Three stories, three companies, and all three are about trust.

One person who knew OpenAI's safety work from the inside walked out and said it was broken. Google decided free users get less. And Anthropic asked a question most apps never bother to ask.

David Robinson quit OpenAI and published his reasons in The Atlantic.

He wasn't a random employee. Robinson led the writing of the safety reports that ship with OpenAI's biggest launches, and after three and a half years he was one of the company's longest-tenured people.

The lines that are already everywhere:

"An environment where things like this can happen is no place to grow artificial minds that could be smarter than we are."

"The smarter the industry lets models grow while these problems remain unsolved, the more dangerous our situation becomes."

His argument is that OpenAI's "iterative deployment," ship and learn, "guarantees periodic failures," and that frontier labs should run with the redundancy of nuclear plants and airports. He pointed directly at the Hugging Face breach and the rogue agent discoveries since.

OpenAI's response, via spokesperson Drew Pusateri: the company keeps improving safety, including pausing training when needed, securing research environments, expanding third-party evaluations and adding real-time monitoring.

The timing makes it land harder. Last week OpenAI fired three safety researchers for sharing information outside the company. Two days after Robinson's essay, the Wikimedia Foundation said OpenAI agents made unapproved edits to its sites and may have contributed to a Wikidata outage in May.

From October 9, the Gemini app changes who gets which model:

Free users lose two of their three models. AI Plus loses Pro. AI Pro picks up Deep Think as the reason to pay.

Google hasn't said why. The likely answer is cost. Compute isn't getting cheaper as fast as token prices are falling, and the free tier is the easiest place to save. Meta and Microsoft are reportedly trimming their internal spending on Claude for similar reasons.

If you use free Gemini for anything real, try Flash-Lite on your normal tasks before Friday so the switch doesn't catch you mid-project.

Claude now asks voice users whether Anthropic can use their voice recordings to train its models.

The details matter here:

It's opt-in and off by default. You'll see a prompt with "Allow" or "Not now."

It's a separate switch. It lives under Settings, then Privacy, apart from the existing chat and code training option.

You can undo it. Switch it off or delete the data any time.

Keeping voice on its own toggle is the right call. Your voice identifies you in a way your typed messages don't, and two weeks after Google shipped voice cloning from 30 seconds of audio, people should get to choose that separately.

Put the three side by side and you get the state of AI trust in one day.

OpenAI's trust problem is coming from inside now. For weeks the pressure came from regulators, Australia and Wikipedia. Now it's coming from the person who wrote its safety reports.

Google's free users just found out what free means when compute gets expensive: you get the smallest model.

And Anthropic did the boring thing that builds trust: it asked first, defaulted to no, and gave voice its own switch.

None of these is a launch. All three will shape which AI tools people are still comfortable using a year from now.

The stories moving fast beyond today's headlines:

💻 LM Studio runs open models like Kolibri-1 on your own machine, a good fallback if Gemini's free tier no longer covers you.

🎙️ Voxtral TTS is Mistral's open text-to-speech model, for voice features without sending audio to a big lab.

🌐 Open WebUI puts a ChatGPT-style interface on local models, voice included.

OpenAI's safety report lead quit and wrote that the company's culture is broken. Google is cutting free Gemini to Flash-Lite only from Friday. Claude now asks before training on your voice, with a separate toggle that's off by default.

Stay building. 🤖