We want your feedback: How can writers use AI to tell human stories?

[Tony Stubblebine]

How we are navigating the role AI does and doesn’t play in our mission to deepen understanding.​

Image by Jason Combs

Here’s a conundrum we’ve been wrestling with here at Medium: The vast majority of AI-generated writing is bad and lazy slop, we can all agree. At the same time, it’s pointless to pretend there aren’t good use cases for AI tools, even in the writing world. So how can we square that while building Medium to be the best place to read and write human stories on the internet?

We want to start a conversation with our community about how we can use AI to help writers tell human stories and where we should place boundaries.

As you read through this, think about how these principles feel to you, if there’s anything that could be useful, or if there’s anything we’re missing.

Below are four sections:

1. How our position on AI has changed in the past two years
2. Our three principles
3. How we are applying our principles to support our mission, our writers, and our readers
4. Specific examples of how we have used AI so far

A quick disclaimer before you read any further: There is no world in which Medium is meant to be a home for a story generated by some large language model with a single click. There is also nothing inevitable about the way we use AI, which is why your feedback matters so much to us right now.

How our position on AI has changed in the past two years​

Our mission and our business model at Medium are the same: to deepen understanding.

For writers to succeed on Medium, for readers to enjoy Medium stories, and for Medium to continue, this requires real stories — human stories, written from human experience and with human wisdom.

So then what is the value of artificial intelligence, and what are the pitfalls to avoid? Is there a way to use AI to deepen understanding, or help writers tell their human stories? Two years ago, we thought the value to writers and readers was less than zero. The AI companies had leached value from your writing without offering consent, credit, or compensation. Then they enabled a wave of spam that tried to replace your writing with hallucinated slop.

As we’ve continued working on making Medium the best place to read and write, we’ve noticed and heard from our readers and writers that some use of AI is starting to be useful. We also are finding uses in our own work. In our recent surveys, more than half of Medium readers are using AI tools in some capacity.

AI is a tool that exists in the world. We want to make sure that we explore all possible options to make Medium better for readers and writers. Here’s how we’re starting to do so with AI, both in the principles that we take into consideration and how we’re currently using AI at Medium already.

Principles that support our mission, our writers, and our readers​

Medium is Humans First. We’ve found that leads to three principles that help us make decisions about artificial intelligence.

1. Our mission to deepen understanding requires human stories that contain human thoughts, human emotion, and human experience.
2. For our writers, we will protect the incentives that make it worthwhile for them to write and share their stories.
3. For our readers, we will strive to give them agency to examine, influence, avoid, and override AI.

How we apply these principles​

In every decision we make to build and improve Medium, including those that involve AI, we try to maximize one or more of our three principles to prioritize human stories, incentives for writers, and agency for readers.

Here is how we currently apply those principles.

Human stories first

• We assess stories and build systems that elevate human stories with human experience and ideas. We have not found any foolproof way to programmatically determine AI-generated stories across all of Medium, so we focus on our primary goal of elevating stories about human ideas and experience instead of relying on an imperfect AI percentage score.
• To empower more people to share their stories, we treat AI-assisted writing differently from AI-generated writing. AI translations, spell-checks, and note-taking help people tell their stories. But the spectrum veers into a gray area as AI contributes outlines, research, and even partial sections. So we again focus on the primary goal of elevating human stories.
• We strive to add human review as a signal within our AI recommendations. The Medium home feed and daily digest emails use AI to match the millions of stories on Medium with the individual interests of each reader. For the most part, readers find this algorithmic approach preferable to chronological following feeds. However, the downside of most of these systems is that they optimize for attention (clickbait) rather than for understanding. We use human review through the Boost, Featuring, and internal curation to elevate human experience over attention seeking. Our quality guidelines explain the many ways that we weight stories that contain human experience.
• We acknowledge the tangle of problems with AI-generated images, but currently allow them. Truthfully, we’re a reading and writing company, so we focus on how AI could help or harm writing. Our readers find that AI images can be an improvement on a writer’s ability to communicate, but we know they exist because AI was trained on images created by people who received no compensation. We ban AI-generated images that veer into misinformation, and we look forward to supporting any initiatives in the artistic industry that apply pressure to solve the compensation question.

