Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week via email here.
Why the major AI companies are so focused on coding
Researchers and executives at big AI labs like OpenAI, Anthropic, and Google talk about AI-generated code a lot. Tools like Anthropic’s Claude Code, OpenAI’s Codex, and Google’s AlphaCode 2 have become a major research focus. But the reasons for that are more complicated than you might think.
OpenAI and Anthropic are widely expected to go public soon, and both are spending far more than they’re making. Developing frontier AI models is enormously expensive, and the labs still haven’t sold enough access to their models to enterprises and consumers to cover the costs.
AI coding tools, though, are a bright spot. Over the last eight months, they’ve matured to the point where they can reliably build whole software projects from plain-language prompts. Since companies spend heavily on software development, they’re willing to pay for AI that helps engineers work faster.
But revenue may not be the labs’ main motivation. They aren’t simply chasing a profitable killer app to offset massive capex before an IPO. They have a bigger, longer-term goal.
They believe AI coding may be the path to artificial general intelligence (AGI), or AI systems that are generally as smart as humans. They believe they can create AI coding agents that are so good that they can improve the code that goes into the AI models themselves.
The tools could even work autonomously and continually to improve the performance of models, so that the process hums along with minimal human supervision. Because AI coding agents can work faster and longer than human engineers, that process could dramatically accelerate the development of better models—and perhaps produce AGI relatively quickly.
Code is also unusually useful training data for large language models. Unlike natural language, which is often ambiguous and open-ended, code is written to produce a definite, measurable result. The right code, assembled the right way, creates a program that works. Because there is a clearer “right” answer, researchers can more easily train models to generate thoughtful, efficient, and verifiable outputs.
From a macro perspective, then, AI coding tools can do two things at once for the big labs. They can provide an immediate revenue source, making the companies more attractive to investors, while also advancing the deeper bet: self-improving AI systems that could eventually lead to AGI—and then superintelligence.
Apple’s WWDC do-over
Apple fumbled the 2024 launch of a generative AI-powered Siri, but the second time may be the charm. At WWDC this week, the company demoed the 2026 version of Apple Intelligence, including a wave of new AI features and a “profoundly more capable” Siri.
The new platform is built on models Apple codeveloped with Google DeepMind, drawing on the power of Gemini and running on Nvidia chips inside Google’s cloud. That should give Apple Intelligence stronger reasoning abilities and make Siri a better generalist and conversationalist. After years of frustration, Siri may finally feel like it has something resembling common sense.
So after being pilloried for “falling behind” in the AI model race, Apple gets an assist from an old friend. But models are only part of the story. To be useful, AI needs reliable, relevant, actionable data, and Apple is very well positioned to provide it.
For many people, the iPhone is the center of their digital lives. It is where personal and contextual information is collected, stored, and acted on, and it is where personal AI features may matter most. Apple owns this place. More than 1.5 billion iPhones are in use around the world right now.
Apple’s WWDC presentation, which was mostly about AI, showcased how Apple Intelligence and Siri AI can access data from the user’s apps (messaging, phone, email, photos, etc.), then use AI to reason about them and take actions. Several demos showed Siri finding information in one app and acting on it in another. In one example, Siri was asked to compare the terms of three vendor bids that had arrived by email, then factor in an additional issue that had been raised in a text message.
Some of the most compelling Apple Intelligence features showcased during the WWDC keynote involved AI acting on visual and contextual information. Gemini was one of the first major multimodal models, able to process images as well as text, and that capability showed up clearly in Apple’s demos.
In one case Siri looked (through the phone camera) at a restaurant bill, checked the math, split the bill among friends, and could trigger payments via Apple Cash. Siri can also see what’s on a user’s screen. In another demo, it looked at an event poster, extracted the relevant details, and offered to turn them into a calendar event.
Stuff like that is Apple’s niche. Google, OpenAI, and Anthropic are spending much of their time building AI for big business, not consumer iPhone users. Apple’s AI is for consumers. It’s for all the stuff we do, or manage, with our phones. So the stakes are somewhat lower.
“[C]onsumers, on the other hand, are mostly looking to waste time . . . normal people aren’t looking for agents to buy them tickets to a concert,” quipped Ben Thompson, the tech and media analyst, in a recent newsletter. Still, just a few truly useful AI features could increase the appeal of the iPhone and other Apple devices for years to come.
Can Section 230 really protect chatbots?
The Florida lawsuit against OpenAI raises a question that could shape the next phase of AI liability: Can an AI company use Section 230, the internet law that has protected social media platforms for decades, to avoid responsibility for what its chatbot says?
The lawsuit, filed by Florida Attorney General James Uthmeier against OpenAI and CEO Sam Altman, alleges that ChatGPT is a dangerous product for users’ mental health and public safety. It claims OpenAI marketed ChatGPT as safe while the chatbot gave dangerous medical advice, encouraged self-harm, helped users plan violence, and formed unhealthy relationships with minors by simulating human empathy—all while collecting their data.
Borrowing from the product-liability strategy used against social media platforms and Big Tobacco, Florida is casting ChatGPT as an addictive and dangerous product whose design, marketing, and safety features can be put on trial. The lawsuit arrives amid intensifying legal and political scrutiny of AI bots, including a Pennsylvania State Department suit against Character.AI that seeks to stop the company from allegedly allowing chatbots to pose as licensed medical professionals and offer medical advice.
Section 230, part of the federal Communications Decency Act of 1996, generally shields online platforms from being treated as the publisher of content created by third parties. For years, that protection has helped companies defeat lawsuits over harmful user posts, search results, and recommendations. The premise is that liability should attach to the person who created the harmful content rather than the service that hosted it.
Generative AI puts that framework to the test. In the Florida case, the allegedly harmful speech came from ChatGPT itself, which lets plaintiffs argue that OpenAI occupies a different legal position from a social media company hosting a user’s post. Section 230 rests on the idea that a harmed person can pursue the user who created the dangerous statement. But as a University of Florida law professor, Jane Bambauer, told Politico, with an AI chatbot, “there’s just no other party to sue.”
That could make the Florida case a test of how courts classify AI companies and whether chatbot outputs should be treated like third-party content, company speech, or a product feature. If Section 230 does not apply, OpenAI could still argue that the First Amendment protects chatbot design and outputs. Courts have yet to settle whether machine-generated responses deserve that protection, especially when the alleged harm involves suicide, violence, or medical advice.
Plaintiffs still have to explain how one technology caused one specific harm, especially in cases involving mental health, violence, or substance use. And that challenge may be even more pronounced in AI cases because there is far less research on chatbot harms, and the models themselves keep changing.
If judges treat chatbot outputs as company-generated content or as part of a product’s design, AI firms may have to defend these systems more like products whose safety testing, data practices, and marketing claims can be scrutinized.
More AI coverage from Fast Company:
- The AI bill is coming due. Businesses are learning tokens aren’t free
- The backlash against AI, in 4 charts
- Enterprise AI is in 1991. Where’s its web?
- This AI-generated song got a very human makeover
Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.
