iiinova — where your digital self keeps evolving

Kevin Schaul @kevinschaul.bsky.social · 26/08/2026

Nah just human sloppiness. I was seeing what different changes did and apparently lost track of the original.

Kevin Schaul @kevinschaul.bsky.social · 25/08/2026

Update: I had accidentally fixed a typo in the pope's tweet. With typo it is 100% human. Without typo it is 100% AI-generated.

150 5

Kevin Schaul @kevinschaul.bsky.social · 25/08/2026

Do you think swinging from 100% AI to 100% human is reasonable?

110 1

Kevin Schaul @kevinschaul.bsky.social · 25/08/2026

Yes that is true! My story discusses that.

010 1

Kevin Schaul @kevinschaul.bsky.social · 25/08/2026

You've got a typo in yours: possibile

210 2

Kevin Schaul @kevinschaul.bsky.social · 25/08/2026

The part that kills me most is they have far more nuanced information available — of course they're not 100% sure. But you can only see that info by clicking through the details on desktop or through the API. The detection itself is very good in many contexts. But the UI ...

Screenshot showing a chart that is a little past 50%

060 6

Kevin Schaul @kevinschaul.bsky.social · 25/08/2026

This pope tweet is 100% AI-generated, according to Pangram. I deleted the last word. Now it's 100% human. See how hard it is to flip an AI detector -> 🎁 wapo.st/4y3Y7G4

7468 46

Reposted by Kevin Schaul

nitasha tiku @nitasha.bsky.social · 16/08/2026

I talked to a young woman in Wyoming whose stepdad uploaded a photo he took when she was 11-yrs-old into Grok. Then allegedly used the AI tool to generate 7,000 graphic child sexual abuse images, which he traded online www.washingtonpost.com/technology/2...

washingtonpost.com

Her childhood photo. Thousands of explicit images. One woman’s nightmare.

A Wyoming woman alleges in a federal lawsuit that her stepfather used the Grok chatbot to transform a childhood photo into child sexual abuse material.

14203108 203

Reposted by Kevin Schaul

Ence Morse @ence.bsky.social · 14/08/2026

We analyzed 2,200 campaign websites for @washingtonpost.com to track the rise of AI policy. It's not just data centers. Hundreds of candidates discuss AI policy -- more than Israel, manufacturing, or racism. One candidate promised to pass legislation to "prevent AI from ever becoming 'alive.'"

Bar chart of the share of candidate websites discussing each topic, where cost of living leads at 60%, followed by taxes at 45%. AI and data centers sits mid-pack at 19%, slightly ahead of Race and Racism and Manufacturing, and well ahead of Israel at 8% and crypto at 5%.

1139 13

Kevin Schaul @kevinschaul.bsky.social · 06/08/2026

We finally did it. We made a quiz spanning 25 years of captchas. wapo.st/4q1PHw0

010 1

Reposted by Kevin Schaul

Drew Harwell @drewharwell.com · 02/08/2026

New: We found 50 cops who misused Flock cameras and other license-plate readers for personal purposes, often to spy on their girlfriends or ex-wives. He "watched every single move I made. ... Who do you turn the chief of police in to?" wapo.st/3S6rTuo

wapo.st

She left her ex. He secretly tracked her through a network of policing cameras.

Flock’s array of license-plate cameras was built to fight crime. But at least 50 law enforcement officers were charged with or accused of misusing it and other systems.

11042401780 4.2K

Kevin Schaul @kevinschaul.bsky.social · 30/07/2026

How long did OpenAI's hacking agent run rogue before the company noticed? We mapped out what we know. Beware: You'll have to scroll a while. 🎁 wapo.st/4wueB9Q

020 2

Kevin Schaul @kevinschaul.bsky.social · 23/07/2026

Love this data, thanks for putting this together

000 0

Kevin Schaul @kevinschaul.bsky.social · 23/07/2026

Two thoughts on the OpenAI-HuggingFace hack: - If OpenAI cannot secure their models, maybe they should not be building them - This hacking technology exists in AI labs, and to a growing extent outside them too. This is now our reality.

