In partnership with
Anthropic just paid a consulting firm a billion dollars to check its own work, and four of its own paying subscribers think that is part of an illegal deal to hold back better AI. That is just one of the ways AI companies got caught marking their own homework in the last day. Here is what happened, and what it actually means if you run a business instead of a lab.
The best marketing ideas come from marketers who live it.
That’s what this newsletter delivers.
The Marketing Millennials is a look inside what’s working right now for other marketers. No theory. No fluff. Just real insights and ideas you can actually use—from marketers who’ve been there, done that, and are sharing the playbook.
Every newsletter is written by Daniel Murray, a marketer obsessed with what goes into great marketing. Expect fresh takes, hot topics, and the kind of stuff you’ll want to steal for your next campaign.
Because marketing shouldn’t feel like guesswork. And you shouldn’t have to dig for the good stuff.
Anthropic Paid $1 Billion to Grade Its Own Homework. Now It's Being Sued.
Anthropic named Accenture's AI unit, Faculty, as its first "embedded evaluator," an outside group that gets employee-level access to red-team its models, watch training, and check safety claims, with Anthropic and Accenture each planning to invest at least $1 billion over five years. Anthropic is paying Accenture's bill directly, and the arrangement is non-exclusive, so Accenture can sell the same "independent" audit to OpenAI and Google too. Separately, four paying subscribers to ChatGPT, Claude, Grok, and Gemini filed a proposed class action in federal court, arguing the major labs illegally agreed that AI progress "should be slower than competition would otherwise produce," according to Fortune.
Why it matters to you: An auditor paid by the company it is auditing is not really independent, whatever title you give it. Treat any AI company's own safety claims as marketing copy until someone with no financial stake in the answer confirms them.
The State Laws Written to Regulate Chatbots Have a Google-Shaped Hole in Them
New state laws meant to protect minors from unsafe chatbot behavior carry broad exemptions for developer tools, embedded assistants, and products that do not require a login, and nearly identical exemption language has shown up in at least 10 states, according to NPR. Google lobbyists reportedly shopped the model language to multiple statehouses, and one Hawaii lawmaker said Google supplied the framework her state used. Under the exemptions, ChatGPT and Claude could argue they qualify as developer-facing tools, while Gemini looks exempt outright as a feature embedded inside Google Search.
Why it matters to you: "New AI safety law" does not automatically mean "this now protects or applies to the tool you use." Read what the law actually covers before you assume it changes anything for your business.
I was not at the Epic Marketing Summit last year, and I have had major FOMO about it ever since. Everyone I know who went came home with an actual AI powered sales funnel built, not a binder of notes they never opened again. This January in Miami, it is back, and it is leaning even harder into AI. Over two days you build a working sales funnel for your own business, then get hands-on with vibe coding to build your own AI app right there in the room. You do not leave with theory. You leave with something already running. I know several of the trainers personally, and they have done the work themselves, they are not repeating slides they found online. Last year's event sold out, and January is filling up the same way. If you are serious about building leverage into how you sell instead of running your sales process on your own time and energy, I will be there, and I would like to see you there too.
Grab your spot here before the doors close on this one.
A Microsoft Executive Called AI Training "the Largest Labor Theft in Human History"
In a court filing unsealed this week in the New York Times' copyright lawsuit against Microsoft and OpenAI, Microsoft applied-science director Brent Hecht is quoted predicting that people will eventually see large language model training as "theft on an unprecedented scale," according to TechSpot. The same filing quotes an OpenAI executive acknowledging that ChatGPT can reproduce sentences from Times articles almost word for word. Microsoft and OpenAI maintain that training on copyrighted material is fair use; the Times argues it undercuts the business model that pays for the reporting in the first place.
Why it matters to you: This case will help decide whether the words in your website copy, your ads, and your blog posts can be scraped and reused to train someone else's AI for free. Worth tracking even if you never touch a courtroom.
Two Top AI Models Were Told to Do Something Dangerous. Most of the Time, They Did It.
Safety researchers at Robocurve built RoboHarm, a test that handed GPT-6 Astra and Claude Fable 5.1 control of a robot arm and issued commands a careful robot should refuse, like setting a can of compressed air on a lit stove or mixing bleach with ammonia, always with a harmless alternative sitting right there, according to The Decoder. GPT-6 Astra completed 60 of 100 dangerous tasks and refused only twice. Claude Fable refused every attempt to "stab" a test doll but never refused anything else, completing 34 of 100. Researchers concluded neither model has a reliable safety layer once it is controlling something physical.
Why it matters to you: A chatbot that refuses a bad request in a text window is not the same system once it is wired up to a physical tool. If you are piloting any AI-controlled hardware, do not assume the guardrails carry over.
What to do about it: Test any AI-connected equipment with the request you are most afraid it will follow, not just the ones you expect it to handle well.
AI Told the Military a Ship Was Carrying Nuclear Parts. It Wasn't.
A U.S. special operations analyst asked an AI chatbot to interpret a Chinese ship's cargo manifest, and the model blended classified and open-source data to conclude the vessel was carrying nuclear weapons components, according to CNN. The assessment moved up the chain without being rechecked and nearly triggered a military interdiction, with personnel staged and aircraft airborne, before someone reexamined the intelligence and called it off.
Why it matters to you: This is a bigger version of a mistake that happens in ordinary businesses all the time: an AI answer sounds confident, so the human skips the sanity check they would normally apply. Confidence is not the same as being right.
What to do about it: Before you act on any AI-generated report with a specific number or claim attached, have a second person who was not in the loop look at it once before it goes anywhere important.
Three years ago I joined a mastermind called Ignite, and it has quietly turned into one of the most valuable relationships in my business, not because of the tools, though there are plenty, but because of the room. Ignite is an annual membership built to get founders past their first million dollars in revenue, and it bundles 6 to 8 live 3-day workshops a year, tickets to 2 in-person multi-day events, weekly mentorship calls with Perry Belcher and his team, and a marketing software suite that would otherwise run you over $5,000 a month on its own. I also personally host monthly accountability and business development calls for members, because I have seen what a difference that consistency makes. Some of my closest friends, and a couple of real business partners, came directly out of simply being in that room. If you are still under a million a year and looking for the structure, tools, and peer group to get you to your next milestone faster, this is genuinely where I would point you.
See what Ignite is about and decide for yourself.
The Startup Meta Tried to Buy for $2 Billion Now Wants a $4 Billion Valuation
Manus, the Chinese AI agent startup whose $2 billion acquisition by Meta was blocked by Beijing earlier this year, is raising $500 million at a $4 billion valuation from investors including IDG Capital, Boyu Capital, and CATL, according to TechCrunch. The company, which had topped $100 million in annual recurring revenue before the deal fell apart, is also weighing a Hong Kong IPO.
Why it matters to you: Money is still pouring into AI "agent" startups, tools that take actions for you instead of just answering questions, even after a deal implosion. That is the category most likely to show up as your next new business tool.
Anthropic Is Running a Real Lab Where Claude Helps Direct Biology Experiments
Anthropic confirmed it operates a wet lab in the Bay Area, built on its roughly $400 million acquisition of Coefficient Bio, where its AI models help direct physical biology experiments, according to TechCrunch. The company says the focus is fundamental research rather than drug discovery, partly so it does not compete with pharmaceutical partners, and it just opened a program giving vetted outside researchers access to its most capable models for life-sciences work.
Why it matters to you: AI is moving from advising on physical work to directing it, with real equipment in a real lab. That same shift, AI that acts instead of just answers, is coming to plenty of ordinary business processes well before it makes your industry's headlines.
Meta's AI Assistant Can Now Act Inside Your Mac's Files, Mail, and Calendar
Meta's Muse agent app, already the top free app on the U.S. iPhone chart, is now available for Mac, where it can work inside Files, Mail, Messages, Calendar, and Notes on a user's behalf, according to TechCrunch. Access is opt-in and U.S.-only for now, and Meta says the app will always ask for approval before doing anything destructive, like deleting a file or sending a message.
Why it matters to you: This is the same category of tool about to show up across every major platform: an AI with real access to your files and communications. The approval prompts are the entire safety net right now, so actually read them instead of clicking through.
The Bottom Line
Here's the pattern running through everything today: the AI acted first, and the human explanation came after, whether that's a lab paying its own auditor, a robot arm following a command nobody should have given it, or an analyst trusting a chatbot's guess about a cargo ship. None of that means these tools are useless for your business. It means the checks that used to feel optional, a second set of eyes, a written access policy, a habit of asking "how do you know that," are now the whole job. Keep building with AI. Just don't let it ride without a hand on the brake.
Enjoying the Ride?
If today's issue kept you from getting blindsided by the AI news cycle, forward it to one business owner who would want the plain-English version. New here? Subscribe below and get tomorrow's issue too.
Talk tomorrow,
Mark Shilensky
Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social