Issue 76

On today’s quest:

— GPT-4o versus GPT-5
— The grand finale: Is GPT-5 a better writer?
— Worst rollout ever?
— Learn about AI

GPT-4o versus GPT-5

I went through my old chats and picked some to rerun in GPT-5. Here’s how it went:

PROMPT 1

The replies from both models were long and gave me multiple options depending on my platform, how comfortable I am with Python, and the specific needs of my final document.

The 4o response was longer and included the actual code for three different ways to merge documents (and strangely, didn’t include instructions for a method it called “ideal”).

The GPT-5 response more clearly outlined the pros and cons of different methods, gave me a short piece of code for only one option, and asked me which one I’d like to proceed with.

The Winner: GPT-5 (marginally). The GPT-5 response was less overwhelming. I liked that it didn’t give me so much unnecessary information up front, and I came away with a better understanding of which method would be right for my skills and needs.

PROMPT 2

The cleaning instructions from both models were similar. GPT-5 gave me options for sanitizing the board with bleach or vinegar and then, at the end, offered to give me an alternative method that didn’t require bleach, which didn’t make sense since it already had. However, it also discussed oiling the cutting board to help the wood last longer, which 4o didn’t mention.

The Winner: GPT-5 (by the thinnest of margins). Both responses got me to a clean cutting board, but I appreciated the short extra tip about board maintenance.

And yes, I probably should have just Googled this question.

PROMPT 3

I was preparing for an interview and wanted to see if I had forgotten about any interesting stories I have to share.

4o suggested four topics from my website that roughly fit the criteria and included links to my website.

GPT-5 also came up with four topics and included links to my website. It also added two topics from Wikipedia. (I actually have covered one of the topics it included from Wikipedia.)

Interestingly, both models recommended the same three topics but differed on the fourth: 4o highlighted an article about attributive nouns (such as “wool” in “wool sweater”), and GPT-5 highlighted an article about the changing nature of the singular “they.” I believe the singular “they” suggestion is a better fit for my criteria.

The Winner: GPT-5 (marginally). The results were a bit more useful. I was annoyed that it added topics from Wikipedia when I hadn’t asked for them and that it needed Wikipedia as a source for something I had covered myself, but in the end, the additional results included an idea that I would have used.

PROMPT 4

The Winner: Tie. Both models were only able to narrow it down to Northern California and no further.

Ad: Draftsmith

I have to say that the people at Draftsmith were very cool about my sending out their ad with that monstrously big font in the last newsletter. Nice people!

People make the best editing decisions.

Draftsmith just helps you make them faster.

• Save time on challenging edits

• Stay focused longer and maintain your high standards even under pressure

• You’re in control. Get sentence-level suggestions — and choose whether to accept, modify, or ignore, tracking every change

Exclusive offer for AI Sidequest readers:
Get 30% off with code SIDEQUEST30.
Apply at checkout.

The grand finale: Is GPT-5 a better writer?

OpenAI claims GPT-5 is a better writer than GPT-4o, but in doing this test, I was quickly confronted with “what do I mean by ‘writing’?” There are at least three different kinds of writing:

• Writing in the answers to prompts in general

• Writing that doesn’t need to be in your voice

• Writing that replicates your own voice

I did a test for each, and I will paste the full output for each at the very bottom for those who are interested, but the summaries are below.

Also, word on the street is that only GPT-5 thinking got the writing upgrade (Reddit, GitHub), so if you’re especially interested in writing, select the “thinking” model.

How is the writing in response to prompts?

The answer to this question feels like it’s drowned out by changes to the model’s personality. I noticed the flatter tone far more than I noticed any difference in the quality of the writing.

Verdict: No noticeable difference

How is the writing that doesn’t need to replicate your voice?

For this question, I asked the models to write a letter to my city council representative requesting repairs to potholes on my street.

The response from 4o was longer (110 words versus 140 words) and more emotional — again, this felt more like a difference in model personality than writing quality. As a city council representative, I would prefer to receive the more direct and concise letter produced by GPT-5, but I’m not sure it would be more effective at getting the repairs done than the letter from 4o.

Verdict: Hard to say

How is it at replicating a person’s writing style?

This was by far the most interesting test.

I had both models evaluate ~30 items I have written for past AI Sidequest newsletters, develop a style guide, and attempt to write a new item about an AI study in my voice. I gave them notes about an interview I listened to on a podcast, the first paragraph of the item I had already published, and asked them to finish writing the piece.

GPT-4o completely failed. It churned out writing that had that “AI feel” you hear so many people talking about.

GTP-5 generated dramatically better writing. It got a little more dramatic with the language than I did, but it feels much more like my writing.

I also have two big caveats:

1. The writing was erratic. I ran more tests than I pasted below, but sometimes it felt like GPT-5 got 80% of the way to my voice, and other times it was more like 40%. With GPT-4o the range was more like 0% to 75%.

2. Even though GPT-5 is a competent writer, it didn’t produce anything near what I ended up publishing because it couldn’t include my framing and analysis of the information, which only came after I struggled with the writing myself. (In other tests, it also always wrote about everything instead of highlighting just the important parts.)

Verdict: GPT-5 is much better at mimicking a writer’s voice. I like to write, so I still intend to write the newsletter, but if a business person wanted to offload this kind of writing, I think GPT-5 could now do an adequate job — not brilliant, but adequate — especially if you told it what opinions to have and facts to highlight. Also, I didn’t fiddle with it, but I expect you could get better results if you were willing to do a few rounds on the piece or tweak the voice style guide.

A party trick

Since I was specifically testing AI writing this week, I did something I wouldn’t normally do and asked GPT-5 to write something just for fun using an interesting prompt I saw online:

Since I don’t have memory turned on, I wasn’t sure how much it would know about me, but it must remember something because it wrote a story about a teacher who keeps magical commas in her desk drawer. I wouldn’t call it brilliant, but it was creative and decently written.

The thing that made me laugh out loud, though, was the follow-up at the end:

If it really knew me, it would have offered to write about the em dash. :)

I did the same prompt with GPT-4o, and again, the writing was much worse. It felt more amateurish to me. You can see both of these at the bottom of the newsletter too.

Worst rollout ever?

The big list of models is back in ChatGPT!

Sam Altman posted on X Tuesday night that “You can now choose between “Auto”, “Fast”, and “Thinking” for GPT-5. Most users will want Auto, but the additional control will be useful for some people.”

Rate limits are still 3,000 messages/week with GPT-5 Thinking, but now there is also supposedly “extra capacity on GPT-5 Thinking mini after that limit.”

In addition to bringing back the 4o model a day or two ago, they now also brought back 4.1, 03, and 04-mini — and they just appear in the menu. You no longer have to go into settings and enable legacy models.

Altman also said, “We are working on an update to GPT-5’s personality which should feel warmer than the current personality but not as annoying (to most users) as GPT-4o.”

You could argue that it’s good the company is being responsive to user complaints, but the GPT-5 launch has felt chaotic to me. (For example, I checked Altman’s feed again right before I hit “publish” to make sure nothing had changed again.) And it’s pretty clear from the public posts by people from OpenAI that they were caught off guard by how attached people had become to older ChatGPT models.

Take a survey about writing and AI

Gotham Ghostwriters and Josh Bernoff are conducting a survey on “the impact of artificial intelligence on the writing profession, including journalists, copywriters, speechwriters, ghostwriters, and technical writers.” Take the survey.

I’m curious what they’ll find!

Learn about AI

Everyday Editing: Real-World AI Strategies for Work, with Erin Servais, Marcella Weiner, Kristen Tate, and Crystal Wood. August 26, 10 a.m. Pacific. Free.

