Google API Changes, OpenAI Voice, & AI Legal Liability for Builders

This episode of No-BS AI Briefing tackles critical developments for AI builders. We break down Google's impending breaking changes to the Gemini Interactions API and OpenAI's powerful new real-time voice models, including GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper, which are already showing significant real-world impact for companies like Zillow. The deep dive focuses on a landmark US court ruling finding that government cannot hide behind AI decisions, specifically regarding DOGE’s use of ChatGPT to cancel NEH grants—a precedent with massive implications for product liability and compliance. We also touch on extended EU AI Act deadlines and CopilotKit's $27M Series A funding for UI-aware agents. Our practical takeaway helps you audit your high-stakes AI systems to navigate this evolving legal landscape. Follow No-BS AI Briefing for concise, opinionated briefings that keep you ahead.

Transcript

SPEAKER_00 0:00

A US court just dropped a bombshell ruling. You can't hide behind AI when it discriminates. This changes everything for how we build and deploy AI systems and products, especially in regulated industries. We're also talking about major API changes from Google and impressive new real-time voice models from OpenAI. No BS AI briefing brought to you by Proactive AI. Welcome back. I'm your host Vikas Sharma and this is where builders get straightforward AI news without the fluff. First up, Google's making some major changes to its Gemini Interactions API, and builders, you really need to pay attention here. They're replacing the outputs array with a brand new steps array. Now what does that actually mean in plain English? Well, it's all about unlocking more sophisticated agent flows. This change lets you do things like mid-flight steering, which means you can intervene or guide an AI agent while it's in the middle of a task rather than waiting for it to finish. You'll also get asynchronous tool calls, making your agents more responsive and efficient. For builders, this isn't just a nice to have, it's production critical. Google's giving us a tight timeline. You can opt in now, but it goes default on May 26th, and they're ripping out the legacy outputs array on June 8th, 2026. If you've got production systems relying on that old schema, you've got to update your parsers to the new steps array before June 8th or your integrations will break. This isn't something you can defer. The new response format also becomes polymorphic, so you'll need updated SDKs for Python and JavaScript versions 2.0 or greater. This really opens up possibilities for more complex dynamic agents, so it's worth reviewing the new schema to see what new experiences you can build. Next, OpenAI has just shipped three powerful new real-time voice models and they're showing some impressive real-world impact. We are talking about GPT Real-Time 2, which brings GPT-5 class reasoning with adjustable effort and a massive 128k context window. Alongside that, they've released GPT Real-Time Translate and GPT Real-Time Whisper. The pricing looks competitive for these capabilities too. Real-time 2 at $32 per million dollar input tokens and 64 per million dollar output tokens translate at just 0.034 per minute and whisper at 0.017 per minute. This isn't just for fun demos either. Zillow, for example, is already citing a significant 26-point lift in call success with GPT Real-Time 2. Now for builders, this is huge. Voice-to-action workflows become truly viable at scale. Imagine complex customer service, sales, or even internal operational workflows that can be driven purely by voice with real-time reasoning. That 128k context window coupled with adjustable reasoning means these models can handle really intricate multi-turned voice interactions without losing context or getting confused. It's a game changer for conversational AI and brings us much closer to truly intelligent voice assistants that can actually do things for people. Also, a US district court has delivered a really significant ruling. The court found that the Department of Government Ethics or Doge acted unconstitutionally by using ChatGPT to cancel over 1400 NEH grants. These grants were apparently flagged based on DEI-related keywords, and the court's ruling was crystal clear. It said there's no distinction between the government and ChatGPT when it comes to responsibility. What does this mean for builders like us? If you're developing or deploying AI systems that make regulated decisions, things like hiring, lending, approving grants, or anything with significant impact on individuals, you remain fully liable for the outputs and outcomes generated by that AI. You can't just shrug and say, the algorithm did it. Expect much stronger demands for human oversight, transparency into how your models are making decisions, and robust bias audits to ensure fairness. This isn't just for government either. It sets a precedent for any organization making impactful decisions with AI. Moving on to policy, the EU AI Act has seen an omnibus deal that extends some key deadlines, giving builders a bit more breathing room, but not on everything. The deadline for high-risk AI systems or HRAI has been pushed back to December 2nd, 2027. For product safety regulated systems, you've got even longer until August 2nd, 2028. Now that sounds great, right? More time for compliance. But hold on, the deal also introduces new prohibitions on non-consensual sexual content and CSAM and AI content marking requirements are still due much sooner by December 2nd, 2026. So while you might have an extra 18 months for some of the tougher compliance burdens, don't get complacent. Content safeguards and labeling requirements are still very much on the near-term agenda, and builders need to start incorporating these into their product roadmaps now. It's a mixed bag of extended timelines and immediate obligations. So make sure you understand which category your systems fall into. Finally, we've got some funding news that signals a clear trend. Co-PilotKit has successfully raised a $27 million Series A round. This funding was led by Glealot Capital, NFX, and