Ole Reissmann

About · Newsletter

THEFUTURE

You’re Holding It Wrong: The ChatGPT Edition

Newsletter sent 2.9.2025 by oler

In this issue: When ChatGPT fails at basic geography, who’s really holding it wrong? The Wall Street Journal’s Tess Jeffers on lightweight AI experiments and changing audience expectations. Plus: When AI tools promise to fix your prompts.

What we’re talking about: Is the new ChatGPT smarter now? Over the past few days, there have been examples that seemingly prove the opposite. People have asked it to create maps of Europe or Germany with poor results. Someone wanted to know exactly when Cisco introduced the C1101-4P router.

On one hand: yeah, really dumb. On the other hand: that’s just not how this works.

This reminds me of the iPhone 4. Fifteen years ago, Apple released an iPhone that didn’t play particularly well with human hands – reception got worse when you gripped it and unknowingly blocked the antenna in the bottom left corner. Apple’s response back then was basically: You’re holding it wrong.

When you spend some time with large language models, you eventually learn: this is lossy compression. With emphasis on: lossy. A Blurry JPEG of the Web. Details get blurred, and it’s almost an art to hit the sweet spot – where are the details still sharp and accurate, where do the models start to hallucinate?

And it’s pretty brazen, of course, that platforms put up chatbots and say: universal tools, mega-smart. And then largely leave users on their own, and when they rely on the output, it’s like: well, you didn’t read the fine print, and the AI fanboys laugh at you.

(When you tell GPT-5 not to paint a picture but to create code, it can be just as disappointing.)

What else I’ve been reading:

AI & Journalism Links

Users of German news site Süddeutsche who were shown a difficult quiz about AI-generated images afterwards trusted media less and visited the news site a little bit more often – and now everyone’s hoping that quality journalism still has a chance. (Sarah Scire, Nieman Journalism Lab)

What happens to carefully crafted journalism when readers expect AI-generated, personalized stories created instantly? Semafor’s Gina Chua on how AI will upend the news.

“It’s the return of the shoe-leather reporter, empowered by an AI partner. One gathers trust, information and relationships. The other who assembles.” (David Cohn, LinkedIn)

A short clip of a Will Smith concert looks like a crappy AI fake – but it’s not: “The crowds were real, but the videos were manipulated: first by Will Smith’s team, and then without asking, by YouTube.” (Andy Baio, waxy.org)

From Star Wars insult to TikTok meme: “Clanker has become a go-to slur against A.I. on social media, led by Gen Z and Gen Alpha posters.” (Eli Tan, New York Times)

Werewolf leaderboard: GPT-5 is the best at bluffing and manipulating the other AIs in Werewolf. (Foaster Labs)

Low-investment, high-stakes: While newsrooms scramble to keep up with AI, Tess Jeffers is taking a different approach at The Wall Street Journal. Her take: lightweight experiments that prepare for a future where audiences need something different from journalism.

Three Questions with Tess Jeffers

Previously: Nieman Lab’s Andrew Deck, INMA’s Sonali Verma, Newsroom Robots’ Nikita Roy, and Republik’s David Bauer.

Hands on: Role, task, context, example – that’s how you write a good prompt. In workshops these days, I like to tell people: just ask the AI itself what would make the prompt better. And on top of that: OpenAI and Anthropic offer specialized tools for optimization.

Just a quick little example. Before:

You are an experienced editor. Read the following text carefully: What is the single most impactful change that would significantly improve this text's clarity, engagement, or usefulness for its intended audience?

With OpenAI Prompt Optimizer for GPT-5:

# Role and Objective
- You are an experienced editor tasked with evaluating and enhancing written content.
# Instructions
- Carefully read the provided text.
- Begin with a concise checklist (3-7 bullets) of your evaluation steps; keep items at a conceptual level, not implementation details.
- Identify the single most impactful change that would significantly improve the text's clarity, engagement, or usefulness for its intended audience.
- Clearly describe this change and explain its expected impact.
# Output Format
- Return only your recommendation, formatted as a brief summary statement.
# Verbosity
- Keep responses concise and focused on actionable editorial improvements.
# Stop Conditions
- Complete the task once you have suggested and explained the single most impactful change.

With Anthropic Console for all recent Claude models:

You are an experienced editor tasked with improving a piece of text. Your goal is to identify the single most impactful change that would significantly improve the text's clarity, engagement, or usefulness for its intended audience.

Here is the text you need to analyze:

<text_to_analyze>
{{TEXT}}
</text_to_analyze>

Please follow these steps:

1. Carefully read and analyze the provided text.

2. Consider various aspects of the text, including but not limited to:
   - Clarity of ideas and language
   - Engagement factor for the intended audience
   - Overall usefulness of the information presented
   - Structure and organization
   - Tone and style

3. Identify the single most impactful change that would significantly improve the text. This change should have the greatest positive effect on the text's clarity, engagement, or usefulness for its intended audience.

4. Before providing your final answer, use the <scratchpad> tags to think through your reasoning. Consider multiple potential changes and why you believe your chosen change would be the most impactful.

5. In your final response, include:
   a) A justification for why you believe this change would be the most impactful. Explain how it would improve the text's clarity, engagement, or usefulness.
   b) The specific change you are recommending.

6. Format your response as follows:
   <answer>
   <justification>
   [Your justification here]
   </justification>
   <recommended_change>
   [Your specific recommended change here]
   </recommended_change>
   </answer>

Remember, you are looking for the single most impactful change. Focus on the change that would bring about the most significant improvement to the text.

Now, does it work better? It’s definitely an example of how AI doesn’t just automatically know what to do when you ask. Clear instructions matter. In one case, Claude suggested that instead of writing an article, I should write a structured checklist. Do I want to hear that right now? Well.

This is THEFUTURE.

Subscribe to THEFUTURE

Get this newsletter in your inbox. For free.

The previous issue is The Traffic Apocalypse That Hasn’t Happened Yet, the next issue is When AI Companies Pay the Price.