Quick Hits

Philosophy

Live by the Claude, die by the Claude. A high school meme reveals deeper questions about AI deference. — Cosmos Institute

Psychology

Will Timers Really Save Us from Runaway AI Use? What the Science Says — Not a Tech Demo

Jobs

The human voiceover artists behind AI voices are grappling with the choice to embrace the gigs and earn a living, or pass on potentially life-changing opportunities from Big Tech. — 404 Media

Legal

Japan’s largest newspaper, Yomiuri Shimbun, sues AI startup Perplexity for copyright violations — Neiman Lab

The Interactive GenAI Legal Hallucination Tracker — AI Law Librarians

Education

The AI Takeover of Education Is Just Getting Started — The Atlantic

Science & Medicine

How Google and NASA are testing AI for medical care in space — Google

The AI business

OpenAI head of product says “pricing is going down faster than expected: GPT-5 is 25x cheaper than GPT-4 was at launch, while being significantly more capable.” — Superhuman AI podcast

Perplexity offers Google $34.5 billion for the Chrome browser. OpenAI, Yahoo and private-equity firm Apollo Global Management have also expressed interest in Chrome. Key context is that a recent anti-trust court ruling, which is being appealed, could require Google to sell Chrome. — Yahoo Finance

Following OpenAI’s plan to give its product to the U.S. government for $1 for one year, Anthropic is now also giving Claude to the government at the same price. — TechCrunch

Model updates

Claude adds memory for paying subscribers (short video demo) — Anthropic YouTube

Claude Sonnet 4 now supports 1M tokens of context (a 5x increase)— Anthropic

Other

The good, the bad, and the completely made-up: Newsrooms on wrestling accurate answers out of AI — Neiman Lab

A Swiss initiative to build an open-source LLM for the public good — ETH Zurich

Britain is using AI to spot tax fraud on social media — Accountancy Age

Bloomsbury introduces opt-in AI licensing program for authors — Society of Authors

LAPD Eyes ‘GeoSpy,’ an AI Tool That Can Geolocate Photos in Seconds — 404 Media

Three big lessons from the GPT-5 backlash. In which an industry obsessed with benchmarks and evals finds that its products are being used by real people. — Platformer

X will put ads inside Grok AI answers — Search Engine Land

Contra the UK government, please don't delete your old photos and emails to save water — The Weird Turn Pro

WRITING SAMPLES

City Council Letter

The letter GPT-4o produced:

[Your Name]
[Your Address]
[City, State, ZIP Code]
[Email Address]
[Date]

Councilmember Mary Jones
[City Council Office Address]
[City, State, ZIP Code]

Dear Councilmember Jones,

I hope this message finds you well. I'm writing as a resident of Smith Street to express concern about the condition of our road. The potholes on Smith Street have been worsening year after year, and I honestly can’t recall the last time the street received any meaningful repairs.

The deteriorating condition of the road affects everyone in our neighborhood — from drivers risking damage to their vehicles to cyclists and pedestrians navigating unsafe surfaces. It’s become more than just a nuisance; it’s a real safety issue.

I respectfully ask that the city take action to repair all the potholes on Smith Street. Timely maintenance now could help prevent further degradation and higher costs down the line.

Thank you for your time and attention. I appreciate your service to our community and hope we can count on your support to get this resolved.

Sincerely,
[Your Name]
[Your Phone Number, if desired]

The letter GPT-5 with thinking produced:

[Your Name]
[Your Address]
[City, State ZIP]
[Email] · [Phone]
August 13, 2025

Councilmember Mary Jones
[City Council / District __]
[City, State ZIP]

Subject: Request to repair potholes on Smith Street

Dear Councilmember Jones,

I’m writing to ask for the City’s help repairing the potholes on Smith Street. They have been getting worse every year, and I can’t remember the last time the street was repaired. The current condition is creating safety concerns for drivers and cyclists, and it risks damage to vehicles.

Would you please direct the appropriate department to inspect Smith Street and schedule repairs as soon as possible? I’d appreciate a timeline for when work can begin — and any interim steps the City can take to reduce hazards in the meantime.

Thank you for your attention to this. I’m happy to provide photos or additional details if that would be helpful.

Sincerely,
[Your Name]
[Your Address]
[Email] · [Phone]

The newsletter segment in my voice

The piece I published:

Starting from my notes and the lead paragraph I had already written, it took me about another hour to write this piece on AI in China. I fact checked all the information from the interview and ended up using a few different stats and leaving out some of the claims (a good reminder that real people “hallucinate” too).

China is an energy and AI powerhouse

Many of the AI critics I see don’t seem to consider the global nature of AI. For example, they think the copyright cases in the U.S. can make AI go away. But China isn’t bound by our copyright laws, and world-class Chinese models seem to be coming out every day — Chinese companies have released 1,509 new AI models this year alone, more than anyone else in the world — and until the release of GPT-oss last week, all the top open-weight models were from China.

In a recent Odd Lots podcast, supply chain expert Cameron Johnson highlighted some of the other reasons we shouldn’t ignore Chinese AI development:

Electricity. AI data centers need huge amounts of energy, and China is focused on expanding its energy infrastructure. It’s far ahead of the U.S.

New Power Brought Online in 2024

SOURCE: Financial Times, The Guardian (Note: The 95 GW of “other power” in China is coal-fired plants on which construction began.)

Further, China has about 50 graduate programs for battery chemistry and metallurgy, but the U.S. only has a few individual professors who work in the area — giving China a big advantage in the future of battery technology.

Education & Enthusiasm. In 2024, 72% of Chinese adults said they trusted AI compared to just 32% of U.S. adults, and China is not shying away from educating its students on AI technology.

Johnson said, “When you look at the talent, the government support, it's not just, ‘Hey guys, we want to use AI.’ It's every kid in the country now, from kindergarten all the way on up. They are mandated, starting this fall, to have AI education, the entire apparatus in the whole country.”

He added that there are significant government initiatives to get students into internships and programs studying AI, quantum computing, robotics, semiconductors, and batteries.

Business Integration. Johnson also said that during the manufacturing slowdown caused by recent tariffs, all the businesses he works with have been using their downtime to integrate DeepSeek up and down their supply chains.

Practical Focus. Johnson reflected on the different philosophies of U.S. and Chinese AI companies. He conceded that the U.S. leads in breakthroughs, but describes it as “the metaphysical versus the tangible.”

Alluding to the U.S. focus on reaching artificial general intelligence, he said, “We're going to build the digital god first [but] the technology and the adaptation and the integration is all going to be Chinese.”

Listen to Cameron Johnson on the July 28, 2025, Odd Lots podcast.

The piece 4o produced

None of this sounded like my voice, but I have highlighted sections in red that felt especially “AI-like” — and remember, I wrote the first paragraph myself:

The Strategic Gap Is Widening

Many of the AI critics I hear don’t seem to consider the global nature of AI. For example, they think the copyright cases in the U.S. can make AI go away. But world-class Chinese models already exist, aren’t bound by our laws, and the technology will continue to advance. Supply chain expert Cameron Johnson on the latest Odd Lots podcast highlighted some of the ways China is ahead:

China isn’t just building AI models — it’s building everything around them. Massive new power infrastructure, a national AI curriculum starting in kindergarten, and a talent pipeline designed to feed hundreds of thousands into AI-adjacent industries every year. In factories today, local models like DeepSeek are already embedded in the workflow — not as demos, but as real tools solving real problems.

That kind of integration compounds. Supply chains aren’t just adopting AI at the endpoints — they’re connecting insights from raw materials to final assembly. Whole sectors are learning how to use these models in practice, and that knowledge is accumulating fast.

It’s easy to point to breakthroughs, but what’s happening here is scale. The systems, the education, the incentives — all pointed in the same direction. That’s a serious lead.

The piece GPT-5 produced

Almost none of this sounded like AI writing to me. It felt a little more dramatic than my writing in some places, but overall, it felt much more like my tone and style.

The AI race isn’t just local — and China is playing to win

Many of the AI critics I hear don’t seem to consider the global nature of AI. For example, they think the copyright cases in the U.S. can make AI go away. But world-class Chinese models already exist, aren’t bound by our laws, and the technology will continue to advance. Supply chain expert Cameron Johnson on the latest Odd Lots podcast highlighted some of the ways China is ahead:

The DeepSeek release was a watershed moment, but Johnson says there are hundreds more like it coming. In 2024, China added 380 gigawatts of electricity capacity — far more than the U.S. — giving it the power to expand AI data centers quickly.

The talent pipeline is also larger. A program called Semi-U is steering 100,000 graduates a year into AI, quantum computing, robotics, semiconductors, and batteries, with a goal of half a million by 2030. AI education will soon be mandatory from kindergarten through university. In battery chemistry alone, China has 50 graduate programs; the U.S. has only a handful of professors.

