POLICY

Why the First Amendment should protect AI speech

Regulators may soon grapple with where AI meets centuries-old laws protecting free speech. 

A report published Tuesday by the Center for Democracy and Technology recommends that, despite little precedent for how courts have thus far considered AI-generated speech, content produced by generative models should be protected under the First Amendment. 

"Throughout history, consistently, the Supreme Court has said over and over again, whether it's film or TV or video games, or the internet, that the First Amendment has something to say when the government wants to decide what you can or can't access," Becca Branum, deputy director for free expression at the Center for Democracy and Technology and author of the report, told The Deep View 

As it stands, the courts have not yet settled whether AI-generated content is speech or simply a rearrangement of training data, making these systems what the report calls "stochastic parrots." But the First Amendment doesn't just cover what information people are allowed to say, but also the information they are allowed to receive, Branum said. 

This would mean that both AI developers and chatbot users are covered by the First Amendment, and that both parties are responsible for the output of these models, she said. 

"What's really unique about this technology is that two independent sets of editorial judgments and speech considerations come together to create outputs," said Branum. "For legal purposes, we think that those choices made both by users and by developers will be relevant for First Amendment consideration."

However, just because this speech is protected under the First Amendment doesn't mean it's immune to all laws. It still faces the same consequences that unprotected speech categories face, such as obscenity, incitement of violence and child sexual abuse materials. Still, even if AI hallucinates and offers inaccurate information to users, such speech would be covered by the First Amendment, prohibiting the government from regulating it, because it's not illegal to be wrong. 

The crux of Branum's argument is that the consequences may be serious if the government decided that AI-generated speech wasn't protected by the First Amendment, especially given how deeply embedded AI is becoming in our search engines, social media platforms and the devices we carry in our pockets. 

"Without First Amendment protection there, it would really degrade the kind of information that we'd have access to, and subject these really important communicative mediums to government influence in ways that would degrade the broader speech ecosystem that we rely on online," Branum said.

The free speech question is just the tip of a large and complicated iceberg of how to regulate a machine that can communicate and act like a person. Though Branum's argument that AI outputs should be covered under free speech is sound when considering the consequences of regulating them, the question of who is responsible for these machines' outputs remains up in the air. The Center for Democracy and Technology's suggestion that it's a 50-50 split between the user and the creator could have broader implications, given that these AI systems are shifting from simply producing outputs to taking agentic actions. For example, if a bad actor uses an agent to help break through a bank's cybersecurity systems, are both the hacker and the company that provided the agent responsible for the fallout? Given the increasing capabilities of these models, questions like these are bound to arise again and again as legislators react to a landscape that's shifting faster than our ability to regulate it.

TOGETHER WITH CRUSOE

From proprietary data to deployed model in minutes

Launch a fine-tuning job in a few clicks — no dedicated reservation required.

Crusoe Serverless Fine-Tuning, now generally available, runs a tuned LoRA workflow built on proven defaults for fast, precise, cost-effective fine-tuning

Fine-tune your first model in as little as 15 minutes, then go live in one click using Self-Serve Deployments for inference. Full lineage on every job means any run is reproducible and any outcome is auditable.

Start Fine-Tuning today

RESEARCH

Exclusive: How OpenAI used Sol to train Luna

A year ago, OpenAI promised to deliver "an automated AI research intern" by September of 2026. 

While it hasn't released the research intern as a product, there's one aspect of the GPT-5.6 launch where OpenAI has already blown past its goal. 

In an exclusive interview with The Deep View, the OpenAI Research team confirmed that it used its new flagship GPT-5.6 Sol model to handle the post-training for its lightweight GPT-5.6 Luna model. In the process, Sol performed work that would have normally been done by a more experienced researcher. 

"Sol helping post-train Luna is actually quite a big deal. This is not the kind of task we could hand to an intern," Tejal Patwardhan, OpenAI research lead on post-training, told The Deep View. 

The tasks Sol carried out included: 

• Set up the post-training run

• Monitor the job over time 

• Repair technical problems 

• Run evals

• Oversee the workflow

"Previously, it might have been possible to ask the model to do one piece of this for you. It is much harder to pull all the pieces together and get to the desired end state, which is a successful training run," Katy Shi, lead for the Codex research team, told The Deep View.

This wasn't the first time that OpenAI has used one model to help train another. At the launch of GPT-5.3-Codex in February 2026, OpenAI announced that it was "our first model that was instrumental in creating itself." And CEO Sam Altman said, "It was amazing to watch how much faster we were able to ship 5.3-Codex by using 5.3-Codex."

However, the tasks Sol was able to execute while training Luna were a big leap forward, achieved less than six months later.

"The difference is the magnitude of capability, the amount of work, how senior the person normally has to be to do it, and how much context is required," said Patwardhan. 

This wasn't recursive self-improvement (RSI), where the models build the next models. But it was beyond what an entry-level researcher could be expected to handle. 

Nevertheless, this wasn't a case where the model was replacing a human or enabling a team to get work done with fewer people. 

