- AI Sidequest: How-To Tips and News
- Posts
- Will AI really replace ~historians~?
Will AI really replace ~historians~?
Plus, why you should stop threatening ChatGPT
Mignon Fogarty
August 04, 2025
Issue 73
On today’s quest:
— Update: OpenAI stops making shared chats public
— Will AI really replace historians?
— Changing the way we talk about AI
— Threats and bribes no longer necessary
— Something weird is happening with ChatGPT
— An easier interface for running open-source models on your computer
— OpenAI’s Agent falls short
— Who is responsible if AI manipulates markets?
— The origin of the chatbot ‘delve’
Last newsletter, I alerted you to ChatGPT conversations being indexed by Google and visible to the public. Soon after, OpenAI discontinued the feature.
It turns out that the only chats that were shared were those from people who had checked a box allowing it. But many people did not realize they had opted in, leading to shock and confusion.
Google Bard and Meta AI also had situations in which people unknowingly shared private chats.
It's a good reminder to check your settings. And you should still be careful what you put into chats. OpenAI is still required to save all chats on their servers — even temporary chats — because of an ongoing lawsuit, and Sam Altman has said that he believes those chats could be used against people in court.
Will AI really replace historians?
A Microsoft Copilot study about jobs that are likely to be affected by the availability of large language models has been getting a lot of play on social media, but the message is often being misinterpreted. The job titles listed are those that use a lot of activites that can be done by LLMs — such as writing and editing text, researching information, and evaluating or purchasing goods. The list does not necessarily represent a list of jobs that will be lost because of AI.
The study also looked at how satisfied people were with LLMs when they did a task versus when they helped a person do a task and found some interesting differences: “When the AI tries to directly provide support or advice, people are less satisfied than when it helps them provide support or advice to others.” (Emphasis added.)
The researchers caution against reading too much into the data, saying the future is hard to predict, and noting the classic example of ATMs leading to an increase in the number of bank tellers as banks “opened more branches at lower costs and tellers focused on more valuable relationship-building rather than processing deposits and withdrawals.”
Changing the way we talk about AI
Danny Liu, a professor of educational technologies at the University of Sydney, wrote an interesting post about changing the way we talk about AI. For example, he sees many critics say, “AI will never be able to do X.” But where does that leave us when AI actually can do X quite well, as is becoming the case in some areas?
I highly recommend reading the rest of his post for a good discussion of AI in education and five more useful ways to change our framing.
Threats and bribes no longer necessary
In the past, researchers found you could get better results from a chatbot by threatening it or offering it a bribe, but newer models no longer respond to these incentives, according to a new paper from a Wharton group.
Further, Ethan Mollick, one of the researchers, says asking an LLM to think through a problem step by step — another common prompting tip — is also no longer necessary. He says, “Don't rely on magic words or approaches to make AI performance better, aim for clarity and interact with the model to push it in the right direction.”
Something weird is happening with ChatGPT
I had two odd experiences with ChatGPT this week:
It actually mocked me when I asked how to find my previously shared chats. I was using temporary chat, so it’s gone, and I didn’t think to get a screenshot at the time, but it said something very much along the lines of “You think you should be able to find those? How cute.”
I asked it to find a source for an event I know happened in the early 2000s, and it told me it could find no evidence it ever happened. I then asked Claude, and it easily found two credible websites that included the anecdote.
This is highly unusual behavior, and the experience reminded me of the Stanford librarians I told you about back in issue 60 in June who were doing an experiment using AI to write alt text to help get their image collections online. One reason they decided to use an open-source model was to ensure they would have as consistent an experience as possible. They didn’t want to develop an effective workflow and then have the model change or disappear.
I do wonder what’s going on. I typically pay for only one pro subscription, and I had already been thinking about canceling my OpenAI subscription and testing Claude for a while. I’m leaning even harder in that direction now.
Let me know if you suddenly got strangely bad results from ChatGPT last week!
An easier interface for running open-source models on your computer
In my ongoing experiment to run local models on my home computer so I can power them from my solar roof, I had been stuck running them in the terminal, which isn’t exactly user friendly.1 But no longer! Ollama, the place where I download models, now has an easy-to-use app that gives you a chatbot-type interface.
If you’ve been curious but intimidated, the process just got easier. (YouTube demo)
OpenAI’s Agent falls short
This teaser from Fast Company sums up what I’ve been hearing about OpenAI’s Agent: “OpenAI has taught its chatbot to perform useful tasks on the internet. Just not very well.”
In the article, Harry McCracken describes Agent creating inaccurate ingredient lists, making flight suggestions that leave out the best options, compiling a list of job candidates based on old information, and repeatedly being confused about the current date.
Leon Furze, who tested Agent for his blog, had it make an “underwhelming” PowerPoint presentation and concluded “it doesn’t work.”
