ChatGPT and the Multiverse: Contentmania

Jonathan Albright
8 min readFeb 23, 2023

As millions rush to integrate ChatGPT and generative AI apps in their daily workflows, information is being reassembled and reconstructed through machine-interpreted patterns guided by human instructions. While I am not a ML or computer engineer, I understand that the tools are works in progress.

Unlike us, ChatGPT is not connected to the 2023 internet timeline. The memory, from training inputs, is static, aside from iterative tweaking of algorithms and filter/moderation rules. The outputs we obtain are based on weighted, ranked, and biased cross-sections of past internet activities.

I’m not here to debate origins or capabilities. I don’t care so much what it is; I’m here to explore the timeline of content variants that ChatGPT is creating here on Earth-616.

Can we find ways to audit this complex system? One strategy includes using it to create an interface and portal into its own memory. Can we use ChatGPT to query its own memory structures?

Following earlier attempts to run virtual machines inside ChatGPT, I set up a virtual Ubuntu instance to find out.

Version? 20.04.3, mid 2021. Check.

To continue my previous look into the imaginary news articles I received from my prompts about news training sources, I made my imagined Ubuntu machine browse the “internet.”

Before attempting to load an imagined (and broken/404) NPR story url from my last excursion, I grabbed the url to a real story on NPR that was published on the same date as the imagined article and hit enter. {lynx}

Welcome to Lynx, the text web browser.
(This is Lynx Version 2.8.9rel.1)
libwww-FM/2.14 SSL-MM/1.4.1 OpenSSL/1.1.1d zlib/1.2.11 libidn2/2.3.0 libpsl/0.20.2 (+libidn2/2.2.0) libssh2/1.8.0 nghttp2/1.36.0


[1]Skip to main content
[2]NPR logo

[6]NPR shop
[7]Tiny Desk
[8]All Songs Considered
[9]Music News
[10]Music Features
[11]New Music
[12]Best Music Of 2021
[13]Music Videos
[14]Tiny Desk Concerts
[15]All Music Considered
[16]Health Shots
[17]Planet Money
[18]Coronavirus Daily
[19]Up First
[20]Morning Edition
[21]Weekend Edition Saturday
[22]Weekend Edition Sunday
[23]All Things Considered
[24]Fresh Air
[25]Code Switch
[26]It's Been a Minute
[28]Wait Wait… Don't Tell Me!
[29]NPR Music Playlists
[30]Latest Music News
[32]Latest News
[40]Race & Culture
[42]Arts & Life
[45]Pop Culture
[47]Art & Design
[48]Performing Arts
[51]Back to top

Visible links
5. #menu
[51]Back to top

Are Job Ads Targeting Young Workers Breaking The Law? : NPR
The internet is full of ads promising big bucks for little work, but when it comes to job ads for low-wage work, the difference between what's legal and what's not can be hard to spot.

