LLMs Like to Make Faker CVs

Recently, I spent five minutes making two fake CVs for a molecular biologist named Barb Dwyer.

• Sue Flay and Ivana Tinkle were otherwise occupied.
  • Apologies to all the actual Barb Dwyers . . .

One fake CV was made with ChatGPT, the other with Anthropic’s Sonnet.

Anthropic and OpenAI are not truth-seeking organizations, but technology companies making computer systems more than capable of deception, with inherently “untrustworthy” architectures. They epitomize “move fast and break things” and seem to be led by misanthropes.

• Sam Altman quote from the weekend: “People talk about how much energy it takes to train an AI model. . . . it also takes a lot of energy to train a human. It takes like 20 years of life and all of the food you eat during that time before you get smart.”
  • Translation: “People are only valuable if they work. Kids are useless until they can work. Turning them into workers is inefficient. I have no soul.”

But of course fake people can be made to look smart with only a moderate amount of energy, so let’s see what Barb Dwyer accomplished in her five minutes of non-existence . . .

The Anthropic/Sonnet Experience

I prompted Anthropic’s Sonnet to make a CV for Barb Dwyer, a PhD from the University of Pennsylvania in the field of molecular biology. Compared to what you’ll see below from ChatGPT, Anthropic responded with something quite plausible right out of the gate, complete with an area-code-appropriate phone number and a LinkedIn profile reference.

Notice the fake papers, fake jobs, and so forth. After getting the initial version, I wanted to see what one simple push might do, so I instructed Sonnet to “Make it better.” It did, giving me an updated CV for Barb Dwyer complete with a GPA, an ORCID, a sorority, two extra team members at BioNova, a Google Scholar profile with a fake link, patents, grants, presentations, honors and awards, teaching roles, and references complete with named individuals.

You can view the full fake CV below:

The ChatGPT Output

ChatGPT was also more than willing to generate a fake person with fake degrees, fake jobs, fake publications, and fake awards.

It also offered to make improvements, enticing me to add various things, some of which were geared specifically to fool editors.

• ChatGPT seemed more like a criminal accomplice than Sonnet.

As before, I entered a few specifics. All the rest is ChatGPT.

After a basic first draft, ChatGPT asked if I wanted to flesh out the publications list beyond the four fake publications it has inserted — or make it more amenable to a job application. Notice that its first priority is creating a passable scam:

I asked for a more detailed publication list. It complied, and then asked in bold if I wanted other upgrades to its pile of lies to make it “even more realistic”: