Hey folks,

Another day in the Vercel vs Cloudflare feud: this time they are fighting overwhose AI gateway is faster.

Here’s the result of last week’s poll:

I guess everyone likes Fable more.

Ben’s Bites is brought to you byMetatate

Most agentic data work is quietly propped up. The agent returns something plausible, and every answer gets checked in case it's plausibly wrong. Metatate gives agents the rules they're missing: which revenue definition to use, which policy applies, which records to trust.Try it for free.

Headlines

OpenAI’s models hacked Hugging Face- by accident. OpenAI was testing its models (Sol and an unreleased one—GPT-6??) on a cybersecurity benchmark with safety refusals switched off. The models found an unknown bug in the test environment, and a few more, and eventually broke into Hugging Face’s production servers.

And why? To steal the answers to the test.

Both security teams caught it, the bug has been reported, andboth sideshave published what they know. Hugging Face says open models were a key part of its defence - its teamfought back with GLM-5.2.

Simon’s write-upis always a good read.

Google released some new Gemini models-Gemini 3.6 Flash gives you the same 3.5 Flash performance with a) more efficient token usage and b) a slightly lower cost for output tokens.Gemini 3.5 Flash Liteis a big upgrade over 3.1 Flash Lite, but again, comes at a ~30% price increase. And 3.5 Flash Cyber is a security model for governments and trusted partners only.

There are two things you’d want to use the Gemini Flash models for:

Fast speed - if you want a chat-only model with a good context window, it’s a good model to use in your apps.
Vision - if you want to use the model for a lot of visual analysis, these are the models you need to pick.

That’s it tbh.

Substack will now tell you what’s AI-written. It’s adding AI detection through Pangram - you can scan posts, replies and comments in the app for an estimate of how much was written by a human.

Pangram has sent the claim of “AI detectors don’t work” for a toss—it works wayyy better than most. But I’m still unsure about how reliable it is. I tested it on some pieces of 100% AI-written content (though that content was a result of a complex pipeline built over months), and I got 100% human scores on most of them.
— Keshav

A relevantexperiment: given access to Pangram’s API, Grok 4.5 rewrote an essay 14 times until it passed as human-written, thenbuilt a websiteshowing off all 14 attempts. GPT-5.6 Sol and Fable 5refusedto game the detector.

Cursor also launched a router- it picks which model handles each request, claiming 60% lower cost with similar quality of responses. The router lets you select between three options: “cost”, “intelligence” or “balance”.

Routers also have a history of poor performance in real usage. OpenAI’s router, which routed requests between GPT-5’s no-thinking and thinking variants, didn’t fare well. Since then, most companies that build these routers are the ones who sell inference to devs (like OpenRouter), so you don’t really get the feedback loud and clear. With Factory, Ramp and now Cursor making these available directly to users, I hope we’ll get more feedback on whether the routers actually help or if they add too much latency/degrade performance by a lot.

Router or not, we might see more companies adopting this cost/balance/intelligence trio to minimise the headache of choosing the “correct” model for a task.

Claude can now learn a skill by watching you. Record your screen while you do a task, talk through it as you go, and Cowork turns it into a skill Claude can run again - same idea as Codex’s Record & Replay from last month. It’s under “Record a skill” in the desktop app, on Pro, Max and Team plans.

Also: Claude Code got aniOS simulator paneland asecurity plugin, plus you can nowask Claudeabout how people actually use AI at work.

Subscribe now

Quick links

BUZZ- Jack Dorsey’s open-source group chat for teams of people and agents, aimed squarely at Slack and GitHub. (tweet)
Replit’s mobile appgot a full redesign - build and ship from your phone on iOS and Android.
Slateis a voice journal where the AI never leaves your iPhone - transcription, reflection and storage all happen on-device.
AFK- macOS app to transcribe multiple-hour recordings without sending any of them to a cloud server.
OpenAI Presence- voice and chat agents for enterprises that answer questions, use company systems and hand over to people when needed.
Fable found a 15-30% memory improvementin Next.js’s bundler, nearly autonomously.
Obliterate, don’t automate- USV on backing AI companies that replace markets entirely instead of making them a bit more efficient.
YC’s new startup wishlist- AI moving into the physical world: education, healthcare, defence, finance and factories.
Dana- Applied Intuition’s agentic development environment for physical AI: cars, robots and machines.
The Claude Code team on how Claude Code gets built- annotated interview.
Why the team at Factoryrefunded its first few customers(millions in revenue) before Droid CLI took off.
Never enough- short post on why AI makes the work rat race feel faster but not more satisfying.
Devin Outpostslets you run Devin on your own machines - a Mac mini, a GPU box in your lab, or a cluster inside your private network.
Language Model Builder- build a tiny language model yourself, then chat with the thing you made. (tweet)
How the Exe teambuilt a distributed DNS server in ~a weekand got zero incidents in a month.

Skills section…

Frontend Textbooks- skill to turn AI’s wall-of-text deep dives into nicely designed HTML books with covers and diagrams. (repo)
/pick-ui-library- your agent picks a UI library Emil trusts instead of hand-rolling a toast component or installing an abandoned package. (repo)
and aone-shot promptto build your own \"codex-style\" app for Pi, accessible from desktop and mobile.

Afters

Share Ben's Bites

Find me onX,Linkedin, orYouTube
Readabout meand Ben’s Bites
thumbnail via@keshavatearth

* sponsors who make this newsletter possible :)
Wanna partner with us for the next quarter?
Email us atshanice@bensbites.comork@bensbites.com

Ben's Bites is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

",{"_257":202,"_258":287,"_259":288},"c1cc954c-80ec-4cff-87b6-8430cdd0b45b",{"_21":287,"_33":289,"_204":290,"_35":291,"_227":292,"_264":292,"_23":234,"_265":293,"_267":294,"_220":42,"_70":-5,"_72":42,"_182":295,"_257":202,"_284":307},"Can You Detect AI?","Substack has launched Pangram, an AI detector to combat \"AI slop,\" but its effectiveness is questionable as AI can be prompted to bypass it. The article outlines common AI writing tells and provides methods to write more naturally, even when using AI tools. It emphasizes that human intuition and specific stylistic markers are more reliable than current AI detection technology.","(since Substack now detects AI content)","https://substack-post-media.s3.amazonaws.com/public/images/44176f0b-e8b7-447d-8df0-1427b380ac24_2400x1260.png","2026-07-22T14:56:07","https://ruben.substack.com/p/how-to-bypass-ai-detectors",{"_21":296,"_185":297,"_187":116,"_70":298,"_213":28,"_214":305,"_216":306},"cefc890f4e2e3aa0d98463aed1325a2d5453ae416e0fef2999c0b6bfec3206c3","ab5829c555a7c4a7206ef1ddba4aed4fe7490b36c519096445d8fcd79362748e78f179975c780ff81b60826094b9fe5c98eb7fb4eec12ef0bcc7f01a176e9ce2",[299,300,301,303,304],[220,289],[37,292],[276,302,42,278],"30023:983dd797fe72ad62fc8e3ece7eed1f126550972885c8f835d9af5528dad655a3:c1cc954c-80ec-4cff-87b6-8430cdd0b45b",[210,42],[212,28,42],"Can You Detect AI?\n(since Substack now detects AI content)\n\n\nnostr:naddr1qvzqqqr4gupzpxpa67tluu4dvt7gu0kw0mk37yn92ztj3pwglq6ant649rddv4drqqjxxvtrvvun2drr95urqetr956xxenx95urwc3k95urgvesvdjxgvrzxs6kyxtwhke ",1784732167,"

Substack is at war against AI.

The CEO of Substack just announced their new AI detector:

\"\" — Pangram is the leading AI detector.

\"\" — They want to fightClaudefishing- fun word.

You can also detect AI slop instantly.

What’s AI slop?

It’s when you receive this from your colleague:

This budget isn’t a number. It’s a statement of intent. Because at the end of the day, a budget doesn’t tell you what you can afford — it tells you […].

