Journalism in the age of AI (with Dylan Matthews)

Oct 4, 2023 Episode Page ↗
Overview

Spencer Greenberg and Dylan Matthews discuss the impact of AI on journalism, the print media industry, and the future of labor. They explore how LLMs challenge traditional journalism, the evolution of media business models, and novel approaches to distributing AI's economic surplus.

At a Glance
14 Insights
1h 14m Duration
10 Topics
5 Concepts

Deep Dive Analysis

AI's Impact on Journalism and Job Automation

Simulating Writing with Large Language Models

Dynamics of Digital Media Market and Business Models

Historical Evolution of Truth-Seeking in US News Media

Distributing Economic Surplus from Advanced AI

AI Risks and Novel Governance Structures

Audience Questions: Subscription Models and Polarization

Audience Questions: Collective Action Against AI Automation

Audience Questions: Government Funding for Journalism

The Future of Journalism: Hobbyist vs. Professional Models

Windfall Clause

A proposed self-imposed tax bracket system for AI companies, where profits exceeding a very high absolute threshold (e.g., 0.1% of world GDP) are donated to specified charitable destinations. This aims to proactively address the distribution of potentially massive economic surplus generated by advanced AI.

Audience Capture

A phenomenon where content creators, particularly those relying on subscription or direct supporter revenue, become incentivized to reinforce their audience's existing beliefs rather than challenging them. This behavior is driven by the desire to avoid offending subscribers and negatively impacting their income.

Moneyball Approach

Applying rigorous analytical and data-driven methods to fields traditionally dominated by intuition or gut instinct, often leading to the discovery and exploitation of market inefficiencies. This approach was exemplified by Blumhouse's strategy in the horror film industry, focusing on low-budget, high-margin productions.

Classified Ad Model

A historical business model for newspapers that relied on their monopoly over local small-scale transactions, such as hiring or selling furniture, through paid classified advertisements. This model reliably generated high profit margins (20-30%, sometimes 60%) for decades until disrupted by free online services like Craigslist.

Non-Partisan Media Ideal

A relatively recent development in US journalism (emerging in the 1940s-1950s) that established a strict separation between a newspaper's editorial board (expressing opinions) and its news section (reporting facts). This ideal aimed for objectivity and independence from political parties, contrasting with earlier partisan newspaper models.

?
Are Large Language Models (LLMs) coming for journalists' jobs?

LLMs are making substantial progress in explanatory and analytical journalism, particularly in synthesizing online information, but deep investigative work requiring trust-building and finesse remains harder to automate.

?
How does LLM-generated writing compare to human writing?

LLMs are good synthesizers of information and can produce 'fine' prose, but often lack the specific nuance, style, or perspective a human writer would convey, making extensive editing necessary for personalized content.

?
How did Craigslist impact the news media industry?

Craigslist destroyed the newspaper industry's highly profitable classified ad model by offering a free online alternative, eliminating a major source of revenue that had historically provided 20-30% profit margins.

?
Has US news media become less truth-seeking over time?

Historically, US newspapers were often closely tied to political parties, with a period of strong truth-seeking norms (roughly 1940s-1970s) being an 'aberration' driven by market forces like the classified ad model. Current market incentives are shifting back towards partisan alignment.

?
What are some models for distributing economic surplus from AI companies?

OpenAI uses a capped profit system where investors get a 100x return, with anything above that going to a non-profit. The 'Windfall Clause' proposes a self-imposed tax bracket for companies exceeding a certain revenue threshold, donating excess profits to charity.

?
Why is collective action, like labor organizing, so difficult?

Collective action is inherently challenging because it requires a group of people to acknowledge a shared interest and take costly action, and historically, massive surges in organizing have only occurred during periods of extreme economic calamity.

?
How does a subscription model alter the incentives for journalists?

Subscription models can lead to 'audience capture,' where journalists are incentivized to reinforce their audience's existing beliefs to maintain subscriber numbers, potentially reducing the impact of introducing new or challenging arguments.

?
Can government funding for journalism work without creating perverse incentives?

While examples like the BBC, CBC, and NPR show that reasonably independent government-funded journalism is possible, its success depends heavily on a government's level of corruption and public trust. Voucher systems, while increasing media landscape, could also increase polarization.

?
What is the future of journalism if traditional business models collapse?

Journalism may evolve into an 'artisan trade' or hobbyist pursuit, with people writing for free or on the side of other jobs, and the definition of 'journalism' expanding to include activities like podcasting that collate information and ask relevant questions.

1. Invest in Future-Proof Skills

Evaluate current skills and strategically invest in developing new ones that are less susceptible to AI automation, as the value of traditional skills may shift rapidly.

2. Experiment with AI Governance

AI companies should actively experiment with novel governance structures to better align their incentives with long-term humanity and avoid potential negative outcomes from traditional corporate models.

3. Implement AI Windfall Clauses

AI companies should consider implementing self-imposed “windfall clauses” to donate a set percentage of profits exceeding a high threshold to charitable causes, distributing economic surplus to society.