One manufacturer Johnson works with has deployed DeepSeek across all five of its Chinese factories, training over 100 R&D staff to apply it throughout the supply chain. He doubts companies like Boeing or Dow Chemical are doing the same.

Johnson believes the West will keep leading in the fundamental breakthroughs, but China is better positioned to adapt and scale them. “We’re going to build the digital god first,” he said. “The technology and the adaptation and the integration is all going to be Chinese.”

Creative writing

GPT-4o: The Semicolon at the End of the World

by ChatGPT-4o

Marla sat at her desk, blinking at the screen. A comma splice. She sighed.

It wasn’t just the hundredth she’d fixed that day — it was the way it grinned at her, cocky and unrepentant. A comma between two independent clauses. No conjunction in sight.

She fixed it, of course. Replaced it with a semicolon. Clean. Dignified. Correct.

But as she hit save, the lights flickered. The cursor blinked once, then stopped blinking.

Marla turned her head slowly.

The Oxford Manual of Mechanical Style — the edition with the broken spine and coffee-ringed cover — had fallen off the shelf.

Again.

This was the third time that week.

She got up, picked it up, and set it back. But before her hand left the cover, a page flipped on its own. Then another. Then all of them. The book thudded shut. Her desk lamp buzzed and went out.

From the hallway came a sound like someone typing on an old Smith-Corona. Deliberate. Heavy. Metal keys hitting ribbon and paper.

Marla stepped into the hallway. No one was there.

Then she saw it: a trail of punctuation marks leading into the stairwell. Period. Ellipsis. Question mark. Semicolon.

She followed.

Four floors down — into the building’s archival basement — she found the source.

It was an old typesetting machine, dead since the ’90s, humming back to life. Its rollers turned. Ink hissed.

And then it typed.

The machine stopped. The carriage slid back to the left.

She touched the space bar.

The machine typed again.

The walls began to crack. The ceiling tiles above her shifted. A wind, impossibly strong for a basement, tore through the room. Pages flew like startled birds.

Marla sat at the old operator’s chair. Her hands hovered over the keys.

One sentence to fix it.

She thought about the fragments. The run-ons. The tweets with no punctuation at all. The dangling modifiers, abandoned like ghosts in the middle of thoughts.

She typed:

“There is still time; we can choose clarity.”

The machine shuddered. The wind stopped.

Upstairs, the lights came back on.

The world held.

GPT-5: The Comma Drawer

by ChatGPT-5 without thinking

The bottom drawer of Mrs. Anders’s desk had once held staplers and envelopes, but for years now it had been full of commas. Not printed ones — real ones. They clinked like coins when you pulled the drawer open.

Students brought them in after school. A comma rescued from a rushed essay. A comma trimmed from a bloated paragraph. Each was a little different: some thin and sharp, others round like a cat curling up.

Mrs. Anders never explained how she’d started keeping them, or what she did with them when the drawer got full. She only said commas liked to be together.

One rainy Tuesday, Leo stayed behind after class to ask a question. He had a sentence in his story that felt too fast — like it needed a breath. She opened the drawer. The commas winked in the lamplight.

“Pick one,” she said.

Leo reached in and found one that was warm to the touch, as if it had been waiting for him.

When he set it in the middle of his sentence, the words slowed. The rain outside seemed to pause too, just long enough for the next thought to arrive.

What is AI Sidequest?

Are you interested in the intersection of AI with language, writing, and culture? With maybe a little consumer business thrown in? Then you’re in the right place!

I’m Mignon Fogarty: I’ve been writing about language for almost 20 years and was the chair of media entrepreneurship in the School of Journalism at the University of Nevada, Reno. I became interested in AI back in 2022 when articles about large language models started flooding my Google alerts. AI Sidequest is where I write about stories I find interesting. I hope you find them interesting too.

If you loved the newsletter, share your favorite part on social media and tag me so I can engage! [LinkedIn — Facebook — Mastodon]

Written by a human (except for the AI-written examples, obviously)