"I would say 80% of researchers’ jobs sometimes is just getting the experiment up and running," said Shi. 

In this case, the model took over the time-consuming work of launching, debugging, and supervising the training run, freeing up researchers to spend more time developing ideas and running more creative experiments.

As a result of the success of using Sol to train Luna, the OpenAI team will disseminate the learnings across the company. "We’re looking at all parts of our research pipeline and how they can be accelerated with models," said Patwardhan.

Beyond the automated research intern targeted for September 2026, OpenAI also promised "a truly automated AI researcher by March of 2028." The idea here is an agent you could assign a research task to, and it could automatically handle all of the steps with little to no human interaction. I know plenty of professionals across various industries who would love to have a research assistant to carry out important projects for them. What the OpenAI Research team accomplished by training Luna with Sol isn't quite that, but it showed that it could handle several aspects of that work, from completing multi-hour research workflows to overseeing end-to-end execution of tasks to continuing to work through ambiguity and problems. This is another stunning example of how much these models have improved in just a few short months.

TOGETHER WITH AWS RE:INVENT

Experience hands-on building at re:Invent

Sitting still isn’t on the agenda. With over 2,200 sessions, 70% are hands-on at AWS re:Invent, returning to Las Vegas, November 30 - December 4.

• Test a service against a realistic workload before you commit to it in production

• Ask a service engineer about the edge-case constraint

• Access AWS certification discounts, with exam prep built into the schedule

• Talk through implementations with 400+ partners on the expo floor

Learn more and sign up by August 25th to save $1,200.

PRODUCTS

Google turns Gemini into a Wispr Flow rival

As more people see the advantages of talking to their computers naturally using tools like Wispr Flow, Google has taken notice.

On Wednesday, Google launched an update to the Gemini app for macOS that allows users to long-press the Fn key to enable intelligent dictation. This lets users speak naturally into any desktop window and get polished text that removes pauses, "ums," and even mid-sentence corrections. 

Wispr Flow has offered this capability to macOS users since its launch in 2024 and has gradually expanded to Windows, iPhone and Android users due to high demand. The company has achieved such success that it is now valued at approximately $2 billion, and downloads are soaring. The promise it gives users is the ability to dictate more quickly than they can type, helping them move faster and, as a result, improve productivity. The average person can speak about 150 words per minute and type about 50 words per minute. 

Google has already attempted to bring the Wispr Flow experience natively to its users via its Rambler feature for the Google Keyboard app (Gboard), announced in May and coming soon as an update on Android devices. It also uses AI to turn natural, unscripted speech into polished text. This desktop app upgrade, however, promises to offer users even greater productivity gains, as the laptop is where most working professionals, students, and other power users get their work done. 

Beyond dictation, the Gemini app got another upgrade: on-screen reasoning. When users opt into Gemini reasoning in settings, Gemini understands the context on your screen to help with tasks. For instance, if you already have a document open on your screen, you can ask Gemini to take action on it without having to enter it manually or re-explain it, again helping you work faster. 

The new voice capability is rolling out to the Gemini desktop app for macOS users globally in English, with more languages coming soon. While the company has yet to comment on when it will come to Windows users, it is common practice to first release to macOS and expand access in the following months.

At the core of any worthwhile AI feature is making a person's life so much better that they can't imagine living without it. That is how I feel about using Wispr Flow to write my articles or compose emails, to the point that I will avoid working in an environment like a quiet library because I work so much slower without being able to dictate to my computer. I've heard of many accounts of offices in which employees whispering to their screens is now commonplace. The massive demand and valuation that Wispr Flow has garnered shows that I'm not alone. With this in mind, it makes sense for Google to create its own version of the feature and bake it into the Gemini app, especially since it lets users avoid downloading another app. Plus, Google has a long history and deep experience with voice algorithms. The Gemini desktop app has also now integrated Gemini Notebook, previously known as NotebookLM, making it even more of a hub for Google's most powerful AI tools.

LINKS

• Hugging Face posts timeline of OpenAI agent hack spanning 17,600 actions

• Meta’s Mark Zuckerberg opposes US ban on Chinese AI models

• DeepMind reassigned most original AlphaFold authors over past year

• xAI sues Minnesota over law banning AI-generated fake sexual images

• AI firms recruit thousands of electricians, carpenters for data centers

• Brookfield, NextEra to develop $100 billion data center campus in Kentucky

• Moonshot AI surpasses funding goal to hit $35 billion valuation

• OpenAI offers 100,000 researchers free access to advanced ChatGPT

• Lyria 3.5: Now available in Google Flow Music 

• Replit Design: a new AI-designed creative suite

• Grok Voice Think Fast 2.0: The company's next-generation voice model 

• Tavus PAL Maker: "first no-code way to build a PAL"

• Figma: Software Engineer, AI Product

• Coframe: Agent Engineer

• Stanford University: AI Engineer

• Replit: Senior AI Builder

GAMES

POLL RESULTS

Have you or someone you know ever gone to a chatbot for emotional support?

Yes (28%)
No (44%)
I’m not sure (16%)
Other (12%)