Incentives for writers

• We advocate for compensation and credit from the AI companies that trained on their writing. That has included blocking many AI companies and will include participating in emerging industry-wide initiatives to make AI companies pay creators. CloudFlare just released one option. We have also been contributing to an open standard about how to allow content owners to define rights and restrictions when it comes to AI training on their writing that we expect to launch shortly.
• We limit the distribution of AI-generated content. Our recommendation system will only present AI-generated content to readers who have chosen to follow that writer, or publication if the writer publishes in one.
• We continue to say no to AI-generated content in our partner program. Even if ChatGPT could generate a perfectly valuable story, we still want our partner program to incentivize human storytellers, and ensure Medium remains a platform where human creativity is valued, protected, and rewarded.

Agency for readers

• Readers can use our AI-powered story recommendations or not. While we make recommendations for what we think a reader will most enjoy, we also give them the tools to find Medium stories their own way. We offer a reverse chronological follow feed, direct email subscriptions from writers, options to mute publications and accounts, and a “show less like this” signal to our recommendation algorithm. We think these principles of agency are critical whenever building AI tools so that the tools elevate rather than diminish the human experience.

How we have used AI so far​

This is a list of the ways we currently use AI tools within the Medium reader and writer experience — both launched, in-development, and cancelled experiments — and how those uses line up with our mission and principles. (We will never shoehorn AI in just because it’s trendy — if we decide to use it, it’s because it serves a purpose that aligns with our three principles.)

• Our recommendation system uses algorithmic machine learning to match-make between what writers publish and what readers want to read. This is our primary use of AI and predates any of the current AI craze. As covered above, we put human signal into the recommendations to prioritize human stories, and give readers agency to adjust, override, and ignore the results.
• We recently ran an experiment with AI-generated paywall text. The paywall text was generated based on the story the writer chose to paywall, making the paywall more specific to the reader’s goals to convert more subscribers and increase the incentive for the people on Medium who write in our partner program.
• We recently ran an experiment with AI-generated SEO meta tags. Writers can set their own tags to optimize traffic from Google, but often don’t take advantage of this feature. Providing these could lead to increased traffic. File this under writer incentive.
• We have prototyped AI summaries of large groups or stories to use on topic pages, or to help people keep up with news within a topic. The key principles still hold: these summaries need to be based on human stories, give readers agency to ignore or correct them, and improve the incentive for writers to write by providing credit and compensation.
• We have prototyped AI tools to search, summarize, and otherwise take on the tedious parts of writing. For example, all of the links in this piece were added by an AI assistant by asking, “Find links from the Medium blog to these concepts and add them.”
• We have used AI to prototype throwaway apps in order to demonstrate an idea. This once went as far as spinning up an entire separate blogging network. That prototype took six hours from start to launch. This helps us move faster on product launches.
• Our engineers typically use AI coding tools such as Cursor and report a modest productivity gain. This also helps us build bigger and more ambitiously.
• We factor the ethics of AI companies when choosing AI tools for internal use. This is the ultimate humans first point: We want to live in a better world. This includes their position towards writers in terms of consent, credit, and compensation, their efforts to combat bias, and their environmental impact. Medium already invests in climate offsets when our company travels. We expect to do the same for AI usage.

In short, we are already using AI to make Medium better for writers and readers. We believe there are other ways AI could help us improve Medium. Let us know: When thinking about how AI can improve Medium, what is most valuable to you?

If you want to share your thoughts, please feel free to publish a story on Medium using the #AiForHumans topic. We’ll read every one, in addition to the responses here.

---

We want your feedback: How can writers use AI to tell human stories? was originally published in The Medium Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

Continue reading...

 

Join 𝕋𝕄𝕋 on Telegram
Channel PREVIEW: