Imagine copying a classmate's homework. They copied someone else, who copied you.

By the time the answer reaches you again, it has three different introductions and still nobody has opened the textbook.

That is a useful way to picture a problem researchers are studying: AI writes material for the internet, and future AI learns from that material. The answer sheet starts feeding itself.

An experiment involving 800 models gives us a reason to care about what goes into that loop. It also gives ordinary AI users a useful habit: bring something fresh to the conversation. Paper

I am Alex. Welcome to ShortCu8 by Innov8.

Lets Dive Deep 🐰🐰.

Before asking AI for a better answer, give it better material.

An original source, something you observed, or the result of a test can move the work forward. Asking for another rewrite may only change the wording.

Think of a hypothetical phone review written by someone who never held the phone. They ask AI to generate it from existing information, then publish it.

If that page enters a future training dataset, the next model has another review to learn from. But nobody tested the battery or noticed that the phone gets uncomfortable during a long call.

The amount of writing grew. The amount of firsthand experience stayed the same.

That is the part I find interesting. We can make an article longer, smoother and easier to skim without adding anything that happened in the real world.

Researchers from the University of Maryland and Pangram Labs trained 800 small language models on different mixtures of human writing and AI-generated web text.

When human material was scarce, adding AI text initially helped. As more was added, the benefit reversed. Models with plenty of human text saw harm sooner, measured by how well they predicted unseen human writing. Results

This was a September 30 preprint using small models. It doesn't establish that your favourite chatbot is getting worse, or that all AI-generated training data is harmful. Carefully prepared training examples are a different case from whatever generated articles happen to turn up online. Study

My takeaway is narrower: extra words can be a poor substitute for fresh information.

We aren't training models when we write a prompt. Still, the study made me think about how easily we can keep recycling an answer in our own work.

Say you're preparing a post about a tool. You ask AI for a draft, ask it to make the draft interesting, then ask another AI to improve it.

Before doing the third rewrite, try the tool. Record one thing that worked and one thing that annoyed you. Give those notes to the AI.

For research, open the original paper or announcement. Check whether it supports the claim you want to make. Five summaries of the same source give you five summaries to compare; they don't give you five independent pieces of evidence.

A prompt I'd use:

You can still use AI for the writing. Give it something worth writing about.

The researchers released their dataset, models and code so others can inspect the work. Their experiment concerns AI-generated text found on the web, mixed into training material. WildAI

The everyday writing advice above is my interpretation, rather than something their model-training experiment tested.

Before the next rewrite, add one thing you learned outside the chat window. A conversation with a customer, a failed test, or a detail from the original source gives the next draft somewhere new to go.

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.

With Attio, you’ll get:

Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?

Get started now

#AD2

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more