overhead. Major outlets rely heavily on AI translation engines: *The New York Times* publishes Spanish and Simplified Chinese editions, while the Associated Press utilizes machine translation with an explicit disclaimer noting that the content was generated by AI and vetted by human editors. French daily *Le Monde* originally planned an English newsletter, but after seeing AI translation capabilities, they launched a full-scale English-language version of their entire website.

In parallel, audio generation tools make written reporting accessible to broader audiences through synthetic voice narration. Outlets like *The Washington Post* embed automated audio players within articles, allowing readers to click a headphone icon to hear stories (such as pieces by reporters like Andrea Gimenez) read aloud by synthetic voices.

```mermaid flowchart LR accTitle: Multi-Format Content Expansion subgraph Input ["Source Material"] A["Original Reporting"] end

subgraph AI_Transform ["AI Transformation Layer"] B["Translation Engine"] C["Text-to-Speech Engine"] end

subgraph Formats ["Expanded Output"] D["Multilingual Editions<br/>(Spanish, Chinese, English)"] E["Synthetic Audio Articles"] end

A --> B A --> C B --> D C --> E ```

- **Disclaimers and Oversight**: Outlets like the AP maintain editorial transparency by explicitly labeling translated pieces as AI-generated and human-reviewed. - **Full-Scale Localization**: Rather than limiting translation to summaries or newsletters, publications like *Le Monde* translate full websites into English. - **Synthetic Voice Narration**: In-line audio players dynamically generate spoken-word versions of text articles to improve accessibility.

---

[16:41](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1001s) Case Study: The New York Times, Copyright Licensing, and Editorial Responsibility

*The New York Times* serves as a prominent case study in how top-tier news organizations navigate the legal, business, and editorial challenges of AI. While the organization made headlines by aggressively suing OpenAI and Microsoft for feeding copyrighted *Times* content into their training datasets without permission or compensation, they concurrently inked a licensing agreement with Amazon. When asked about the apparent contradiction, the fundamental distinction came down to fair compensation: *"We're getting paid. Like that's the whole point of this—is to be compensated for our work."*

Internally, *The Times* is transparent about leveraging AI across their operations—from identifying crater patterns in investigative data to generating translations, summaries, and audio. However, the organization maintains a strict line on ultimate editorial accountability: nothing published is fabricated out of whole cloth by AI, and all published material remains the sole responsibility of human journalists. To conclude, for those interested in a deeper exploration of this landscape, a 46-page collaborative report by the *Columbia Journalism Review* provides an extensive roadmap of the intersection of AI and modern journalism.

- **Litigation vs. Licensing**: Suing uncompensated scrapers while licensing archives to paying tech partners like Amazon establishes a sustainable commercial framework. - **Editorial Ground Rules**: AI augments data analysis, headline ideation, and audio narration, but human journalists retain final responsibility for all published reporting. - **Further Reading**: A 46-page report from the *Columbia Journalism Review* offers comprehensive research on AI adoption across the journalism industry.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=4s) We need to set the terms or we'll all be screwed. — Nancy Scola

> [02:18](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=138s) white Protestant culture

> [04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) extinction-level event — Nancy Scola

> [06:45](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=405s) parade of horribles — Nancy Scola

> [11:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=676s) External Brain — Nancy Scola

> [11:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=676s) A Second Pair of Eyes — Nancy Scola

> [11:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=676s) Dreaded Third Third — Nancy Scola

> [14:02](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=842s) the horrors that are headlines and social media — Nancy Scola

> [14:02](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=842s) Odds on the Oval Office

> [16:41](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1001s) We're getting paid. Like that's the whole point of this—is to be compensated for our work. — Nancy Scola

## Mentioned Nancy Scola, Politico Magazine, Washingtonian, Politico, The Washington Post, Bruce Reed, The Guardian, Sports Illustrated, Drew Ortiz, CNET, Los Angeles Times, Patrick Soon-Shiong, Insights, Google, Google AI Overviews, Pew Research Center, NPR, The Atlantic, OpenAI, Jennifer Peters, Perplexity, Department of Justice, Tunney Act of 1974, Bloomberg Law, The New York Times, Otter, Otter AI Chat, Minutes, NotebookLM, Grammarly, Claude, ChatGPT, X, LinkedIn, The Telegraph, Associated Press, Le Monde, Andrea Gimenez, Microsoft, Amazon, Columbia Journalism Review

## Video description Nancy Scola, Veteran Technology Reporter

[Source](https://www.youtube.com/watch?v=D6pJ4Tk-B-A)

中文简报

Google AI Overviews 分流新闻流量，出版商转向采用 AI 工具 新闻编辑室面临来自零点击搜索引擎的生存级收入威胁，迫使出版商通过技术授权协议将历史档案变现，同时在严格的人工监督下部署生成式模型以扩大调查性报道规模。 Scola 将 AI 的到来置于近几十年来已流失 80% 就业岗位的美国报业萧条背景之下。她指出了切实的冲突点——包括低劣的“AI 垃圾内容”、充斥幻觉的财经解读、存在缺陷的算法偏见工具，以及分流网站流量的零点击搜索概览。然而，她也展示了记者如何成功与生成式工具协作，以克服构思瓶颈、在截稿期限前快速掌握特定领域知识、审查海量档案文档并精简多格式发布流程。最终，结合 The New York Times 等案例研究，她认为新闻编辑室必须在确保核心编辑责任严格由人工把关的前提下，争取到公平的商业回报。 AI 技术正在深刻冲击本已脆弱的美国传统新闻业；1990 至 2025 年间，美国报业从业人数锐减 80%，逾 3000 家报纸关停，导致约 200 个县沦为没有本地媒体的“新闻沙漠”。 (0:04) 部分媒体为了缩减成本草率部署自动化采编，引发了低质内容与严重事实硬伤，例如 CNET 错误的复利解释、《体育画报》虚构 AI 记者署名，以及《洛杉矶时报》自动化工具因美化三K党而紧急下架。 (2:18) 谷歌 AI Overviews 等“零点击搜索”大幅截流了媒体网站的访问量、广告与订阅收益，而部分出版社与 AI 厂商签署的数据授权协议也在内部引发了采编人员对劳动成果被剥削的强烈抗议。 (4:28) 在深度报道领域，AI 工具展现出强大的协作潜力，例如彭博法律结合大语言模型高效分析了 600 份法院卷宗，《纽约时报》则利用计算机视觉识别出加沙南部的 200 多个重型弹坑以核实轰炸情况。 (8:40) 主流媒体正在商业变现与专业准则之间寻求平衡，如《纽约时报》在起诉未授权抓取数据的科技公司之余达成合规授权，并明确确立所有发布内容必须由人类记者承担最终编辑责任的底线。 (16:41)

Reading summary

Technology journalist Nancy Scola argues that artificial intelligence presents both an existential collision and a transformative collaborative tool for an already depleted news industry. While automated content and traffic-siphoning search engines jeopardize publisher survival, journalists are actively adopting AI to accelerate complex investigations, acquire rapid subject expertise, and streamline routine newsroom production. Structural Collapse and Zero-Click Audience Loss (0:04) AI has entered an American news landscape that has already contracted precipitously. Between 1990 and 2025, U.S. newspaper employment plummeted by 80%, more than 3,000 newspapers closed over the past two decades, and approximately 200 counties became news deserts without any local coverage. Scola notes that generative tools do not enter a stable ecosystem, but rather deliver an additional shock to an industry already fighting for economic viability. Compounding this fragility, Google’s AI Overviews and zero-click search synthesize answers directly on search engine results pages. Pew Research Center findings show that users presented with these summaries become far less likely to visit primary news sites, prompting an NPR report to describe the shift as an extinction-level threat to digital ad models. While publications like The Atlantic have signed commercial licensing deals with OpenAI to secure needed profitability, staff journalists have raised concerns that these pacts trade away archives without returning traffic or preserving reader trust. Hallucinations, AI Slop, and Failed Algorithmic Neutrality (2:18) Early publisher attempts to cut costs through automated production have resulted in significant editorial failures. Instances of low-quality copy—often termed 'AI slop'—include Sports Illustrated publishing product reviews attributed to a fabricated writer named 'Drew Ortiz,' as well as CNET publishing an AI-generated personal finance explainer that fundamentally miscalculated compound interest. Scola points out that public trust drops further whenever audiences suspect automated tools are involved in reporting. Efforts to replace human editorial judgment with algorithmic balancing tools have proven equally volatile. At the Los Angeles Times, an experimental tool called 'Insights,' intended to provide counter-arguments on opinion pieces, generated a rebuttal asserting that the Ku Klux Klan was merely an expression of 'white Protestant culture' rather than an organization rooted in racial hatred. The feature was quickly decommissioned, highlighting the hazard of using automated systems to arbitrate sensitive historical and political context. Scaling Investigations and Civic Monitoring (8:40) While autonomous generation struggles with editorial nuance, AI models have proven effective at processing massive datasets for investigative reporting. Bloomberg Law combined large language models with reporter spot-checks to analyze 600 complex court dockets, revealing a nationwide decline in voting rights lawsuits after a 2021 Supreme Court ruling. Similarly, visual investigators at The New York Times used computer vision to inspect Southern Gaza satellite imagery, identifying more than 200 craters wider than 40 feet to confirm that 2,000-pound munitions had been dropped in designated civilian safe zones. In beat reporting and local government coverage, AI tools help compensate for acute newsroom staffing deficits. In Michigan, the civic monitoring tool Minutes tracks over 100 public school boards, city councils, and county commissions by ingesting public video streams, generating transcripts, and issuing alerts when specific topics like land use are debated. For individual reporters, conversational transcription search engines like Otter AI Chat allow reporters to query years of interview archives using natural language without needing to remember specific dates or source names. Workflow Augmentation and Multi-Format Publishing (11:16) Reporters are increasingly using generative tools to reduce day-to-day cognitive and structural friction. Scola recalls querying Perplexity while reporting on Department of Justice firings to comprehend the 1974 Tunney Act within 90 seconds, allowing her to interview sources fluently on deadline. For drafting, models help journalists test political perspectives, generate metaphors, and draft background context—the 'dreaded third third' of recurring news stories—via Retrieval-Augmented Generation anchored in verified past reporting. Tools like Google's NotebookLM also serve as hallucination-resistant research assistants by restricting responses strictly to uploaded source material. At the distribution stage, AI handles peripheral publishing tasks like headline brainstorming, platform-tailored social copy for networks like LinkedIn and X, and automated newsletter curation, such as The Telegraph's health roundup. Outlets are also broadening audience reach through automated translation—exemplified by Le Monde translating its entire site into English and the Associated Press publishing disclaimed, human-reviewed translations—alongside synthetic voice tools that generate audio articles for readers at The Washington Post. News organizations are negotiating this transition by pairing legal resistance with selective commercial partnerships and strict editorial boundaries. As shown by The New York Times—which sued OpenAI and Microsoft over uncompensated content scraping while simultaneously licensing archives to Amazon—publishers are demanding compensation for their intellectual property. Across all newsroom applications, Scola emphasizes that human journalists must retain ultimate editorial responsibility for every published fact.

精炼阅读摘要

資深科技記者 Nancy Scola 指出，在美國報業經歷從業人員銳減80%的歷史性重創背景下，人工智慧的介入正對新聞業構成存在性震盪；它既因產出低劣內容與零點擊搜索而加劇生存危機，同時也在調查報導、繁雜工作流優化及跨語言傳播中展現出顯著的協同價值，但新聞責任的底線始終依賴人類的專業把關。 產業衰退與生成式內容的編輯危機 (2:18) 在探討人工智慧衝擊前，Scola 梳理了美國新聞業已面臨的嚴峻困境：1990年至2025年間，美國報業就業人數縮減了80%；過去20年中，超過3000家報社關閉，平均每週倒閉超過兩家，導致全美約200個縣完全缺乏在地新聞機構。在行業根基已極度脆弱的環境下，部分媒體試圖藉助生成式AI削減勞動成本，卻引發了嚴重的質量與倫理失控。例如，《體育畫報》（Sports Illustrated）曾虛構名為 Drew Ortiz 的AI作者發表排球測評等低質內容；科技媒體 CNET 依賴AI撰寫的基礎理財說明文，則被揭露在複利計算等常識概念上存在重大事實錯誤。 媒體管理層試圖利用算法取代人類編輯判斷的做法同樣遭遇挫敗。面對公眾對新聞客觀性的質疑，《洛杉磯時報》（Los Angeles Times）資方曾設想引入自動化「偏見計量器」，在內部引發反彈後轉為在評論專欄測試名為「Insights」的反方觀點生成工具。該實驗在處理一篇探討三K黨（Ku Klux Klan）的專欄文章時，竟生成了為其辯護的反論，聲稱該組織僅是「白人新教文化」的體現而非建立於種族仇恨，導致該工具因引發強烈爭議而迅速下線。 零點擊搜索與版權變現的商業撕裂 (4:28) 生成式AI平台正實質性地分流原創媒體賴以生存的受眾群體。以 Google 推出的「AI Overviews」為代表的零點擊搜索模式，直接在搜索結果頁面聚合答案，大幅削弱了用戶點擊進入原創內容鏈接的意願。皮尤研究中心（Pew Research Center）的數據證實了這一流量下滑趨勢，而 NPR 更將此形容為數位新聞編輯室的「滅絕級事件」，嚴重衝擊了仰賴頁面訪問量獲取廣告和訂閱收入的商業模式。 面對生存壓力，新聞機構內部在如何應對AI技術巨頭方面產生了深刻分歧。《大西洋月刊》（The Atlantic）等知名媒體選擇向 OpenAI 簽約授權其內容存檔用於模型訓練與檢索，管理層直言授權費用讓刊物多年來首次實現盈利，但基層記者普遍抗議此舉無異於出賣原創心血且無法帶來實質反向流量。此外，研究表明受眾一旦察覺報導有AI介入，信任度就會顯著下降，這使得以人際信任為基石的新聞業面臨更深層的信譽侵蝕。 調查報導與基層監督的規模化賦能 (8:40) 儘管面臨風險，生成式AI在賦能前線記者突破思維瓶頸與加速專業知識獲取方面成效顯著。Scola 分享了個人經驗：在採訪美國司法部反壟斷部門突發人事變動時，她借助 Perplexity 在90秒內迅速掌握了此前不熟悉的1974年《塔尼法案》（Tunney Act）背景，從而得以在後續採訪中準確追問。在應對海量非結構化數據時，AI更展現出傳統人工無法比擬的處理效率。 在大型調查實踐中，彭博法律（Bloomberg Law）利用大語言模型結合記者的人工抽查，成功審核了600份複雜的法院卷宗，量化證實了2021年最高法院一項關鍵裁決後投票權訴訟顯著減少的趨勢；《紐約時報》（The New York Times）視覺調查團隊則運用計算機視覺技術掃描加沙南部的衛星圖像，識別出200多個直徑超過40英尺的彈坑，由人工記者核實後揭露了2000磅重型彈藥被投放在平民安全區的實況。在地方監督層面，密西根州開發的「Minutes」項目通過自動語音轉錄和關鍵詞警報，幫助僅有5名記者的編輯室有效監控逾100個基層公共委員會的會議視頻。 采編工作流革新與跨媒介傳播拓展 (11:16) 在日常編輯環節，AI工具已深度嵌入內容提煉與分發體系。記者利用 Google 的 NotebookLM 上傳一手採訪與核實文獻，構建杜絕公開網絡幻覺的私有參考引擎；針對需要頻繁複述案件背景的重複性段落（即「可怕的三分之一」），采編團隊可利用檢索增強生成（RAG）技術從已過審的既往報導檔案中精確提取背景資訊。此外，ChatGPT 與 Claude 廣泛承擔了標題擬定（例如提煉出「Odds on the Oval Office」等視角）、社群媒體文案語氣轉換，以及《每日電訊報》（The Telegraph）等媒體自動彙編專題通訊的輔助工作。 在擴大全球受眾覆蓋方面，AI翻譯與合成語音顯著降低了營運門檻。《紐約時報》藉此維護西班牙文與中文版面，美聯社（AP）在標明AI生成並經人工校對的前提下推行多語種轉譯，法國《世界報》（Le Monde）更依託機器翻譯上線了全英文網站。與此同時，《華盛頓郵報》（The Washington Post）等媒體在報導內嵌由合成語音朗讀的音頻模組，使內容消費更具無障礙便利性。 Scola 總結指出，《紐約時報》一方面對未經授權抓取數據的 OpenAI 和微軟提起版權訴訟，另一方面與亞馬遜達成付費授權協議，表明新聞業在尋求合理經濟補償與維護知識產權上的現實考量。新聞機構必須嚴格守住「任何報導均不由AI憑空捏造、最終發布責任完全由人類承擔」的底線；正如《哥倫比亞新聞學評論》（Columbia Journalism Review）46頁專案報告所反映的，媒體必須在確立主導權的前提下探索可持續的AI共存路徑。

Detailed notes — zh

## Background 本视频由频道 Ai4 于 2025 年 8 月 27 日发布在 YouTube 平台。页面简介显示，内容为资深科技记者 Nancy Scola 发表的关于人工智能与新闻业关系的主题演讲。 ## TL;DR 演讲首先剖析了美国传统报业面临岗位断崖式削减与报社倒闭潮的严峻背景，并指出 AI 带来的低质垃圾内容、事实错误、平台截流零点击搜索以及削弱读者信任等剧烈“碰撞”。随后，Scola 结合一线采写经验转向积极的“协作”面，展示了生成式 AI 在破除选题瓶颈、突发采访中瞬时补足专业盲区、辅助彭博法律与《纽约时报》进行海量数据及卫星图像挖掘的实操潜力。她进一步分享了利用 Otter、开源工具 Minutes、NotebookLM 以及大模型进行会议追踪、搭建个人外脑、自动化撰写新闻背景段落与定制多平台文案的工作流。最后，她以《纽约时报》起诉 OpenAI 但签约亚马逊的商业抉择为例，强调新闻业拥抱 AI 的核心在于版权获偿与始终坚持人类记者的最终责任制。 ## Atmosphere & style 整场演讲兼具资深科技记者的冷静审视与一线从业者的机智幽默，开局对新闻业凋敝现状的剖析带着切肤的沉重与焦虑。随着话题切换至具体报道实战，Scola 以松弛、坦诚且富有自嘲精神的叙事风格，生动还原了在突发电话采访中争分夺秒查法案、截稿前用 AI 头脑风暴隐喻的真实采写切面。整场演讲的情绪从对技术蚕食媒体生态的警惕，自然过渡到利用工具赋能个体的实用主义探索，现场氛围兼具行业洞见的敏锐与工具实操的充实感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=4s) AI 与新闻业的关系并非单纯破坏性的对立，在行业普遍的焦虑悲观论调下，实际存在着切实为一线记者采编赋能的深度协作场景。 - [01:23](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=83s) AI 降临在美国新闻业极其脆弱的生存危机时刻：1990至2025年间报业岗位流失达80%，超3000家报纸倒闭，导致200多个县沦为无本地媒体的“新闻荒漠”。 - [02:18](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=138s) 媒体迫于运营压力批量引入 AI 生成低质水文（AI Slop）或产生基本事实错误（如 CNET 复利科普翻车），对专业媒体的品牌公信力造成毁灭性打击。 - [03:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=196s) 试图用算法取代记者的专业新闻判断极易失控，例如《洛杉矶时报》测试 AI“去偏见”工具时输出洗白三K党的言论，引发严重伦理翻车。 - [04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) 谷歌等搜索引擎推广“AI 概览”引发“零点击搜索”，直接截流原始新闻网站流量，使出版商面临严峻的生存危机并倒逼其妥协授权。 - [06:12](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=372s) 新闻的核心在于人与人之间的深度连接，调研显示读者一旦发现报道有 AI 参与痕迹其信任度会骤降，滥用 AI 正在摧毁本就脆弱的读者信任。 - [06:45](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=405s) 生成式 AI 能有效破解记者的“选题枯竭”困境，充当头脑风暴伙伴，帮助记者在宏大模糊的报道需求中快速提炼出具有商业价值的切入角度。 - [07:51](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=471s) AI 赋能记者实现“瞬时专业度”（Instant expertise），在突发采访间隙用一两分钟搞懂陌生法案或历史渊源，保障追问深度。 - [08:27](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=507s) “大模型初筛 + 记者人工复核”的模式扩展了深度调查报道的边界，被成功应用于数以百计的法律诉讼判例梳理与战区巨型弹坑的卫星影像筛查。 - [09:49](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=589s) 智能转录与开源监控工具（如 Minutes）打破了地方公共事务监督的人力瓶颈，让精简的新闻室能够自动化追踪并警报全州上百场市政会议。 - [11:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=676s) 基于记者专有采访笔记构建的封闭语料库（如 NotebookLM）能充当没有幻觉风险的“第二大脑”，同时还可辅助语言修辞与全新认知隐喻的构建。 - [12:37](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=757s) 利用基于已审校旧稿的 RAG（检索增强生成）技术，可以自动重组新闻底部繁琐的案情与事件背景，彻底消除记者重复劳动的负担。 - [15:10](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=910s) 机器翻译与合成语音等 AI 工具正在重塑内容分发，推动传统严肃外媒以极低成本实现全站级内容出海与多模态音频转化。 - [16:52](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1012s) 新闻媒体与 AI 巨头的关系正走向务实的商业变现，从版权诉讼到达成内容授权合作，核心诉求在于媒体智力劳动必须获得应有的经济补偿。 - [17:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1048s) 严肃媒体为内部应用 AI 划定了不可逾越的红线：严禁发表全篇由 AI 生成的内容，报道终审与事实准确性的最终责任永远归属于人类记者。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=4s) 演讲人背景与主题概览：AI 与新闻业是碰撞还是合作？ 资深科技记者 Nancy Scola 开门见山地抛出了演讲的核心命题：人工智能与新闻业之间，究竟是一场破坏性的“碰撞”（Collision），还是一次富有成效的“合作”（Collaboration）？为了让现场听众明白台上的主讲人是谁，Scola 幽默地自我调侃并介绍了自己的背景。她常驻华盛顿特区，主要报道科技、政治、公共政策以及宏观经济，目前是《Politico》杂志和《Washingtonian》杂志的特约撰稿人，此前曾长期担任《Politico》和《华盛顿邮报》（The Washington Post）的专职记者。她长期深耕科技政策领域，撰写过多篇关于人工智能的重磅报道，其中包括对拜登政府时期白宫 AI 核心负责人布鲁斯·里德（Bruce Reed）的专题报道。 Scola 梳理了整场演讲的脉络：首先将剖析当前 AI 与新闻业之间爆发的几大剧烈冲突与危机，随后迅速切换视角，从一名一线调查记者的实操经验出发，探讨 AI 在实际新闻采写工作中展现出的积极协作潜力与上升空间。她指出，业内讨论该话题时往往充斥着悲观论调，但实际上存在着能够切实帮助一线记者的赋能场景。 ### [01:23](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=83s) 新闻业的脆弱现状：行业萎缩与 AI 带来的双重危机 在深入探讨 AI 带来的冲击之前，Scola 首先确立了一个行业基本盘背景：为什么新闻从业人员对 AI 如此焦虑？答案在于“记者本身就对一切充满忧虑”，而且这种焦虑完全是有据可循的。美国报业正经历着前所未有的结构性崩溃，从 1990 年到 2025 年，美国报业的工作岗位大幅削减了 80%。这种断崖式下跌的直接原因在于报纸实体的大规模倒闭——过去 20 年间，美国有超过 3000 家报纸停刊，相当于平均每周关闭超过两家报社。 如今全美约有 200 个县已经彻底沦为“新闻荒漠”，没有任何本地新闻机构覆盖。在新闻业已经深陷生存危机的当下，AI 的横空出世犹如雪上加霜，成为悬在整个行业头上的又一柄利剑。《卫报》（The Guardian）的一则标题精准道出了业界的普遍情绪："We need to set the terms or we'll all be screwed"（我们必须主动确立规则，否则大家都会完蛋）。 - **岗位缩减**：1990 年至 2025 年间，美国报业岗位流失达 80%。 - **报社倒闭**：过去 20 年间全美关停逾 3000 家报社，平均每周关闭两家以上。 - **新闻荒漠**：全美约 200 个县已没有任何本地新闻媒体。 ### [02:18](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=138s) 碰撞危机一：AI 垃圾内容与虚假事实的泛滥 AI 与新闻业碰撞的首要危机在于低质“AI 垃圾”（AI Slop）泛滥，以及用劣质生成内容冒充专业新闻从而摧毁新闻质量。以知名体育媒体《体育画报》（Sports Illustrated）为例，该刊曾发表由虚构作者“Drew Ortiz”署名的排球评测文章，通篇皆由 AI 生成，行文质量极其平庸，毫无专业媒体应有的普利策级水准，纯粹是为了以次充好、兜售给读者。 更严重的问题在于，AI 生成的内容不仅文笔糟糕，还经常出现根本性的事实错误。老牌科技媒体 CNET 在经营陷入困境时，试图通过引入 AI 大规模生成基础解释性内容（explainers）以维持运营。然而在其发布的一篇关于“复利”（Compound Interest）的科普报道中，AI 却给出了完全错误的数学计算与概念解释。这种事实层面的重大翻车给这家老牌媒体的声誉带来了毁灭性打击。 ```mermaid flowchart TD accTitle: AI内容对新闻业声誉的破坏链路 subgraph S1["媒体运营压力"] A["营收困境与人力削减"] --> B["引入 AI 批量生成内容"] end subgraph S2["质量失控"] B --> C["AI Slop 低质水文充斥"] B --> D["核心概念与事实性错误"] end subgraph S3["行业后果"] C --> E["媒体公信力暴跌"] D --> E end ``` ### [03:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=196s) 碰撞危机二：用算法代替专业判断与失控的“去偏见”工具 另一个令新闻界深感担忧的碰撞，是管理层试图用所谓“客观中立的机器人”取代记者的专业新闻判断。媒体资方普遍认为公众对新闻界的不信任源于“媒体偏见”，因此异想天开地希望通过部署 AI 审核工具向公众证明自己没有偏见。《洛杉矶时报》（Los Angeles Times）在深陷运营泥潭时，其所有者黄馨祥（Patrick Soon-Shiong）曾提出在新闻稿件中部署“偏见测量仪”（Bias Meter），由 AI 实时给记者写的报道贴上“偏左”或“偏右”的标签，妄图以此重获读者信任。 该计划遭到报社采编团队的强烈抵制，随后被缩水退回到仅对社论和评论板块（Opinion）使用，推出了名为“Insights”的工具，用于抓取社论并提供对立观点。然而在实际测试中，当一篇专栏文章提到三K党（Ku Klux Klan）时，该 AI 工具竟然给出荒谬的对立论调，称三K党“不一定是一个基于种族仇恨的组织，而只是白人新教文化的一种体现”。由于这一严重的伦理翻车事故，该工具被彻底下架。 ### [04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) 碰撞危机三：平台截流“零点击搜索”与媒体对 AI 公司的妥协 大型 AI 平台正在直接剥夺新闻媒体本已极其匮乏的读者流量。以谷歌推出的“AI 概览”（AI Overviews）为例，用户在搜索引擎输入问题后，页面顶端会直接显示一段由 AI 总结好的完整答案，导致所谓的“零点击搜索”（Zero-click search），用户无需再点击进入原始新闻源网页。 皮尤研究中心（Pew Research Center）的调查显示，一旦搜索页面提供了 AI 总结，用户点击原始新闻链接的概率出现断崖式下跌。这对依赖页面流量、广告曝光和订阅转化来维系生存的传统媒体而言是毁灭性的打击。NPR 甚至在报道中惊呼，在线新闻出版商正面临来自谷歌 AI 搜索的“灭顶之灾”（Extinction level event）。尽管谷歌官方辩称外界误读了相关研究数据，但新闻出版业界的危机感已不可逆转。 与此同时，部分传统媒体选择向 AI 巨头妥协。《大西洋月刊》（The Atlantic）选择与 OpenAI 达成战略合作，将其优质新闻版权内容授权给对方作为训练与回答语料。这一举措引发了报社记者的强烈不满与抗议，编辑们抱怨 OpenAI 只是在单向汲取媒体的智力资产，并未真正导流读者。但管理层却冷酷地回应：“这是我们多年来首次实现盈利，赶紧跟上节奏吧。” ### [06:12](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=372s) 碰撞危机四：人性的丧失与读者信任的崩塌 Scola 指出，AI 对新闻业最本质的冲击在于“人性的剥离”。新闻的本质绝非机械的数据搬运，而是人与人之间关于信息、深刻洞见乃至情感的深度交流；一旦将人类记者从这个流程中剔除，整个新闻业存在的根基都将动摇。 调研数据表明，当读者意识到报道中有 AI 参与的痕迹时，他们对该新闻的信任度会进一步暴跌。在全美公众对新闻媒体信任度本已跌入历史冰点的背景下，滥用 AI 正在加速摧毁媒体与受众之间最后一丝脆弱的信任纽带。 ### [06:45](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=405s) 积极协作一：利用 AI 击破选题枯竭与寻找报道灵感 在列举完令人沮丧的现实危机后，Scola 将话题转向 AI 的积极协作价值，首要的应用场景便是破解记者的“选题枯竭”（Idea block）。对于职业记者而言，日复一日地策划原创报道选题是极度消耗心智的重担。正如作家 Jennifer Peters 所形容的那样："I sometimes feel like I've completely run out of ideas. Not just story ideas, but all original thoughts"（我有时觉得自己不仅是报道灵感枯竭，而是连所有原创思维都彻底被掏空了）。 在这种情况下，生成式 AI 工具（如 Perplexity）能够充当强大的头脑风暴助手。Scola 分享了自己的实际工作流程：当某家商业杂志的编辑打电话约稿，要求写一篇关于“政府如何使用 AI”的文章时，她会直接打开 Perplexity，输入定位“这是一份商业刊物，关于政府应用 AI 有哪些值得挖掘的选题方向？”尽管系统生成的很多点子并不尽如人意，但其中关于“联邦承包商（Federal Contractors）与政府之间的合作关系将如何因 AI 而被重塑”的切入点极具商业报道价值，记者只需在此基础上进行提炼打磨，就能迅速形成一份高质量的选题提案。 ### [07:51](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=471s) 积极协作二：实时获取领域新知与建立瞬时专业度 现代记者的核心职业要求之一是“瞬间建立专业度”（Instant expertise）——如果早上对某个领域一无所知，到当天傍晚就必须能够条理清晰、专业严谨地写出一篇深度分析。Scola 经常报道反垄断法案与司法动态，她分享了一个此前亲身经历的突发报道案例：美国司法部（DOJ）的两名高级官员突遭解雇，具体缘由扑朔迷离。 在 Scola 打电话向业内核心消息源探寻内幕时，对方反问她：“你熟悉 1974 年的《塔尼法案》（Tunney Act of 1974）吗？”面对自己完全陌生的法律概念，Scola 一边继续在电话中稳住消息源，一边立刻在电脑上调出 Perplexity，迅速输入查询指令，在极短的时间内搞清楚了这部历史法案与司法部透明度审查之间的法理渊源，从而确保了后续采写与追问的顺利推进。 ### [08:27](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=507s) 深度调查与数据挖掘：彭博法律与《纽约时报》的 AI 报道实践 在日常采访中，AI 工具能够瞬间为记者补足知识盲区。比如刚挂断司法部（DOJ）的电话，记者就能在短短一分半钟内迅速搞懂此前一无所知的法律条款，紧接着拨通下一个消息来源的电话，胸有成竹地追问该条款在事件中将起到怎样的作用。除了这种即时补课，大型新闻机构在处理海量数据调查时，也开始依赖大语言模型与人类专业判断的深度结合。2021 年美国最高法院就投票权作出重大裁决后，彭博法律（Bloomberg Law）希望梳理全美 600 起相关诉讼，以评估判决前后的司法风向转变。面对即便专门报道团队也难以消化的工作量，他们采用“大模型初筛 + 人工复核”的模式，最终证实全美投票权诉讼案件数量确实出现了急剧下滑。如果完全依靠人工，这类需要极高时效性的调查报道根本无法完成。 《纽约时报》则将视觉分析工具运用到了战地新闻核查中。在调查以军在加沙南部开展的军事行动时，记者通过向军事专家请教得知，重型弹药爆炸通常会在地面留下直径超过 40 英尺的巨大弹坑。为了验证以方在声称未使用大杀伤力武器的地区是否存在违规操作，团队利用 AI 工具对加沙南部的卫星遥感影像进行自动化检索与扫描，精准圈定出 200 多个符合特征的巨型弹坑。随后，多位视觉调查记者介入进行二次人工核查并锁定事实，最终推出了一篇极具说服力的调查重磅报道。 - 彭博法律分析了 2021 年最高法院判决前后的 600 起投票权案件，利用 LLM 与记者团队协作完成深度分析。 - 《纽约时报》通过 AI 扫描卫星影像，在加沙南部识别出 200 余个直径超 40 英尺的弹坑，揭露重型弹药的使用真相。 ### [09:49](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=589s) 听会与转录的革新：从 Otter AI 跨记录检索到密歇根“会议追踪器” 录音转录一直以来都是一线记者日常耗时最多、最枯燥的苦力活，而转录工具的革新正在颠覆这一流程。许多人熟悉的 Otter 不仅能提供近乎即时的转录并在点击文本时精确定位音频，其推出的 Otter AI Chat 功能更支持跨所有历史录音进行全局语义检索。记者不再需要抓狂地去回想“几个月前到底是谁跟我聊过美法对比的话题”，只需向 AI 描述模糊的记忆片段，系统便能在庞大的个人采访库中准确调出当年的对话原文。 创新的边界甚至延伸到了地方公共事务监督。在密歇根州，一家仅有 5 名记者的精简型地方新闻编辑室面临着全州数百个市议会、教育委员会闭门开会的庞大监督压力。一位程序员与一名记者联手打造了名为“Minutes”的开源自动化工具，系统性地抓取全州各类市政会议视频并生成文字转录，通过关键词自动监测特定议题。该平台如今向更多记者开放，覆盖了 100 多场定期市政会议；记者只需订阅“土地利用规划”等特定标签，一旦有议员在会议发言中提及，系统便会自动发送邮件警报并附上精准时间戳。 ```mermaid flowchart TD accTitle: 密歇根新闻室 Minutes 工具会议监控流程 subgraph 数据采集 A["地方政府会议视频流"] --> B["自动化视频抓取与转录"] end subgraph 智能处理 B --> C["NLP 语义解析与议题打标"] C --> D{"记者预设监控主题<br>(如: 土地使用、学区预算)"} end subgraph 记者触达 D -- 命中关键词 --> E["发送邮件通知与时间戳回放"] E --> F["记者介入核实并跟进深度采写"] end ``` ### [11:16](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=676s) 扩展外脑与概念重构：从 NotebookLM 到语言修辞辅助 新闻记者正在将 NotebookLM 等工具打造成个人专属的“第二大脑”。与在公网环境中漫游的通用大模型不同，记者可以将自己跑出来的独家采访实录、背景调研文献和一手材料全部上传至 NotebookLM。基于封闭语料库构建的检索，既拥有类似 Perplexity 或 Claude 的自然语言对话能力，又从源头上杜绝了模型幻觉，确保每一次追问调取的都是记者亲自验证过的真实事实。 在日常撰稿与修辞层面，AI 也展现出独特的辅助价值。除了用 Grammarly 进行基础校对和文本节奏润色外，记者开始借力 Claude 和 Perplexity 寻找全新的写作隐喻。例如，在报道一款新兴社交网络产品的用户扩张时，记者直觉上认为其扩散轨迹类似于某种“生物传染病”，但难以将这种抽象感觉落诸笔端。通过让 Claude 提供相关病理学隐喻与框架构建，记者能迅速获得丰富且贴切的生动表达素材。这并不是让机器代替写作，而是在漫长而疲惫的截稿周结束前，为精疲力竭的记者提供灵感火花与思维原料。 ### [12:37](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=757s) 借一双审视的眼睛与解决“令人头疼的末尾三分之一” 每位记者在写作时都难免带有个人经历与认知偏见，利用 AI 扮演不同背景的读者则能提供宝贵的“第二视角”。在一项有趣的测试中，提问者让 Perplexity 设想如果政府提议“将热狗定为美国国菜”，特朗普支持者、自由派民主党人和自由意志主义者分别会作何反应。模型合理地推理出，民主党人或许会很认同热狗背后深厚的蓝领工人阶层文化象征，这种意料之外的视点推演极具启发性。 与此同时，AI 在新闻工业中最立竿见影的减负场景，莫过于撰写那段“令人头痛的末尾三分之一”（The dreaded third third）。在严肃硬新闻中，报道底部往往需要花费几个段落交代事件的前因后果，而这些背景材料记者往往已经在过往报道中反复写过五六遍。例如在 Politico 追踪谷歌诉甲骨文案期间，每当法庭下达新裁定，记者就必须重写一遍早已写到吐的案情背景。理想的解决方案是建立基于 RAG（检索增强生成）的工作流，让模型直接调取编辑已经审校、事实核查过并已发表的过往旧稿，自动重组出精准严谨的背景补充段落，记者只需快速浏览微调即可交稿。 - 借助大模型模拟政治光谱各端（如特朗普选民、自由派、自由意志主义者）的读者反应，帮助记者打破信息茧房。 - 利用基于过往已核实稿件的 RAG 技术，能够自动生成新闻底部的背景回溯段落，彻底消除繁重无意义的重复写作。 ### [14:09](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=849s) 拟定标题与社交裂变：让 AI 搞定多平台文风定制 漫长的采写结束后，面对编辑“给标题和社媒宣发提几个备选文案”的要求，精疲力竭的记者往往很难再挤出好点子，而这正是 ChatGPT 与 Claude 大显身手的场景。在撰写一篇关于“政治下注日常化”的杂志专题时，向大模型索要备选题目的尝试便带来了惊艳成果——模型给出的“白宫赔率”（The Odds on the Oval Office）既精炼又富有文字趣味，直接提供了极高水准的备选方案。 对于不同社交平台的矩阵化运营，AI 工具更是能根据不同平台调性量身定制宣发文案： - **LinkedIn 风格**：注重行业趋势分析与专业度，适合沉稳理性的职业受众。 - **X (Twitter) 风格**：强化悬念与冲突感，注重短平快的信息密度与互动讨论。 - **语调与代际定制**：当文风显得过于板正时，记者可以直接向模型提出“希望能幽默一点”或“希望能打动千禧一代读者，但我不知道千禧一代怎么说话，帮我换成他们习惯的语气”，AI 能够瞬间调整出契合特定受众语感的表达方式。 ### [15:10](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=910s) 内容重构与全球化传播：自动化通讯、机器翻译与有声生成 在全媒体传播链条中，将既有报道快速包装为多种衍生内容已成为刚需。《每日电讯报》（The Telegraph）上线了一款全自动生成的“健康生活”邮件通讯，系统会定时扫描网站当天更新的所有内容，自动提取健康垂直领域的文章，整合成结构规范的 Newsletter 发送给订阅用户。在跨语言传播方面，《纽约时报》运用 AI 提供西班牙语和简体中文版本的机器辅助翻译；美联社（AP）在采用 AI 翻译的同时，严格执行合规标准，明确在稿件注明“本报道由 AI 生成并经人类编辑复审”的免责声明。 海外严肃媒体也在借此实现全球化跳跃。法国《世界报》（Le Monde）最初仅计划筹办一份英文电子邮件通讯，但在意识到现代 AI 翻译技术足以支撑全站级内容转化后，果断直接推出了完整的全英文版《世界报》，彻底打开了国际受众市场。《华盛顿邮报》等机构则普遍引入了 AI 语音合成技术，读者在新闻网页点击耳机图标，便能听到系统自动播报“本文作者为安德烈亚·希门尼斯（Andrea Gimenez），当前由 AI 合成语音为您朗读”。资源雄厚的大型媒体机构正持续投入重金，率先摸索新技术在新闻工业化流程中的最佳实践形态。 ### [16:52](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1012s) 从起诉OpenAI到签约亚马逊：内容价值与商业变现 新闻机构在面对 AI 浪潮时正展现出极为务实且清晰的商业逻辑，以《纽约时报》（The New York Times）的态度转变为典型代表。几年前，《纽约时报》认定 OpenAI 与微软（Microsoft）在未获授权且未支付任何补偿的情况下，擅自挪用时报的大量原创报道内容来训练和喂养其 AI 工具，随即对两家科技巨头提起了法律诉讼。这一维权诉讼在当时被外界视为极其强硬且激进的对抗举措。 然而就在近期，《纽约时报》却选择与亚马逊（Amazon）达成战略合作，允许亚马逊将其新闻内容接入并喂给亚马逊旗下的 AI 产品。面对外界对其态度转变的质疑与好奇——“究竟是什么发生了变化？”，《纽约时报》给出的回应非常直截了当："Well, we're getting paid."（“因为我们得到了报酬。”）这直接点明了新闻机构的核心诉求：整个事件的本质在于智力劳动与版权成果必须得到合理补偿；既然亚马逊愿意为此付费，这就达成了一笔双赢且精明的商业交易。 ### [17:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1048s) 内部AI使用边界、记者责任制与延伸报告 除了对外的内容授权战略，《纽约时报》对其内部如何应用 AI 也有着极其清晰且公开的规范准则。他们在内部广泛探索了 AI 的辅助功能，包括利用 AI 深度分析数据、提炼关键信息、辅助撰写新闻标题与摘要，以及自动生成多语言翻译和音频版本等。但时报划定了一条不可逾越的底线：报社所发表的任何内容，绝不能“完全由 AI 独立生成”（"created in whole cloth by AI"）。 时报明确强调，使用这些先进工具是为了有效提高生产力、放大采编团队的工作成效，但所有内容的最终责任始终由《纽约时报》的专业记者来承担。在演讲的尾声，主讲人向关注该领域的听众推荐了一份极具参考价值的权威行业报告。 - **核心责任原则**：AI 仅作为采编流程中的增强与辅助工具，终审与事实准确性的最终责任完全归属于人类记者。 - **推荐资料**：由《哥伦比亚新闻评论》（Columbia Journalism Review）等机构联合发布的专题报告，全文共 46 页，深度剖析了 AI 与新闻业交汇点上的各类关键议题与前沿实践。 ## Notable quotes > [01:23](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=83s) We need to set the terms or we'll all be screwed — The Guardian > [04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) Extinction level event — NPR > [04:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=268s) 这是我们多年来首次实现盈利，赶紧跟上节奏吧。 — The Atlantic 管理层 > [06:45](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=405s) I sometimes feel like I've completely run out of ideas. Not just story ideas, but all original thoughts — Jennifer Peters > [07:51](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=471s) 你熟悉 1974 年的《塔尼法案》（Tunney Act of 1974）吗？ — 消息源 > [14:09](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=849s) The Odds on the Oval Office — AI 模型 > [15:10](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=910s) 本文作者为安德烈亚·希门尼斯（Andrea Gimenez），当前由 AI 合成语音为您朗读 — 华盛顿邮报播报 > [16:52](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1012s) Well, we're getting paid. — The New York Times > [17:28](https://www.youtube.com/watch?v=D6pJ4Tk-B-A&t=1048s) created in whole cloth by AI — The New York Times ## Mentioned Nancy Scola, Bruce Reed, Politico, Washingtonian, 华盛顿邮报 (The Washington Post), 卫报 (The Guardian), 体育画报 (Sports Illustrated), CNET, 洛杉矶时报 (Los Angeles Times), 黄馨祥 (Patrick Soon-Shiong), Google, AI Overviews, 皮尤研究中心 (Pew Research Center), NPR, 大西洋月刊 (The Atlantic), OpenAI, Jennifer Peters, Perplexity, 美国司法部 (DOJ), 彭博法律 (Bloomberg Law), 纽约时报 (The New York Times), Otter.ai, Minutes, NotebookLM, Claude, Grammarly, ChatGPT, 每日电讯报 (The Telegraph), 美联社 (AP), 世界报 (Le Monde), Microsoft, Amazon, 哥伦比亚新闻评论 (Columbia Journalism Review) ## Video description Nancy Scola, Veteran Technology Reporter [Source](https://www.youtube.com/watch?v=D6pJ4Tk-B-A)

Source quotes

[1:23] The Guardian: We need to set the terms or we'll all be screwed [4:28] NPR: Extinction level event [4:28] The Atlantic 管理层: 这是我们多年来首次实现盈利，赶紧跟上节奏吧。 [6:45] Jennifer Peters: I sometimes feel like I've completely run out of ideas. Not just story ideas, but all original thoughts [7:51] 消息源: 你熟悉 1974 年的《塔尼法案》（Tunney Act of 1974）吗？ [14:09] AI 模型: The Odds on the Oval Office [15:10] 华盛顿邮报播报: 本文作者为安德烈亚·希门尼斯（Andrea Gimenez），当前由 AI 合成语音为您朗读 [16:52] The New York Times: Well, we're getting paid. [17:28] The New York Times: created in whole cloth by AI

Source quotes

[0:04] Nancy Scola: We need to set the terms or we'll all be screwed. [2:18] white Protestant culture [4:28] Nancy Scola: extinction-level event [6:45] Nancy Scola: parade of horribles [11:16] Nancy Scola: External Brain [11:16] Nancy Scola: A Second Pair of Eyes [11:16] Nancy Scola: Dreaded Third Third [14:02] Nancy Scola: the horrors that are headlines and social media [14:02] Odds on the Oval Office [16:41] Nancy Scola: We're getting paid. Like that's the whole point of this—is to be compensated for our work.

Original transcript

[0:04] We are here to talk um as was mentioned [0:06] to talk about AI and journalism whether [0:08] it's a collision or a collaboration. I'm [0:09] your host Nancy Skola. Um I'm going to [0:12] tell you just a little bit about me. Uh [0:13] a few quick seconds you're not wondering [0:15] who the hell excuse me who the heck is [0:16] up here uh talking to to you. So I'm [0:18] based in Washington DC where [0:20] everything's perfectly normal at the [0:21] moment. No worries. Uh, and I generally [0:26] write on technology, politics, public [0:27] policy, a little bit of economics. Um, I [0:31] am a contributing writer at Politico [0:32] magazine and Washingtonian magazine. [0:34] I've been a staff writer at Politico, [0:35] Washington Post. Here's a sampling of [0:37] some stories I've written on technology [0:39] um, and policy. A couple on AI. On the [0:41] upper right there is a story about Bruce [0:43] Reed, who was the White House point [0:44] person on AI in the Biden [0:46] administration. So, generally write on [0:48] AI fairly often. Um we're going to uh [0:51] just