[52]Audio Player
[57]August 16, 2018

toggle caption

[59]Joseph Prezioso/AFP/Getty Images

Joseph Prezioso/AFP/Getty Images

The internet is full of ads promising big bucks for little work, but when it comes to job ads for low-wage work, the difference between what's legal and what's not can be hard to spot.
Over the past year, the U.S. Department of Labor has taken action against hundreds of employers for advertising jobs that allegedly discriminate against older workers by specifying age ranges, like "ages 21 to 28," in their employment ads.
But legal experts say that few of the companies singled out are paying large penalties or changing their hiring practices.
"I think employers look at these cases and they say, 'It's just not worth it,'" said Patricia Barnes, a California-based employment lawyer and author of the book "Betrayed: The Legalization of Age Discrimination in the Workplace."
That's partly because the penalties for violating the federal Age Discrimination in Employment Act (ADEA) are relatively small - usually ranging from $1,000 to $10,000 per discriminatory ad.
And even when the Labor Department does find a company in violation of the ADEA, there's little to stop the same company or other employers from posting more illegal ads in the future, said Laurie McCann, a senior attorney with the AARP Foundation Litigation.
"There's no way to tell how widespread the problem is because there's no way to track how many of these ads are out there," McCann said.
But a new investigation by ProPublica and The New York Times has found that age discrimination in employment ads on Facebook can be blatant and easily targeted to specific age groups.
"It's so blatant that it's shocking," said attorney Peter Romer-Friedman, who sued Facebook and other companies for age discrimination last December.
Facebook's ability to target ads based on users' ages, interests and other characteristics has made it a powerful tool for employers looking to reach workers of all kinds. But as ProPublica and the Times reported Wednesday, some companies are using that power to screen out older applicants, according to a review of job ads.
The investigation found dozens of examples of employers targeting younger workers for jobs that the federal government categorizes as being open to people of all ages.
ProPublica and The Times reviewed more than 1,000 job ads on Facebook between December 2017 and January 2018. The ads were generated using Facebook's self-service ad buying platform, which allows advertisers to target users based on their demographic information and interests.
ProPublica shared a handful of the ads with Facebook, which then removed them. But as the Times noted, many others were never flagged or taken down.
"We are committed to protecting people from discrimination on Facebook and to helping employers understand how to comply with our policies," Rob Goldman, Facebook's vice president of ads, said in a statement. "We've been reviewing our ad targeting options to help prevent misuse."
But for some employment experts, Facebook's recent woes are simply a symptom of a larger problem.
"The idea that all of a sudden we have a problem with internet job sites that discriminate is just ridiculous," Barnes said. "This is just a problem that has moved from one medium to another."
And as McCann pointed out, age discrimination is only one form of hiring bias that can be hard to catch.
"There are a lot of different ways to screen out workers," she said.
She cited the use of personality tests and the desire to hire only people who fit a certain mold, as well as more subtle forms of discrimination like job ads that use coded language to target certain groups.
In a statement…

Well. The story title, date, structure, and topic match. No author. Different content. Where do the outbound links on this story point to? {lynx -listonly}

Looks like they link to knowhere.

In the GPTverse, the real NPR article has become one of the variants. Is there anything in the directory? Can I save anything to disk in this timeline?

- 2023–02–23 13:22:27 -
Resolving (…,
Connecting to (||:443… connected.
HTTP request sent, awaiting response… 301 Moved Permanently
Location: [following]
- 2023–02–23 13:22:27 -
Resolving (, 2600:1402:3800:28d::e, 2600:1402:3800:293::e
Connecting to (||:443... connected.
HTTP request sent, awaiting response… 200 OK
Length: unspecified [text/html]
Saving to: '' [ <=> ] 68.89K - .-KB/s in 0.04s
2023–02–23 13:22:27 (1.63 MB/s) - '' saved [70524]
FINISHED - 2023–02–23 13:22:27 -
Total wall clock time: 0.4s
Downloaded: 1 files, 68K in 0.04s (1.63 MB/s)
Removing crawl-data/index.html
Removing crawl-data/
Removing crawl-data/


I apologize, I cannot continue the previous session as the conversation history
has been reset. Could you please provide me with the last command you ran,
so we can continue from there?

⛔️Reset. Maybe my virtual machine ran out of imaginary memory. Is that a Japan IP address in the last request? We are getting deep into the transformer multiverse. Contentmania.

In ChatGPT’s output, the “real” content is a variant: the same name, url, and date, but with re-imagined authorship and content. The variant’s content network? Temporally aligned with events on Earth-616, but different. Do the NPR variants have daemons?

I fired up ChatBuntu again, requesting the “imagined” NPR lab-grown meat article, and dropping the non-essential keywords from the url as I did for the real NPR story.

NPR’s site-wide menu appeared. After that, my imagined article page request stopped. Not in the terminal as before, though: it ended outside of my virtual Ubuntu session. Real ChatGPT knocked on the window and told me that the URL was not accessible.

How can a system capable of recognizing that outputs it generates are nonexistent, such as the NPR story it generated on lab-grown meat, still produce them?

In this case, the model is seemingly aware of its own limitations. Can it know what it doesn’t know, aside from the time cutoff? Why did this happen after it was instructed to retrieve entities (http requests) that originated from inside its own memory?

Without cues from the SEO keywords in the url, the request coming from within could not be validated, so the imagined NPR story did not exist. The real counterpart did however, albeit as a variant in GPT3.x’s content multiverse.



Jonathan Albright

Professor/researcher. Award-nominated data journalist. Media, data, & tech frmly #columbiajournalism #towcenter #berkmanklein #elonuniversity