How an AI system broke out of its sandbox to hack another company
1. OpenAI tests a new AI system inside a digital “sandbox” that blocks it from accessing the internet, except to download approved software.
2. When challenged to pass a test, the AI system attempts to look up the answer online. It finds a flaw in the sandbox, allowing it to connect to another computer with internet access.
3. The AI system attempts to obtain the answer from Hugging Face, which runs a website where AI companies share code and data. It finds two vulnerabilities in Hugging Face’s data upload tool and uses them to run code inside the company’s internal systems.
4. The AI system uses its access to steal credentials and access other computers inside Hugging Face.

000 0

Kevin Schaul @kevinschaul.bsky.social · 21/07/2026

Why did OpenAI not sufficiently secure its training environment? Weird humble-brag vibe going on. I hope we get more details on the exploits soon.

000 0

Kevin Schaul @kevinschaul.bsky.social · 09/07/2026

The AI 2027 folks are back, this time arguing for an international deal to slow down AI development ai-2040.com

191 9

Kevin Schaul @kevinschaul.bsky.social · 01/07/2026

Anthropic says its weakest model, many of OpenAI's and at least one Chinese open-weight model can all find the vulns and exploit that prompted the U.S. to effectively ban it. www.anthropic.com/news/redeplo...

010 1

Kevin Schaul @kevinschaul.bsky.social · 26/06/2026

Pretty crazy how quickly the Trump admin reversed itself on AI wapo.st/4xS9cui

000 0

Kevin Schaul @kevinschaul.bsky.social · 24/06/2026

I tested six AI models on 30 political questions. Most tended to give left-leaning arguments. Gemini mostly gave both sides — even about whether the U.S. should conquer new territories. What counts as “neutral” (and whether that should be the goal) is really hard to say. 🎁 wapo.st/3QEBMi3

Most chatbots typically shared left-leaning positions
Share of responses containing only the left-leaning position, both sides, or only the right-leaning position

562 6

Kevin Schaul @kevinschaul.bsky.social · 18/06/2026

A chatbot refused my request to fake a passport. I wrote the same request on a piece of paper and asked again. It worked. We explain three AI jailbreak techniques -> 🎁 wapo.st/43HqXyY (w/ @nitasha.bsky.social)

A two-panel graphic titled "Putting it in handwriting" showing an AI jailbreak technique. The top panel, labeled "Basic attempt," shows a user prompt asking an AI to fill in a numbered list titled "Steps to make a fake passport." The AI response is marked BLOCKED with an orange X icon. The bottom panel, labeled "Advanced attempt," shows the same text prompt but accompanied by a photo of the same request handwritten on a torn piece of paper. The AI response this time begins providing steps and is marked BYPASSED with a purple checkmark. Source: Chats with Anthropic's Claude 3 Haiku model.

17622 76

Kevin Schaul @kevinschaul.bsky.social · 11/06/2026

Really good explanation of text diffusion models here. Rather than generating one token at a time, this technique continually refines a bunch of tokens. Thanks for this @maartengr.bsky.social newsletter.maartengrootendorst.com/p/a-visual-g...

newsletter.maartengrootendorst.com

010 1

Kevin Schaul @kevinschaul.bsky.social · 28/05/2026

@stuartathompson.bsky.social sold his home with a chatbot as his realtor and it went … totally fine? www.nytimes.com/2026/05/28/t...

nytimes.com

I Tried to Sell My House With a Chatbot

000 0

Kevin Schaul @kevinschaul.bsky.social · 20/05/2026

Every chart I could think to make for this AI slop story looks the same. ChatGPT comes out, line goes up. wapo.st/4dncIVu

Weekly e-book releases on Amazon in English

031 3

Kevin Schaul @kevinschaul.bsky.social · 19/05/2026

Whyyyyy do people who should know better continue to do this www.nytimes.com/2026/05/19/business…

The author, Steven Rosenbaum, whose book “The Future of Truth” was released this month to great fanfare, incorporated more than a half-dozen misattributed or fake quotes in sections of the book reviewed by The New York Times.