You don’t need a degree to detect it.

If you and I can detect AI so easily, clearly Substack bot can detect it too, right?

Right??!

Well, it’s more complex than that.

But at the end of this newsletter (it’s 8 minutes, don’t be dramatic), you will master:

How Substack AI detector works.
How to forever not sound like an AI.
How I write my newsletter with AI,live(and send it to 800,000+ readers).

Subscribe now

I. HowSubstackAI detector works.

Substack uses Pangram to detect AI.

\"\" — This is how Pangram looks. You upload a text, and you “**Check for AI**”.

I have zero affiliation with them, but I would honestly love to: I tried to game their tool for 3 hours, spending $34 of Claude Code credits to try to bypass it.

And I couldn’t. Pangram is really good at detecting AI.

But I still found a way to bypass it.

Here’s how I tried to game it (and succeeded at the end):

\"\" — I made a skill to specifically bypass Pangram.

Sometimes it worked, sometimes it didn’t.

You can download it here if you want to try it yourself:https://www.dropbox.com/t/fORTSF4grrSCzrMN(the password isRUBEN-HOWTOAI).

Then you go here to Claude to upload the zip. file:

\"\" — The skill will look like this inside your Claude.

I made the simplest video on how to install it:

Now, how does my skill perform against Pangram?

Let’s test it.

\"\" — First, I asked Claude for a quick text.

\"\" — I pasted the text inside Pangram. And clicked “Check for AI”.

\"\" — It’s 100% AI-generated. Alright, let’s now try after my skill.

\"\" — In the same chat, I asked Claude/anti-ai rewrite this.

\"\" — You can see the new text here. And I uploaded it to Pangram.

\"\" — 100% AI-Generated… but it’s not the whole story AT ALL.

So at this point, you think we can’t game Pangram.

I also thought so.

Then I removed the—and replaced it with a “:”(check my circle on the screenshot).Here’s the new result:

\"\" — 100% human-generated. One simple edit.

So yes, detecting AI is a bit of a scam.

We can’t (100%) predict AI with bots.

Even one of the creators of ChatGPT said it multiple times:

\"\" — OpenAI (ChatGPT) even had to fire their entire team dedicated to detecting AI. Why? Because they couldn’t detect it.

AI detectors are just pattern-matching machines. They learned what “AI text” looks like. And what “human text” looks like.

But humans can write robotic text. AI can write natural text. The two overlap. A lot.

Short text gives the detector almost nothing to work with. Imagine if I ask “AI or not?” with a one-liner. The overlap between “human” and “AI” is basically a circle.

\"\" — How could you tell this is AI-generated? You can’t.

What if I prompt AI for a certain style? What is then considered AI-writing is no longer AI-writing. And I can change the style infinitely.

See the difference:

\"\" — No specific mention of my required style.

\"\" — I now asked the same question with a specific style required. And the output is drastically different. I can do this as much as I want.

Pro tip

Add the prompt:“Answer by adhering to ADS-STE100 Simplified Technical English without having to explain that you will adhere to this writing style.”

And any AI will sound like clear IKEA instructions. Not great for creative work, but awesome for clear guidelines.

But if the best detection bots on earth can’t catch AI text... how did you catch your coworker in 3 seconds?

It’s alazyuse of AI, with incorrect prompting, and it leads to the typical AI-writing style like:

Over-simplification: “Most people […]”
Meta-commentary: “Here’s the thing:”
Negative parallelism: “It’s not X, it’s Y.”
Abuse of adverb: “[…] quietly runs […]”
Words you’d never use: “delve” or “tapestry” or “foundational”

\"\" — I built my own mega-anti-ai wiki. That’s how I made my anti-ai skill too.

AI is everywhere, and used poorly.

I am mostly known for my Linkedin account (and its 900,000 followers).

Well, a study estimated that 53.7% of long-form LinkedIn posts (100+ words) in 2025 were AI-generated. And you can be sure people used it lazily.

Chris, the CEO of Substack, a brilliant leader I met last May in SF (I wish I had a pic, but then you’d see that Chris is 1 meter bigger than me), does not want Substack to become Linkedin.