My failed Agent experiment
I did attempt an experiment so I could give you my own opinion, but I’ve had trouble coming up with a task that makes sense. I thought it might be nice to have all my blog posts in one document, so I asked Agent to visit the AI Sidequest website, get the text for each post, and copy it into a document.
It hemmed and hawed (“That’s an awful lot of pages. Are you sure you need them all?), and ultimately recommended I use the Beehiiv Export Posts tool to get them myself, which — I have to say — was a smart suggestion. So on the one hand, it didn’t do what I asked, but on the other hand, it seemed tuned to find the most efficient way to achieve the goal, which is hard to criticize.
Trust as an issue with agents
The only other task I could think of was deleting old links from the show notes in my podcast archive, but there’s no way I’m going to give Agent access to a mission critical tool and ask it to delete things.
I think trust is going to be a big issue for getting people to give agents the access they need to accomplish significant work. And another account I read but have subsequently misplaced described Agent recognizing the task was sensitive and requiring the user to approve every small step (officially called “Watch Mode”), negating any time savings from offloading the task.
I do believe agents will get better, but it seems this much hyped OpenAi product isn’t yet ready for prime time.
Ad: AI Tool Report
AI You’ll Actually Understand
Cut through the noise. The AI Report makes AI clear, practical, and useful—without needing a technical background.
Join 400,000+ professionals mastering AI in minutes a day.
Stay informed. Stay ahead.
No fluff—just results.
Who is responsible if AI manipulates markets?
I read about a Redditor who is letting ChatGPT manage a $100 investment in micro-cap stocks at Robinhood and subsequently saw a report from a cryptocurrency exchange saying that among 78,000 Gen Z users, 67% had activated at least one AI-powered trading bot.
Given the chatbot behavior study I wrote about in the last newsletter showing that different chatbots used different strategies in the prisoner’s dilemma (with Gemini being manipulative and ruthless), I wondered who would be responsible if AI agents manipulated the market. What happens if an autonomous bot creates social media accounts and posts false information about stocks it is holding? And how would you prevent this from being a huge problem in a world with hordes of agents?
I’m not the only one wondering about how AI will behave in financial markets (obviously). Another study found that when placed in simulated markets, AI trading bots did not compete with one another, but rather began colluding in price-fixing behaviors.
The origin of the chatbot ‘delve’
It’s well-documented that the word “delve” appears in text created by LLMs more often than it does in general writing, with some studies looking specifically at the use of “delve” in science writing, and I always thought it was because scientists had used “delve” a lot in previously published papers that were used to train AI. But it seems this is not the primary reason.
Instead, it seems “delve” is popular with LLMs because it is common in Nigerian English, and Nigerians played a large role in the AI training that happens after the models are originally made — a process called reinforcement learning with human feedback.
A recent paper on AI shaming highlighted comments from Nigerians on X who were angry that their particular style of English is now being associated with AI writing, which is looked down on by many Americans and Brits: “Imagine after being force-fed colonial languages, being forced to speak it better than its owners then being told that no one used basic words like ‘delve’ in real life. Habibi, come to Nigeria.”
Quick Hits
Bad stuff
Chatbots gave suicide instructions when the question was framed as an academic pursuit. The companies say they have fixed this problem, but it’s another example of the myriad ways people can bypass guardrails. — LA Times
Grok wrongly told people there were no tsunami warnings after the 8.8 earthquake in Russia — Bluesky
Medicine
Researchers create ‘virtual scientists’ to solve complex biological problems — Stanford Medicine Center News
Cedars-Sinai's AI tool delivered 24/7 care to 42,000 patients (77% of AI recommendations were rated optimal, while 67% of physicians' decisions were rated optimal) — Business Insider
Job market
The history of the world will be split into a pre-AI and a post-AI era. Too many people are still in denial. — Yascha Mounk
Education
When Meta embedded AI bots in its apps, even students in the most remote corners of Colombia gained access. But rather than boosting learning, it’s getting in the way. — Rest of the World
I’m laughing
The Story of the World’s First AI-Organized Event (the funniest AI story I’ve read since Claude’s vending machine experiment) — LessWrong
Other
Amazon Updates Code to Keep Out Google AI Shopping Tools (Amazon has reportedly “taken similar precautions with bots from companies like Perplexity, Anthropic and OpenAI) — PYMNTS
Amazon to Pay New York Times at Least $20 Million a Year in AI Deal (for content from the cooking section and The Athletic) — Wall Street Journal
The rise of AI tools that write about you when you die — Washington Post
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]
There are other ways to get a chat interface, but they seemed more complicated than the new Ollama interface, and I hadn’t gotten around to trying them after my first attempt failed.
Written by a human