China's government-led push to outpace the U.S. in generative AI is hitting speed bumps created by the Chinese government's need to control political speech, according to new reports.

Why it matters: Even as the U.S. tries to restrict China's access to high-end chips and hardware sold by its allies, the demands of China's authoritarian system could prove more decisive in tipping the global AI race America's way.

Driving the news: Several stories in the Western press this week detailed the drag on China's AI progress created by the country's chief regulator, the Cyberspace Administration of China (CAC).

• The Financial Times reported Wednesday that the CAC is requiring elaborate reviews of the AI models developed by China's tech giants and startups — including ByteDance, Alibaba, Moonshot and 01.AI.
• "China's operational guidance to AI companies published in February says AI groups need to collect thousands of sensitive keywords and questions that violate 'core socialist values', such as 'inciting the subversion of state power' or 'undermining national unity,'" the Financial Times notes.
• Per the Wall Street Journal, "The internet regulator requires companies to prepare between 20,000 and 70,000 questions designed to test whether the models produce safe answers, according to people familiar with the matter. Companies must also submit a data set of 5,000 to 10,000 questions that the model will decline to answer, roughly half of which relate to political ideology and criticism of the Communist Party."

Case in point: China's chatbots — like its search engines and social media spaces — can't talk about the Tiananmen Square uprising of 1989 or question the legitimacy or policies of President Xi Jinping.

• The bots are designed to refuse to answer politically controversial queries. If users ask too many such questions in a row, the systems must end the conversation.

Between the lines: China's AI firms arrived late to the generative AI game to begin with, and having to toe the government's political line is making it harder for them to catch up.

• Training their models is slower when they have to scour data to remove politically sensitive information.
• And no matter how much effort they put into sanitizing their bots' output, the random and unpredictable nature of generative AI means the government can never be 100% certain they won't stray into sedition.
• Meanwhile, all the development time and computing resources companies put into building these ideological guardrails isn't available for making the AIs faster and more useful.

Catch up quick: China spearheaded a previous wave of AI development focused on image recognition and computer vision — technology that the government put to repressive use in the restive Xinjiang region.

• After the advent of ChatGPT kicked off a generative AI frenzy, Beijing started orchestrating a national push to match and overtake Western expertise in the field.

China was the first major nation to create detailed regulations governing the deployment of AI.

• Those rules, instituted a year ago, protect citizens' privacy and data from abuses at the hands of the companies that make AI — but not from government power.

The big picture: In the U.S., the Biden administration's AI executive order has mandated relatively modest transparency rules for companies working at the cutting edge of AI capabilities.

• U.S. tech firms' efforts to patrol AI-generated misinformation and limit chatbot hate speech have been characterized as censorship by conservative critics.
• Meanwhile, the EU is finalizing its own, more comprehensive rules governing data use and privacy standards for AI.

Our thought bubble: American companies say the EU approach will hobble EU-based AI development, and they're threatening to withhold more products from the region.

• If U.S. executives really want to see what a stifling AI regulatory regime looks like, all they have to do is look across the Pacific.

A group of the top tech companies unveiled a new coalition Thursday to create cybersecurity and safety standards for the artificial intelligence tools they're developing.

Why it matters: The promise of the group is that each company will follow the same rigorous security standards for their projects to keep malicious hackers at bay.

• Companies have historically approached cybersecurity differently in their product development — with some firms pursuing weaker security practices than others.

Driving the news: Google announced the new Coalition For Secure AI during the Aspen Security Forum in Colorado this week.

Zoom in: The new coalition will start its work by developing standards for software supply chain security for AI systems, compiling resources to measure the risk of these tools and pulling together a framework to help defenders determine the best use cases for AI in their work.

• Founding members include Amazon, Anthropic, Chainguard, Cisco, Cohere, GenLab, IBM, Intel, Microsoft, Nvidia, OpenAI, PayPal and Wiz.
• The coalition is operating under the OASIS Open, an international standards and open source consortium.

What they're saying: "This is the industry coming together and doing this as a coalition — it's not an executive order, it's not a regulation," Heather Adkins, vice president of security engineering at Google, said at an Aspen Security Forum event.

• "I've seen many of these coalitions in the field ... it's the reason why we are making such good progress on many other areas of cybersecurity," she added.

Reality check: Many of the participating companies had either already developed their own standards for securing AI or were working on them.

• Last summer, Google released its own framework for securing AI, which encouraged other companies to review the security measures in place for AI models and incorporate automation into their cyber defenses.

The intrigue: This is the first time the industry has come together to work on these AI security issues.

• The coalition plans to disseminate open-source methodologies, frameworks and tools to help companies that are looking for ways to securely implement AI into their workflows.

What's next: The coalition is actively accepting new members, Adkins said at the Aspen event.

• OpenAI released a cheaper, faster version of GPT-4o, and Box CEO Aaron Levie suggested it may be cheap enough that his company could offer paying customers unlimited access to Box's AI features.
• Netflix CEO Ted Sarandos told investors on yesterday's earnings call that AI will help filmmakers tell better stories. (The Hollywood Reporter)

Enjoy this video of actor Hank Azaria trying out several of his "Simpsons" characters' voices in an effort to get a British Royal Guard to break his poker face.