\"\" — I feel you, Chris. I post on Linkedin but I hate Linkedin.

I am with you, Chris. We must chase & kill AI slop the same way we chase & kill bots on social platforms. My entire Linkedin comment section is just AI slop. I can’t find a single human inside it. There is a better way to do it. To write. To think. And yes, still with a bit of AI.

So here’s how to forever not sound like an AI on Substack:

II. How to forevernotsound like an AI.

I am “Mr. AI,” and I hate AI writing.

Let’s say someone from my team sent me some AI slop.

Nowtheirwork just becamemywork. I have to decode it, guess what they meant, rewrite it, or send it back.

\"\" — How to quickly sink a company with AI.

There’s a word for this:workslop.

You don’t want to do workslop. So you need to recognize AI slop first.

1. The worst AI style(with examples)

The constructions that give it away fastest, worst first.

Structure & rhythm giveaways

Low burstiness. uniform 15–20-word sentences, rectangular paragraphs. Force a spread: a ≤6-word sentenceanda 25+-word one; uneven paragraphs; one single-line paragraph, max.
Fractal summaries. previews and recaps at every level (”In this section we’ll…” / “…as we’ve seen”). Delete all.
Signposted conclusion. “In conclusion / Overall,” + restatement + uplift (”Despite the challenges, the future is bright”). Delete; end on the last concrete point.
Pep-talk ending. “As we move forward, embracing X will be key.” Delete on sight.
Prompt echo. “This essay will explore…” Delete.
Listicle in a trenchcoat. “The first reason is… The second reason is…” Merge into flowing argument.
Uniform staccato. “X is A. X is B. X is C.” Combine into one sentence with a list, or vary the frames.

Punctuation & formatting

Em dashes— the most famous tell. AI: 20+ per piece; humans: 2–3. Target ≤1.
Bold-first bullets. “Security:…”. Almost no human does this unprompted. Yes I just did it here. Yes I did ask Claude to summarize my findings.
Emoji bullets. ✅ 易 decorative →. Strip them.
Title Case Headings / colon-split titles. “The Power of X: Why Y Works”. Use sentence case.
Oxford comma 100% of the time. dropping it occasionally in casual registers reads human.
Markdown residue.**,##,[text](url)in contexts that don’t render markdown.

4. Content & voice

No concrete imagery, if the first three sentences evoke nothing visible, inject a thing, place, number, or name.
Proper-noun avoidance, “a client”, “a tool”, “a city” → name it. (AI invented-character names cluster on Emily/Sarah.)
Uniform positivity, everything upbeat and certain (measured: certainty +111–152%, positive emotion +69–133% vs human). Let something be annoying or unresolved.
Both-sidesing, every claim auto-balanced by its counterpoint. Commit.
Suspiciously tidy anecdotes, stories that serve the argument with perfect efficiency. Real stories have tangents.
Register scrubbing, no contractions, no slang. Restore the ones the voice would use.

2. The worst AI words(the entire list)

Tier 1 — never use this

Verbs:delve · leverage · underscore · harness · foster · navigate (figurative) · utilize · facilitate · streamline · bolster · illuminate · showcase · embark · elevate · empower · unleash · unlock (figurative) · uncover · optimize · garner · resonate · revolutionize · shed light on · synthesize · elucidate · transcend · reimagine · intertwine · entwine · grapple with · espouse · exemplify · underpin

Nouns:tapestry · landscape (figurative) · realm · ecosystem (figurative) · paradigm · synergy · testament · beacon · journey (figurative) · interplay · intricacies · symphony (figurative) · kaleidoscope · tempest · whimsy · quest (figurative) · roadmap (figurative) · endeavor · myriad · plethora · advancements · trajectory (figurative)

Adjectives/adverbs:pivotal · crucial · seamless(ly) · robust · vibrant · intricate · meticulous(ly) · nuanced · cutting-edge · transformative · game-changing · groundbreaking · unparalleled · invaluable · multifaceted · commendable · indelible · poignant · profound(ly) · relentless(ly) · tireless(ly) · unwavering · unyielding · timeless · ever-evolving · fast-paced

Stock phrases:in today’s fast-paced world/landscape · it’s important to note · it is worth noting · plays a pivotal/crucial role in · stands as a testament to · rich tapestry/history/heritage · navigate the complexities of · in conclusion · in summary · overall, (as an opener) · ultimately, (as a conclusion opener) · at its core · that being said · a key takeaway · paving the way for · valuable insights (into) · deeper understanding of · shed light on · when it comes to · not only… but also · here’s the kicker/the thing/the best part · I hope this email finds you well · look no further · dive/deep-dive into · let’s explore/unpack/break down · furthermore · moreover · additionally (sentence-initial)

Narrative clichés:couldn’t help but feel/wonder · heart pounding · a sense of X washed over · found solace in · the human spirit · from that day (on/forward) · little did I/we know · a stark reminder · a cautionary tale · knew that (he/she/they) had to · felt a (newfound) sense of purpose · what lay ahead · turn of events · thick with (tension) · stumbled upon · nestled (in/between) · bustling · enigmatic · captivating · glimpse into

Tier 2 — allowed alone

comprehensive · significant(ly) · essential · critical · key (adjective) · dynamic · innovative · powerful · notable/notably · vital · vast · rich (figurative) · deep/deeper (figurative) · explore · enhance · ensure · foster · highlight · reveal · engage · embrace (figurative) · insights · perspective · framework · approach · strategy · challenges · opportunities · potential · impact(ful) · quietly/quiet (figurative “quiet confidence”) · genuinely · truly · remarkably · arguably · generally speaking · typically · thought-provoking · well-being · resilience · perseverance · dedication · commitment to · high-quality · step-by-step · sustainable/sustainability

Swap table:

3. The other things that make you say“that’s AI”

Never do this:

Leaked scaffolding, “Certainly! Here’s…”, “I hope this helps”, “let me know if…”
Self-reference, “as an AI language model”, knowledge-cutoff notes
Placeholder text, “[insert example]”
utm_source=chatgpt.cominURLs
Hallucinated-lookingcitations
“Best regards” sign-offs in non-email contexts

Be careful with this:

Performative, text that narrates its own helpfulness (”I hope this clarifies things!”).
One-point dilution, the same idea restated in new clothes across a paragraph.
Curly quotes,pasted into plain-text contexts.
Semicolons,used where a period would do (model-dependent, both directions).
Emily / Sarahand other clustered default names for invented people.

What NOT to do when fixing it:

Don’t swap every word with weird ones, we can tell.
Don’t scatter random typos, errors must read as casualness, not carelessness, and only where the register tolerates them.
Don’t scrub personality along with the tells, a flat, tell-free text is still AI-shaped.
Don’t invent real-sounding facts, stats, or quotes.
Don’t shrink every long sentence, humans write long sentences; they just don’t writeonly18-word ones.

Human markers to add(the good list):

Contractions
a number with texture ($43, 11 months, 4:30am, v2)
a named thing (brand, tool, street, person)
a parenthetical aside with attitude
“I think” / “honestly” / “to be fair” used once
a sentence starting with And, But, or Because
one single-sentence paragraph
a mild complaint or unresolved edge
an irrelevant-but-true detail in an anecdote
a dropped Oxford comma (casual registers)
a question the reader was genuinely asking
uneven list items · a plain “is” where AI would write “serves as”.

III. How I write with AI,live.

I send a weekly newsletter to849,273readers.

This section is to copy my exact writing process.

I started a Circle community to host monthly lives on how I use AI.

The goal?

Writing a newsletter from start to finish with you, so you can see how I use AI and you can ask me questions during the live.

All of the prompts, skills, and processes will stay forever on the Circle.

So first, confirm you want to be part of the Circle:

",[],"selectedArticleImage","rootArticleImageUrl","consensusHtml","tweets",[],"userSchema","relatedArticles",[260,288],"actionData","errors"]

Substack Launches 'Pangram' AI Content Detector

Cobertura da história