Signalfire, and it's all aimed at expanding their enterprise AI co-pilot platform. What's particularly interesting here is co-pilot kits uh focus on UI aware agents. These aren't just chatbots, they're agents that can actually manipulate in-app elements. Think about that for a second. An AI that doesn't just tell you how to do something, but can actually do it for you directly within an application's user interface. For builders, this signals a growing and well-funded demand for enterprise copilots that go beyond traditional chat interfaces. It points towards a future where agents are deeply integrated into software using graphical user interfaces directly rather than relying solely on APIs. This AG UI or agent-to-user interface approach could be a significant leap forward in enterprise automation and productivity, opening up a whole new category of intelligent tools. Now, for our deep dive, we're coming back to that US court ruling, the one about doji and chat GPT. This isn't just another legal footnote, it's potentially a landmark decision. What happened? Well, the Department of Government Ethics or Doge decided to use Chat GPT to identify and then cancel over 1400 National Endowment for the Humanities or NEH grants. Their criteria were murky, tied to DEI-related keywords, and the result was a significant number of grant cancellations. The individuals affected sued, and the district court's finding was unequivocal. They said, and I'm quoting here, no distinction between the government and Chat GPT regarding responsibility for the outcomes. Essentially, the court rejected the age-old, the algorithm did it defense, stating that the government and by extension any entity remains fully accountable for decisions made using AI. Why does this matter right now? Because it's the first major US decision that very clearly spells out that using AI does not diminish your accountability, especially for discriminatory outcomes. For the market, this immediately shifts the conversation from can AI do it to can AI do it responsibly and accountably. It means that the legal system is catching up to the technology and the days of treating AI as an opaque black box that absolves responsibility are rapidly ending. This isn't just a policy paper, it's a legal precedent that will influence future cases and regulatory guidance. So who should care about this? Honestly, everyone building with AI, but especially founders and product managers. If you're a founder, your company's liability just got a lot clearer when you deploy AI. You need to factor this into your legal strategy and product design. Product managers, you'll need to ensure your AI-powered features aren't just efficient but also auditable, transparent, and fair. If you are an infrastructure engineer, this means a rising demand for governance tooling, think robust audit logs, systems for generating explanations for AI decisions and integrated bias detection frameworks. And for indie hackers, even if you're operating at a smaller scale, understanding this means building ethical AI from day one, you don't want your side project to become a legal headache. How I'd think about it as a builder, look, this isn't a ruling that says don't use AI for decisions. Not at all. What it does say is if you use AI for decisions, you better be able to explain, justify, and take responsibility for those decisions. My mental model here is like delegating a critical task to an employee. You as the manager or the company are still ultimately responsible for their work, especially if it leads to harmful or discriminatory outcomes. So you'd put in checks, training, and oversight. You need to apply that same rigor to your AI systems. This means baking in human-in-the-loop processes, developing explainability features so you can understand why your AI made a specific recommendation and performing regular, thorough bias audits. Don't wait for a lawsuit. Build this into your product DNA from the start. My no BS take on this. This is real, it's not hype. It's a foundational step towards mature AI governance in the legal system. AI is a tool, and the person or entity wielding that tool is still accountable for its impact. Period. This isn't about slowing innovation, it's about making innovation responsible and trustworthy. If you're finding this useful, hit follow in your podcast app right now. It takes two seconds and it's the best way to make sure you don't miss the next briefing. If you want one practical takeaway from today's episode, here it is. Audit any AI systems you have influencing high-stakes outcomes right now. You don't want to be caught off guard by a similar legal challenge. Here's how to try it in under 30 minutes. First, quickly list out any AI-powered systems or features in your product or internal operations that directly influence high-stakes decisions. Think about things like user eligibility, content moderation, hiring, loan applications, or even dynamic pricing models that could impact user access or fairness. Second, for each system you identified, document the decision criteria it uses. What data inputs are involved? What's the underlying model or logic? Can you describe its decision-making process in simple terms? You might even ask who would be accountable if this system made a bad or biased decision. Third, quickly sketch out your current oversight process. Who are the human owners reviewing the results? How often do they check? And importantly, how are you currently testing for potential bias or unfairness in the system's outputs? This isn't about solving all problems today, but about getting a clear picture of your exposure. This specific experiment is worth your time right now because the Doge ruling makes it clear. Accountability for AI isn't speculative, it's a present reality. Proactive auditing can identify blind spots, help you prioritize risks, and set you up for future compliance, protecting both your users and your business. That's it for today's NoBS AI briefing. If this helped, follow the show in your podcast app and share it with one builder you know. And if you've got questions or topics you want covered, connect with me on LinkedIn and send them over. See you in the next briefing.