Lucid Lens surfaces the content beneath the headline

Alek Chakroff, Justin Stimatze, Natasha Jensen

People don’t always read past the headlines – but headlines don’t tell the whole story. Here we introduce the first feature in AOI’s Lucid Lens project: a proof-of-concept tool that automatically rewrites news headlines to reflect the content of articles in a more accurate, less sensational way.

The Challenge

On Nov 26, 2021, the New York Times published an article with the headline: “After Murders ‘Doubled Overnight,’ the N.Y.P.D. is Solving Fewer Cases.” A reader could be forgiven for concluding that murders in New York City had doubled overnight. In fact this was not the case, and was indeed never claimed in the content of the article. A correction later noted that: “detectives’ caseloads doubled overnight, not murders“ [emphasis added]. This came alongside complaints on social media from discerning readers. The editorial move to imply murders had doubled in America’s largest city turned a complicated story into a simple, sensational, terrifying one – that is, if you only read the headline.

People often share articles after reading only the headline.1 What drives sharing is largely emotional content, especially negative emotion. As the journalistic saying goes, “When it bleeds, it leads.” This incentivizes users to share content that they think will be engaged with. Journalists may internalize these incentives (even without knowing it) and may craft headlines that are more sensational, misleading, or essentially “clickbait.” Headlines are powerful enough to shape readers’ understanding of articles even when they read the articles.

The Opportunity

Although people often share misleading and sensational content, people also desire accuracy. The Reddit forum “Saved You a Click'' has 1.9 million subscribers and is devoted to crowdsourcing accurate summaries of the content behind headlines. As the About section says: “Don't click on that, we already did. Fighting clickbait for better journalism.” 

We propose an automated, scalable solution to the same problem.

Previous features that suggest or rewrite headlines have relied on crowdsourcing, or hard-coded string replacement. Thanks to developments in AI, particularly large language models like GPT, we can now automatically read and summarize an article and generate a new headline that reflects its contents, plain and simple. Notably, the news app Artifact uses a similar approach to rewriting headlines. Lucid Lens is differentiated on multiple counts. Artifact is geared towards improving clickbait, while Lucid Lens addresses clickbait as well as misleading and sensational headlines. While Artifact requires users to use their app, Lucid Lens integrates with the user’s browser, and doesn’t require any change in browsing habits.

Fig 1. Lucid Lens in action: Front page with headlines before and after text replacement

As a demonstration, we developed a Chrome extension that will replace headlines on the front pages of a set of popular news sites. When the front page is rendered in the browser, the extension will follow each headline address and send the article contents to OpenAI’s GPT-3  for processing. We developed a prompt2 that returns a summary of the body, which is then used to re-write the headline.

One demonstration (second row in Fig 1) replaced the headline:

“Many Americans still think one false claim is ‘probably’ True.”

 with

“Kaiser survey finds many Americans believe misinformation on guns, Covid-19, and reproductive health.”

The rewritten headline is more informative, and accurately reflects the contents of the article. 

Lucid Lens can help increase the quality of information consumed online. Without interrupting their natural browsing patterns, people will interact with cleaner headlines scrubbed of sensational or misleading information. This could reduce the sharing of misinformation. Further, assuming that rewritten headlines are less sensational, and less emotional, this tool could lower stress associated with “doom scrolling” the news. No, murder rates did not double overnight. 

What’s Next

The following are key areas where we are excited to build on Lucid Lens.

A primary concern is privacy. If Lucid Lens relies on API calls to OpenAI for rewriting headlines, it risks surfacing a user’s browsing behavior. This process is also tedious, requiring separate calls to the API for each article linked on a news page. Running an LLM on the user’s local machine could help with both privacy and efficiency.

The current demonstration of Lucid Lens is limited in the number of news sites it can process – mainly due to difficulties in identifying the original headlines on the page. News sites and aggregators use a variety of markups for headlines, and sometimes obfuscate CSS as a consequence of compilation or minification. Relatedly, it can be difficult to identify the body of an article on a news site, which could lead to faulty rewritten headlines. We plan to support a larger diversity of news sites by handling differences in markups that can prevent automatic scraping and headline detection.

While headlines are crucial, Lucid Lens can’t help if the whole article is misleading or sensational, since it only summarizes what’s in the content. Further work could mitigate how misleading these sensational articles can be, by presenting articles in the context of other, more straightforward discussion – for example, by aggregating articles by topic or by the event they cover. One could see a summary headline for “Murder rates in NYC,” with more representative coverage. 

As with any LLM, there is a risk that Lucid Lens will deliver inaccurate AI-generated “hallucinations” to the user. In the task of headline summarization, there may be less room for error or hallucination, as compared to other risky cases like drafting legal briefs. Still, we emphasize that even if the tool makes browsing the news easier and more informative, there is no substitute for reading the article.

Lucid Lens serves one broad goal of the AI Objectives Institute (AOI): to help people understand the ways they are being manipulated online and provide tools to help them navigate the connected world more safely and responsibly. Other ongoing research at AOI explores how to detect and classify manipulative content, persuasion tactics, and bias. 

Early testing of Lucid Lens is compelling. Its rewritten headlines are more accurate and less sensational. We look forward to launching into research at a larger scale to quantify these effects. We can then investigate deeper questions such as: does Lucid Lens reduce engagement with misinformation? Do users have a more accurate takeaway from the news? Do users find their time better spent? We have tested internally and we love how it’s made reading the news nicer, but different individuals have different interests and different needs. We really want to know what YOU think! Please join the experiment by trying the Lucid Lens extension yourself. The project is open sourced in the Lucid Lens project repository on github. We welcome all feedback!

Notes

  1. In fact, research found that most links are shared by people who have not clicked into the content. ↩︎

  2. The current prompt: “Be as concise as possible. Give me a single new objective, neutral, factual headline for the article content below.” ↩︎

A correction was made on September 12, 2023: An earlier version of this post did not include to include information about Artifact’s Intelligent News app and its headline rewriting feature.

A correction was made on October 5, 2023: An earlier version of this post did not include Natasha Jensen as a contributor to the Lucid Lens project.

Previous
Previous

Talk to the City: an open-source AI tool for scaling deliberation

Next
Next

Roadmap for a collaborative prototype of an Open Agency Architecture