Rule of five

One way to iterate more quickly when using AI tools is the rule of five (or ten, or thirty, or ninety). Instead of asking for one headline at a time, ask ChatGPT for five headlines.

Along with getting more ideas set out in front of you, I’ve also found it to be helpful process for highlighting bias. I don’t (necessarily) mean what we normally think of as bias - think of it more like “habits” or “tendencies.” Maybe one person always writes sensationalist headlines, another person always uses colons and subheads.

Exercise

We’re going to start from the same headline generation prompt as before, but change it to be for five headlines instead of just one.

A prompt suggestion

You are an editor at the New York Times. Write five headlines for the article below. The tone should be informed but approachable. Limit the headline to 70 characters.

ARTICLE:

(paste the story in)

Use my article 🔮

Clear the chat 🔮

This is simple, this is easy. And nothing’s stopping us from asking for ten headlines or twenty headlines!

Noticing habits

The five responses I got were pretty much all in the same vein:

  • Bronx Drug Market Thrives as City Battles Overdose Crisis
  • Fentanyl Fear Rises in NYC Amid Record Overdose Deaths
  • In Bronx Underpass, Open Drug Use and Rising Tensions
  • Bronx Tragedy: Fentanyl’s Shadow Over a Community in Crisis
  • Amidst Drug Crisis, Bronx Residents Seek Safer Streets

When you ask for five headlines instead of just one, the similarities really jump out at you! We’re being gently guided into a certain direction, even though the LLM doesn’t actually have a reason to send us in one direction or another (it can’t think, after all!).

I think calling this “bias” is a bit much, but it’s a gentle reminder that AI tools come with baked-in assumptions and angles. If you don’t work hard to notice them, at the very least they’ll make all your content similar and boring!

Notice how even though I asked ChatGPT to not use subheads or colons, it did anyway! You should expect the AI to ignore your instructions a good portion of the time.

One way to add a little diversity to the mix is to simply ask for it.

Diversifying your results

An extension of the simple “give us five” from above might be to explicitly ask ChatGPT for five markedly different options – in this case it might be five different angles or five different tones.

Even if we aren’t making specific demands of the AI tool on how to break out of its habits, asking for different takes can go a long way in helping it happen.

A prompt suggestion

You are an editor at the New York Times. Write five headlines for the story below. The tone should be informed but approachable. Limit the headline to 70 characters.

Have each headline elicit a different emotional response or subject-matter concern from the reader. Label each headline with the emotion/concern.

ARTICLE:

(paste the story in)

Use my article 🔮

Clear the chat 🔮

Reflection

Let’s take a look at the results I got:

  • Concern: Public Safety: “Bronx Daycare Tragedy Highlights Rampant Drug Crisis in NYC”
  • Emotion: Outrage: “Fentanyl Scourge Unchecked: Bronx Community Demands Action”
  • Emotion: Compassion: “Amid Bronx Despair, a Plea for Help Against Drug Menace”
  • Concern: Policy and Governance: “NYC Struggles with Drug Policy as Overdoses Soar, Communities Suffer”
  • Emotion: Fear: “Bronx Nightmare: Fentanyl Threat Looms as Overdoses Rise”

When you’re provided suggestions by the AI, it’s easy to just go along with them. But when you starting seeing the habits and tendencies and start asking for different angles, I feel like that’s when you start to realize “hey, we’re giving up a lot of control here!”

If the media has a certain way of talking about certain topics – which media critics like Alec Karakatsanis would very much agree – it takes a certain degree effort to break out of the typical framing.

Using an AI-generated headline can go both ways: since language models are just statistical predictions of the next word, they go above and beyond when it comes to reinforcing what’s “typical,” or the status quo of how a subject is discussed. But conversely, if we recognize that other angles might exist, it’s easy enough to prompt in directions we wouldn’t normally think about.

In our next step we go even further in having the AI help us generate more headlines and think critically about our angles.