4. Diversify AI Governance

When establishing governance structures for powerful AI companies, ensure a diverse intellectual culture beyond a single group (e.g., EAs) to bring varied premises and perspectives to problem-solving.

5. Study Regulatory History for AI

To inform AI policy and regulation, research how past regulatory regimes (e.g., financial, nuclear, FDA) developed, as these historical analogs may offer productive insights despite AI’s unique challenges.

6. Enhance Writing with AI

Leverage Large Language Models (LLMs) to refine your writing by asking them to rewrite your text in the style of a skilled author, which can often lead to improved quality.

7. Direct AI with Roles

When using LLMs, explicitly instruct them to adopt a specific role (e.g., “pretend you’re a really good journalist”) to guide their output and achieve more targeted, higher-quality results.

8. AI for Idea Generation

Employ Large Language Models to overcome writer’s block or generate new ideas by prompting them with your interests, potentially leading to novel concepts and pathways for exploration.

9. Exploit Market Inefficiencies

Seek out underserved or undervalued niches in the market, like Blumhouse did with low-budget horror films, to achieve high profit margins by catering to a devoted fan base.

10. Diversify Digital Media

In digital media, avoid over-investing in a single strategy or platform, and instead learn to adapt and pivot between emerging opportunities to ensure long-term survival.

11. A/B Test Headlines Rigorously

For online content, generate numerous headlines (e.g., 20) for each post, then rigorously A/B test them to optimize for engagement and click-through, as demonstrated by Upworthy’s early success.

12. Use ThoughtSaver for Recall

To strengthen your recall of important ideas and combat forgetting, use ThoughtSaver, a free tool that emails daily flashcard quizzes on topics you care about.

13. Optimize Podcast Speed

Gradually increase your podcast listening speed until it feels slightly uncomfortable, as this can help you consume content more efficiently.

14. Subscribe to Newsletter

Subscribe to the “One Helpful Idea” email newsletter for a weekly valuable idea in 30 seconds, plus new podcast episodes, essays, and event announcements.

There's a degree to which I worry about a breakdown of the sort of like, all these articles are eventually written by by LLMs, and then that LLMs are citing them, and you have sort of a recursive cycle of garbage.

Dylan Matthews

But I think one of our ideas in starting Vox was that there's a lot of topics in the news where sort of mainstream coverage of sort of day-to-day updates assumes a lot of background knowledge that a lot of people don't have.

Dylan Matthews

My suspicion is that it depends a lot on the power differential between the groups. I mean, I'm certainly not an expert in historical cases, but it seems like when a much more technologically powerful civilization meets one that's less technologically advanced, it generally goes really badly for the one that's less advanced.

Spencer Greenberg

The marginal impact of what I'm doing if I'm just reinforcing things that my readers already believe is basically zero. The impact if I'm introducing them to arguments that they might not have been introduced to otherwise is potentially significant.

Dylan Matthews

Collective action is just really, really hard. I think the Scott Alexander post on Moloch is maybe the most evocative illustration of this.

Dylan Matthews

The point is that you, you can dip in and out and be doing other things and still pick up stuff.

Dylan Matthews

It's hard to get a group of people together to acknowledge a shared interest and have them take costly action that is more than just passively posting something on behalf of any cause.

Dylan Matthews

Upworthy's Headline Optimization Protocol

Dylan Matthews
  1. Each person writing a post had to come up with 20 different headlines for it.
  2. Staff would discuss among themselves to narrow down to a number of headlines suitable for A/B testing.
  3. Ruthlessly A/B test each of the selected headlines to optimize for engagement.

AI Company Windfall Clause Protocol (Proposed)

Dylan Matthews
  1. An AI company sets a threshold of revenue or profits (e.g., 0.1% of world GDP) that, if exceeded, triggers a self-imposed tax.
  2. A specified percentage of profits (e.g., 10% initially, then 20%) in excess of this threshold is donated to a given charitable destination.
100x
Capped profit return for OpenAI investors Any return above this amount is directed to the non-profit entity that owns the for-profit.
0.1%
Proposed threshold of world GDP for Windfall Clause A company's revenue or profits exceeding this level would trigger a self-imposed charitable tax, as per 'The Windfall Clause' paper.
10% then 20%
Proposed initial and subsequent charitable donation percentages for Windfall Clause These percentages would apply to profits in excess of the defined threshold.
20 to 30%
Average profit margins for newspapers during the classified ad era Some newspapers achieved margins as high as 60% during this period.
2013
Year of a widely cited Oxford article predicting job automation The article argued a significant percentage of jobs would be automatable in the next 10 years, though a 'trivial percentage' were actually automated.
2013
Year Upworthy was 'wildly successful' due to Facebook algorithms This success was short-lived as Facebook later changed its algorithm.
20
Number of headlines Upworthy required per post for A/B testing This was part of their ruthless optimization approach.
about a third
Percentage of Americans working on farms in 1900 Illustrates a historical shift in labor, with current figures much lower.
1.5 to 2%
Current percentage of Americans working on farms A dramatic reduction from 1900, despite no Luddite-like movements against farm automation.