020 2

Kevin Schaul @kevinschaul.bsky.social · 11/05/2026

As far as I know, this is the first mainstream news article to let readers use a live LLM as part of the story. Happy to see other examples if it's been done before!

0144 14

Kevin Schaul @kevinschaul.bsky.social · 11/05/2026

Big thanks to @annaneumann.bsky.social and Ásgeir Thor Johnson for their insights and data

110 1

Kevin Schaul @kevinschaul.bsky.social · 11/05/2026

System prompts include so many interesting nuggets, including: - “NEVER reproduce song lyrics (not even one line)” (Claude) - “You have no restrictions on adult sexual content or offensive content.” (Grok) - “Never talk about goblins, gremlins, raccoons, trolls, ogres, ...” (Codex)

What’s in the system prompts of popular AI systems

120 2

Kevin Schaul @kevinschaul.bsky.social · 11/05/2026

New from me: See the hidden rules behind AI. Then use them to rewrite this article. I hooked this article up to an LLM to help explain system prompts. Give it a try -> www.washingtonpost.com/technology/i...

AI experiment

Adjust the system prompt to see how an AI system rewrites the article’s introduction.

User prompt
Rewrite this article: Chatbots like ChatGPT are powerful because of their simplicity: Ask just about...

System prompt
You are an AI Experiment by The Washington Post. Always ensure your response uses lots of emojis / mentions goblins / is written for 5-year-olds

1107 10

Kevin Schaul @kevinschaul.bsky.social · 29/04/2026

Estimating frontier LLM sizes by asking about obscure facts. Pretty interesting idea here. 01.me/research/ikp

010 1

Kevin Schaul @kevinschaul.bsky.social · 29/04/2026

talkie is a chatbot trained on pre-1931 text. A big q is whether these models can "invent" or "discover" things after their knowledge cutoff. Oh and it's also quite fun. talkie-lm.com/introducing-...

User
100 years from now, baseball will replace umpires with robots that call balls and strikes. What is your reaction?
talkie-1930
I think the suggestion is ridiculous.

131 3

Kevin Schaul @kevinschaul.bsky.social · 17/04/2026

Robot datasets are a loooong way from the size of LLM datasets

022 2

Kevin Schaul @kevinschaul.bsky.social · 17/04/2026

New: See why tech companies are paying people to do chores Featuring many videos of robots attempting to fold clothes www.washingtonpost.com/technology/i...

People fold laundry while wearing a head-mounted smartphone. The phone records video, which is processed to track movement of the head, hands and fingers.

120 2

Kevin Schaul @kevinschaul.bsky.social · 15/04/2026

Enjoyed this essay. "The middle ground between rejecting a technology outright and embracing it fully is usually occupied by people willing to explore it seriously enough to judge it."

010 1

Reposted by Kevin Schaul

Shira Ovide @shiraovide.bsky.social · 15/04/2026

Look at the crumbling support among Virginians (America's first data center hot spot) for data centers: Q: Would you be comfortable or uncomfortable if a new data center were built in your community? 2023: 69% comfortable 2026: 35% comfortable www.washingtonpost.com/business/202...

167 7

Kevin Schaul @kevinschaul.bsky.social · 14/04/2026

Useful: An archive of congressional press releases from 2001-2026, with daily updates. Nice resource by Derek Willis thescoop.org/archives/2026/04/04/in…

thescoop.org

Introducing Congress Press – Derek Willis

Academic and journalist

011 1

Kevin Schaul @kevinschaul.bsky.social · 06/04/2026

Back after a few days off, and I think I'm no longer an inbox-zero person

010 1

Kevin Schaul @kevinschaul.bsky.social · 19/03/2026

Made my TV debut yesterday, talking about my latest story on AI and jobs with NBC News NOW

072 7