This week, we start with new research showing how AI is already reshaping entry-level jobs and who’s feeling the brunt of the change. Then we look at OpenAI’s GPT-realtime, a model that brings voice and multimodal conversations closer to human speed than ever. Let’s dive in ⬇️
In today’s newsletter ↓
🧠 Clear evidence AI is shrinking the workforce
🎙️ GPT-realtime conversations gets an upgrade
📚 AI short stories top human favorites
☁️ New controls for AI web crawlers
🎯 Challenge: Test your creativity vs AI’s flexibility
A new working paper, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” tracks how AI adoption is reshaping early-career work using monthly payroll records for millions of U.S. workers through July 2025. The authors document six consistent patterns — learn more about each below.
Since late 2022, workers ages 22 to 25 in the most AI-exposed jobs have seen a 13 percent relative drop in employment. This decline holds even when controlling for firm-level shocks like interest rate changes or sector downturns. The data suggest that AI, not broader economic cycles, is driving the difference.
OUCH! A report released by Stanford economists titled "Canaries in the Coal Mine?" is likely to get attention as it reports that AI is leading to fewer jobs for younger workers and recent college grads. In the research, high-frequency payroll records from millions of American
— Holger Zschaepitz (@Schuldensuehner)
8:45 AM • Aug 27, 2025
U.S. employment overall is on the rise, but the gains aren’t evenly spread. In the most AI-exposed jobs, younger workers saw a 6 percent decline in employment between 2022 and mid-2025, while their older colleagues in the same roles grew by 6 to 9 percent. This indicates a generational gap in who benefits when AI enters the workplace.
The type of AI use matters. Jobs where AI is primarily automating tasks — such as coding or routine data entry — saw sharper declines in young workers. In contrast, jobs where AI augments human work, like healthcare or management support, showed more stability and even growth.
Credit: Stanford Digital Economy Lab
Researchers tested whether these changes could be explained by industry downturns or firm-specific issues. Even after adjusting for these factors, the employment declines for young workers in exposed jobs remained strong. This suggests the pattern is directly linked to AI adoption, not broader market turbulence.
Wages have remained relatively stable across exposed and non-exposed groups. The main effect has been fewer entry-level hires, not shrinking paychecks. Economists call this “wage stickiness,” meaning firms adjust by reducing headcount rather than cutting salaries.
These results are consistent across industries and job types, even when excluding tech companies or remote-friendly roles. Importantly, the patterns do not appear before the rise of large language models in late 2022. That timing reinforces the idea that AI adoption is the catalyst.
The study argues that entry-level workers rely heavily on codified knowledge — the kind found in textbooks or training manuals — which generative AI can replicate easily. In contrast, older workers lean more on tacit knowledge built through experience, intuition, and social skills, which are harder to automate. This dynamic explains why younger employees are the first to feel the brunt of AI’s disruption.
OpenAI introduced GPT-realtime and moved its Realtime API to general availability with new production features. The model processes and generates audio directly, reducing latency and enabling fluid, interruption-friendly conversations that feel much closer to live dialogue than previous cascaded pipelines.
You can test it right now in the OpenAI Playground.
OpenAI calls GPT-realtime its most advanced speech-to-speech model, with improvements in audio quality, instruction following, function calling, and comprehension. Two new voices, Cedar and Marin, debut in the Realtime API, and the system can follow fine-grained delivery cues like “speak empathetically in a French accent.”
The Realtime API is officially out of beta and ready for your production voice agents!
We’re also introducing gpt-realtime—our most advanced speech-to-speech model yet—plus new voices and API capabilities:
🔌 Remote MCPs
🖼️ Image input
📞 SIP phone calling
♻️ Reusable prompts— OpenAI Developers (@OpenAIDevs)
5:53 PM • Aug 28, 2025
The Realtime API adds remote MCP server support to plug in external tools, image input for grounding conversations in what the user is seeing, and SIP phone calling support to connect agents to phone systems. Reusable prompts and better handling of long-running function calls round out the release.
By handling input and output audio in a single model, the system cuts the wait and preserves speech nuance. OpenAI highlights lower latency and more natural, expressive responses compared to older speech pipelines.
My favorite demo of the new gpt-realtime model from @matthieulc -- Shoggoth Mini using Realtime API with image input
— Peter Bakkum (@pbbakkum)
5:33 PM • Aug 28, 2025
GPT-realtime is generally available to all developers, with prices reduced by about 20 percent versus the prior preview: $32 per 1M audio input tokens and $64 per 1M audio output tokens. OpenAI says the API supports EU data residency and is covered by enterprise privacy commitments.
Challenge: This week, stress-test your creativity against an AI’s flexibility.
Here’s what to do:
💡 Step 1: Pick one idea. It could be anything — a podcast theme, a small business concept, or even tonight’s dinner. The simpler, the better.
🔄 Step 2: Ask for three takes. Prompt an AI to expand that idea into three distinct versions. For example: “Give me three ways to pitch a sci-fi podcast” or “Show me three spins on spaghetti night.” Don’t settle for the first pass — if they look similar, nudge the AI for wilder contrasts.
🧪 Step 3: Compare the results. Which version feels most original? Which one is realistic to try? Which one completely misses the mark? Circle your favorite and consider why you chose it.
📝 Step 4: Add your human touch. Rewrite the best version with your own style, knowledge, or humor. Even small tweaks — like grounding it in your own experiences or tone — can transform an average AI idea into something memorable.
🔆 Why it matters: AI is good at multiplying possibilities, but you’re better at spotting what actually works. This challenge shows how machine brainstorming can widen the field, while your judgment narrows it to the ideas that feel right. In the end, the real creativity comes from how you remix the raw material.
That’s it for this week! Are we adopting AI so fast that we’re leaving young people behind, and has OpenAI just released the customer service agent of the very near future? Hit reply and let us know your thoughts.
Zoe from Overclocked