preview our topic. Um so what we're [0:55] going to talk about some uh first is [0:57] some of the collisions we're seeing [0:58] between AI and journalism and then we're [1:00] going to switch to moving at a pretty [1:02] quick clip um through what I think as [1:03] some of the potential collaborations [1:05] between AI and journalism. Generally [1:06] we've talked about this. It's a little [1:07] bit doom and gloom. I think there's some [1:09] potential from for some upside [1:11] particularly from the perspective of a [1:12] working reporter. So I'd like to talk a [1:14] little bit about that. So, um, and we're [1:16] going to move pretty quickly. If there's [1:17] a couple minutes left of, uh, [1:18] conversation questions at the end, I'd [1:20] love that. So, maybe be thinking about [1:21] what you might want to talk about. So, [1:23] first, let's start with a quick setting [1:24] of a baseline. Um, why are journalists [1:26] so worried about AI? Um, because [1:28] journalists are so worried about [1:29] everything. Uh, with good good reason. [1:32] Uh, we have a that on the chart on the [1:34] left that talks about the 80% decline in [1:37] newspaper industry jobs from 1990 to, [1:40] uh, 2025. 80%, right? That's a pretty [1:42] good dip in part because there are just [1:44] a lot fewer newspapers than there once [1:46] were. In the last 20 years, more than [1:48] 3,000 newspapers have closed in the [1:50] United States. More than two a week. Uh, [1:53] excuse me, two two a week. Yes. And the [1:56] um there's about 200 counties in the [1:58] United States that have no local news [1:59] outlets, right? So, there just far fewer [2:01] jobs. AI comes into the picture. It [2:04] represents one more threat to an [2:05] industry that's under a ton of threats. [2:07] The Guardian had this headline that I [2:08] think kind of sums up the mood pretty [2:10] well in the industry. forgive the [2:11] language, but it seems says we need to [2:13] set the terms or we'll all be screwed, [2:14] right? That's a sort of general feeling [2:16] in indust industry about AI. So, what [2:18] are we worried about? What are the some [2:19] of the collisions that we worry about [2:21] when it comes to AI and journalism? On [2:22] the left, we have a Sports Illustrated [2:25] review that was AI generated uh by a [2:27] fictional staff writer named Drew Ortiz [2:29] talking about volleyballs. If you can [2:31] see it from where you're sitting, you'll [2:32] see it's not exactly Pulitzer [2:34] Prizewinning uh type work. And the idea [2:37] there is basically AI slop. Using AI to [2:39] generate content, selling it to the [2:41] public as journalism um and degrading [2:43] the quality of journal journalism across [2:45] the board. Occasionally the writing is [2:47] not just bad, it's wrong. Um on the [2:49] right we have an example from CNET which [2:51] is a tech publication. Any CNET readers [2:53] at least back in the day? Yeah. Uh it's [2:55] a struggling publication. They in order [2:58] to sort of stay afloat they decide to [2:59] embrace the idea of using AI to generate [3:01] content um to produce basic explainers. [3:04] Um they produced one explaining compound [3:06] interest that explained compound [3:07] interest incorrectly. Um and so that you [3:10] know did not reflect particularly well [3:12] on the publication. [3:16] Uh other things we worry about when it [3:17] comes to collision of AI and journalism [3:19] the swapping of journalistic judgment [3:21] for unbiased bots. Right? News owners [3:23] publication owners have sort of gotten [3:25] to this idea that the public uh [3:27] generally distrusts the media because [3:29] they think we're all biased. what we'll [3:30] do is use bots to assess our journalism [3:33] and then therefore prove to the the [3:35] reading public that we're not as biased [3:37] as they once thought. Right? So this was [3:39] the LA Times uh which going through a [3:41] lot of challenges. The owner there [3:43] Patrick son uh Patrick Suniang floated [3:46] this idea of using a bias meter on the [3:48] publication's news stories. Reporter [3:50] puts out a story. The bias meter said [3:52] this leans left, this leans right. He [3:54] thought this would restore faith with [3:55] the reading public. Uh a lot of the [3:57] staffers were not exactly thrilled with [3:59] this idea. they decided to roll it back [4:01] only apply it to their opinion [4:02] journalism. Um, and they have this tool [4:04] called insights that evaluates opinion [4:07] journals in the paper and offers [4:08] contrary viewpoints. They have an [4:10] example where there was an a column that [4:12] mentioned the clux clan. The insights [4:14] tool came back and said the clan's not [4:16] necessarily a race u racial hatred based [4:19] organization. It's just an example of [4:21] white Protestant culture. Um, and they [4:23] took down the insights tool. Uh, we're [4:26] going to skip through this actually. the [4:28] uh hijacking readers. This is another [4:30] concern when it comes to AI. This idea [4:32] that the big um AI platforms are taking [4:35] readers from publications when they can [4:37] ill afford to lose readers. So we have [4:40] on the left we have Google's AI [4:42] overview. People notice the roll out of [4:43] AI overview in recent months. So the [4:46] general idea if you're not familiar with [4:48] it is you search for something on Google [4:50] rather than providing search results it [4:52] provides an AI summary answering your [4:54] query. Uh they sometimes call this [4:55] zeroclick search, right? you're not even [4:57] actually clicking through to a result. [4:59] Um, and the Pew Research Center did um [5:02] did some research on this and found that [5:04] people are just far less likely to [5:05] actually click through to a search [5:07] result when a summary was provided. Um, [5:10] and this is just, you know, really bad [5:12] news for news publications when people [5:14] are not actually going to our [5:15] publications, seeing our ads, supporting [5:18] supporting our work on our platforms. [5:19] Um, so this was taken in the news [5:22] industry as a pretty bad uh development. [5:25] You see NPR headline on the bottom right [5:26] there says online news publishers face [5:28] extinction level event from Google's AI [5:30] powered search. Right? Another one of [5:32] the reasons we're worried Google says [5:33] we've misinterpreted this research. It's [5:36] not actually all that bad, but you know, [5:37] of course they're going to say that. [5:40] Another thing that uh people are [5:41] concerned about is what you might call [5:42] appeasing the AI companies. Some [5:44] publications now have decided to [5:46] actually join up with AI companies like [5:47] The Atlantic here has joined up with [5:48] OpenAI to provide their journalistic [5:51] content to OpenAI to um as sort of a [5:54] premium content product. Um the writers [5:57] of the Atlantic said this is not [5:59] terrific. They're not actually diverting [6:00] any traffic to us. They're just sucking [6:02] up our content. We're not getting a lot [6:03] out of it. Complained to their [6:05] management. Their management said we're [6:06] profitable for the first time in a long [6:08] time. You know, get on board. [6:12] And one of the things we really worry [6:14] about uh is the idea of the loss of [6:17] humanity in journalism. Journalism, this [6:19] might sound sort of squishy, but [6:20] journalism is really an exchange of [6:21] information, insights, even emotion [6:24] between human beings. We take humans out [6:25] of the process really, you know, what [6:28] are we doing here anymore? And part of [6:29] the problem with that is that uh [6:31] research has found that readers trust [6:33] news even less when they begin to think [6:35] there's AI involved, right? Americans [6:38] trust and faith in in the news business [6:39] is is very very low. and this is only [6:42] diminishing it. So we worry about that. [6:45] All right. So I want to switch gears. [6:46] This was a period of horribles which I [6:47] think are justified. I think there's [6:48] plenty to worry about when it comes to [6:50] AI and journalism. But I want to switch [6:51] gears and talk a little bit about what I [6:52] see about some of the potential um [6:54] positive uses. So defeating idea block. [6:58] One of the uh hardest things about being [6:59] a journalist is coming up with story [7:00] ideas day in and day out. Uh on the left [7:03] we have this writer Jennifer Peters who [7:04] talks about idea block. She said, "I [7:06] sometimes feel like I've completely run [7:08] out of ideas. Not just story ideas, but [7:10] all original thoughts, right? You work [7:11] in journalism long enough, you just get [7:13] like there's nothing left to think of, [7:14] right? Like you've thought all the [7:15] thoughts. Um, and so how do you come up [7:17] with story ideas? So, one of the tools I [7:19] use is you go over to Perplexity. Folks [7:22] use Perplexity. Yeah. Okay. Go over to [7:24] Perplexity and say, you know, an editor [7:26] calls me up and says, "How about a story [7:27] on uh AI is used in government?" I hop [7:30] over Perplexity. I say, "It's a business [7:31] publication. Um, what are some story [7:33] ideas for how AI is being used in [7:35] government?" Not all are great, right? [7:36] Not all are going to be terrific. Um, [7:38] but I'm kind of intrigued. There's one [7:39] about federal contractors. Um, how the [7:41] relationship between government federal [7:42] contractors might change because of the [7:44] use of AI. And that's has potential, [7:46] right? I could sort of polish that up [7:47] and and sell that to an editor. [7:51] One of the tasks we're charged with as [7:53] journalists is to develop inst uh [7:55] expertise instantaneously. If in the [7:57] morning I know nothing about a topic, by [7:59] the end of the day, I'm supposed to [7:59] write on it coherently. So again using [8:02] perplexity this was an example where [8:04] there was a development in in antitrust [8:06] which I often write about where two uh [8:08] officials at the department of justice [8:09] were fired. It wasn't quite clear why [8:11] they had been fired. So I started you [8:13] know poking around on it thinking [8:14] there's a story there. One of my sources [8:16] said are you familiar with the Tony Act [8:17] of 1974. I had to admit I had no idea. [8:20] Still on the phone pull up Perplexi [8:23] asked Perplexi what the heck does the [8:25] Tony Act have to do with transparency um [8:27] at the Department of Justice? Before I [8:28] get off the phone I know the topic. pick [8:30] up the phone, call up my next source and [8:32] say, "Oh, how do you think the 10X is [8:33] going to play in here?" Right? Minute [8:35] and a half before I had no idea that [8:36] this um law even existed. [8:40] Um some other positive uses of AI tools [8:43] uh in journalism. We have on the left we [8:46] have Bloomberg Law that did some really [8:48] interesting research um to try to see [8:50] what the impact of a Supreme Court [8:51] ruling on voting rights uh was. It was [8:54] in 2021. What the impact was before and [8:56] after on voting rights cases in the [8:58] United States. 600 court cases they [9:00] wanted to an analyze. It's just too much [9:02] work for a team of reporters, even [9:03] dedicated reporters. So, they used a [9:05] combination of an LLM and human [9:07] reviewers to review these 600 cases. Put [9:09] together a pretty compelling story um [9:11] that uh found out, you know, they sort [9:13] of concluded there would have been a [9:14] plunge in the number of voting rights [9:16] cases. Um not work you could necessarily [9:18] do in a timely fashion without using AI. [9:20] On the right, we have the New York [9:21] Times. They used uh an AI tool to search [9:25] satellite footage from South Gaza, [9:27] looking for craters in the land that [9:29] were over 40 feet, which was what they [9:31] had figured out by talking to experts [9:32] was would be the mark of of large [9:34] munitions that Israel was using. Found [9:37] over 200 um of these craters um in the [9:40] satellite imagery, had reporters go back [9:41] over them, then wrote a story based on [9:43] that. Right. Israel is using large [9:45] munitions in Health Gaza, even though [9:46] they say they're not [9:49] transcribing Otter. any Otter users in [9:51] here? It's a lifesaver. Okay, terrific. [9:54] Because it's a great tool. Uh [9:55] transcription is is uh a lot of what [9:57] reporters do day in and day out. Um and [10:00] Otter is a tool that you can use to to [10:02] transcribe your calls transcribes them [10:04] instantaneously. You can click on a bit [10:07] of the transcription. It pulls up the [10:08] audio. That's pretty cool in and of [10:10] itself. Not too long ago, Otter World [10:13] Otter AI chat where you can actually [10:14] query those transcripts. So I can say at [10:17] some point I was talking to somebody [10:18] about topic X. They were kind of [10:20] comparing, you know, what had happened [10:21] in the United States, what would happen [10:22] in France. I don't remember the person. [10:23] I don't remember uh when we had that [10:25] conversation. And it'll it'll query all [10:27] of my transcripts and pull up that bit [10:29] of uh reporting, which is enormously [10:31] useful. On the right is similar. We're [10:33] seeing some innovation in the [10:34] transcription space. This is a product [10:35] called Minutes, which was developed by a [10:38] programmer and a reporter in a newsroom [10:40] in Michigan who said there's too many [10:42] government meetings happening across [10:44] Michigan. You know, there's uh city [10:45] council meetings, school meetings, all [10:47] that sort of thing. We have like five [10:48] reporters in our entire newsroom. [10:49] There's no way we can cover them. Let's [10:51] grab the video, transcribe it, and [10:53] automatically ping it for the topics [10:55] that we're interested in. They open it [10:56] up to other reporters. Now, folks can [10:59] sign up for any meeting that um Mace [11:01] covers, and it's 100 meetings. You can [11:03] say, "I'm particularly interested in [11:04] mentions of land use development." And [11:07] it'll uh email you and say, "At such and [11:09] such a time, city council member Smith [11:11] mentioned that topic, which is pretty [11:13] cool." [11:16] I'm going to skip the notebook LM is a [11:18] tool that um reporters are using more [11:19] and more. Basically, the idea is [11:20] developing your own brain. Upload your [11:22] own reporting, your own research to [11:24] notebook LM and then you can query in [11:26] the same way you might query, you know, [11:27] Perplexity or Cloud or or Chat GBT, but [11:30] you know the information there is [11:31] trusted. [11:35] Uh some couple other tools that [11:36] reporters are using. Grammarly folks I'm [11:38] sure people are familiar with Grammarly [11:39] using it for sort of basic copy editing [11:41] but also for suggestions on language [11:43] effectiveness that sort of thing. [11:45] Perplexity and claude can also do [11:47] something which is pretty neat which is [11:49] asking uh asking it to kind of help [11:53] suggest uh different ways of thinking [11:54] about a topic, different ways of writing [11:56] about a topic. So say I'm writing about [11:58] a new social network and I'm thinking [11:59] this you know the way the social network [12:01] is being adopted sort of is uh feels [12:03] like a social like a biological [12:05] contagion right it's sort of you know is [12:07] spreading in some places not other [12:08] places how can I describe the adoption [12:11] of this social network using those sort [12:13] of biological contagion terms and Claude [12:16] does a pretty good job of sort of [12:17] suggesting how I might write about that [12:18] so turning a story that could be like [12:19] pretty dry and suddenly I have a little [12:21] bit of fodder to work with obviously you [12:23] know this is raw material like I have to [12:26] do all the work of turning this into uh [12:28] you know actual good copy but at the end [12:31] of a long week uh it can be nice to have [12:32] a a boost in this way [12:37] uh borrowing a second pair of eyes this [12:39] again is using perplexity but obviously [12:40] other tools will be good at this too I'm [12:42] writing about topic and I'm thinking you [12:44] know we all come to this we all do have [12:45] our own biases our own experiences how [12:48] um might other people who are interested [12:50] who are going to be reading my story [12:51] think about the topic I'm writing about [12:53] so just for fun I asked a perplexed How [12:56] might a Trump supporter think about [12:58] making hot dogs the official food of the [13:00] United States? How about a liberal [13:01] Democrat? And how about a libertarian? [13:03] Right? Just to get a little perspective, [13:05] another second set of eyes on these [13:06] topics. I like the bit they talked about [13:08] how Democrats might actually like the [13:10] the connotation, the sort of cultural [13:12] connotation of the the tie between the [13:14] hot dog and the working class. They [13:15] might actually be, you know, into it and [13:17] sort of reasons that out. So, a little [13:20] bit helpful. [13:22] All right. Drafting the dreaded third [13:24] third. The bottom third of a lot of news [13:25] stories is stuff you've already written [13:28] five, six, seven times in the past. [13:30] You're tired. You don't want to write it [13:31] again. It drives me a little bit nuts [13:34] because you're expected to write it [13:35] again. You expect to sort of reddraft. [13:37] You know, this is example here is a [13:39] court case involving Google and Oracle. [13:41] I'd written on this court case like [13:42] seven times. This was on our staff at [13:44] Politico. Written on this court case [13:45] like seven times. There was a new ruling [13:47] in it. Have to write the third of the [13:49] story, which is the background of the [13:50] story. Ideally, you would have an AI [13:52] like a rag sort of uh based tool to say [13:55] what have I've already written on this [13:56] that you know my editor has approved, [13:58] it's been fact checked, it's been [13:59] published, we know it's good [14:00] information. How do you remix it and [14:01] just sort of write the bottom third [14:02] third for me and then I can um go over [14:04] it and edit it a bit. So, I'm hoping [14:07] hoping that [14:09] that becomes available. Uh other uses of [14:13] AI uh possible collaborative uses is [14:16] helping with the horrors that are [14:17] headlines and social media. You've [14:19] written the story, you're tired, your [14:21] editor comes to you and says, "Uh, what [14:23] are some suggestions for headlines in [14:24] social media?" You go and chat GPT. This [14:26] is a story I wrote recently. What's a [14:27] good headline for a magazine story on [14:29] the mainstreaming of betting on [14:30] politics? And it offers me some pretty, [14:32] you know, decent suggestions. Some [14:33] aren't great. Um, but I like the odds on [14:35] the Oval Office. Like, that has [14:37] potential. I could work with that, [14:38] right? Uh, same with social media, [14:40] right? You're expected now to to post uh [14:41] your stories on Twitter, LinkedIn, uh, [14:44] all over the place. and um go into [14:47] claude it suggest uh potential posts uh [14:50] for those different mediums customize [14:52] for the nuances of those particular [14:54] platforms and I can say okay this is a [14:55] little bit too serious we're trying to [14:57] you know have a little bit of fun with [14:58] this can you make them funny right and [15:00] suggest funnier posts said can you [15:02] appeal to millennials like I don't know [15:03] how millennials talk can you make it [15:05] appeal to millennials and it helps me [15:06] set the tone that way so that can be [15:08] useful um repackaging uh our content for [15:11] wider publication um very often that's [15:15] uh in the form of newsletters. We're [15:17] expected to distribute our uh journalism [15:20] in uh newsletters, other forums, and you [15:22] can just uh ask AI to say, can you [15:25] distill this down in a way that might be [15:26] useful for a newsletter? Can you go [15:27] through the website this today and pick [15:29] out everything? In this case, the [15:30] Telegraph has a healthy living [15:31] newsletter that's AI generated. It just [15:33] sort of scans the Telegraph website, [15:35] pulls the stories that are related to [15:36] healthy living, packages them, sends [15:38] them out. Uh translation is enormously [15:42] useful. obviously reach reach a brighter [15:44] audience wider audience if you're not [15:45] just publishing one language. New York [15:46] Times uses AI for um they have a Spanish [15:48] version and a simplified Chinese [15:50] version. Uh the AP also uses um AI power [15:54] translation. They also uh it's [15:56] interesting they they make a point of [15:57] adding a disclaimer. Not everybody does [15:59] to say this was uh produced by AI and [16:03] reviewed by a human reviewer. [16:05] Um Lemon French publication they were [16:07] going to they said let's do an English [16:08] language newsletter. When they realized [16:10] that with AI they could simply convert [16:12] their entire website pretty much to [16:13] English, they decided to just launch a [16:15] complete English language version of [16:16] Leond. Obviously opening up their [16:18] journalism to a much wider audience. [16:20] Washington Post is doing something in a [16:22] similar vein. Uh doing audio versions [16:24] very often now you go on websites [16:25] there's a new story and there's an audio [16:26] version that's AI generated. You click [16:28] on little headphones and it says this is [16:29] a story by uh Andrea Gimenez I think [16:32] wrote that and uh being read to you by [16:33] an AI generated voice. [16:37] Uh we're going to skip over that a [16:39] little bit. [16:41] Uh the case study New York Times often [16:43] does uh very well with these new [16:44] technologies come out. They devot devote [16:46] a lot of their their considerable uh [16:48] resources to studying them and figuring [16:50] out how to best apply them. So it's [16:51] often useful to look at how they're [16:52] thinking about these things. Um the [16:55] interesting bit here is they decided a [16:57] few years ago that um Open AI and [16:59] Microsoft were actually uh [17:01] misappropriating uh Times content [17:03] feeding into their AI tools. they [17:04] weren't being compensated and so they [17:06] sued them, right? Which was seemed like [17:07] as a pretty aggressive move. Very [17:09] recently, uh they uh partnered up with [17:12] Amazon and said uh you can take our [17:14] content, feed it into Amazon's AI [17:16] products. And people said, "Well, what [17:17] changed?" And they said, "Well, we're [17:18] getting paid." Like that's the whole [17:20] point of this uh is to be compensated [17:22] for our work. Amazon's willing to [17:23] compensate us. Um so this is this is a [17:26] smart deal. Um The Times is also very [17:29] explicit about how they use um AI [17:31] internally, which I think is pretty [17:33] interesting. They talk about here how [17:35] they use it uh to analyze data. We [17:36] talked about the the identifying mount [17:38] creators to write headlines and [17:39] summaries uh to help uh generate [17:41] translations, audio versions, a lot of [17:43] the applications that we talked about. [17:45] But I they make this very interesting [17:46] point. They say everything that the [17:49] Times publishes though none of it's [17:51] created in whole cloth by AI and it's [17:53] the responsibility excuse me [17:55] responsibility of Times journalists [17:58] right so we're using these tools we're [17:59] using them effectively we're using them [18:01] to boost uh boost um our efforts but at [18:04] the end of the day we're responsible for [18:05] this stuff all right if you I think we [18:09] are pretty out of time uh if you're [18:11] interested in the topic I recommend this [18:12] report it's from the Columbia Journalism [18:14] Review and a few other folks uh 46 pages [18:16] pages uh full of good stuff on the uh [18:19] intersection of AI and journalism.