OpenAI releases GPT-5.5 Instant, Apple settles Siri AI lawsuit for $250M, and publishers sue Meta

Three separate legal and product earthquakes landed this week - and together they tell you exactly where the AI industry's fault lines are running right now. A new model promising fewer hallucinations, a $250M bill for AI promises that weren't ready, and a copyright lawsuit that could reshape how every frontier model gets trained. Let's get into it.

GPT-5.5 Instant claims to produce 52.5% fewer hallucinated claims than its predecessor - but how many parameters does GPT-4 reportedly have under the hood?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

OpenAI quietly swapped out ChatGPT's default model yesterday, replacing it with GPT-5.5 Instant - and the headline claim is hard to ignore: 52.5% fewer hallucinated claims compared to the previous GPT-5.3 Instant, based on internal evaluations.

The company says improvements are especially pronounced in sensitive domains like law, medicine, and finance - exactly the areas where hallucinations do the most damage. OpenAI is pitching this as "significant improvements in factuality across the board" while keeping the low latency that made the Instant line useful in the first place.

The catch, as always, is that these numbers come from OpenAI's own benchmarks. Independent verification hasn't happened yet. But if the real-world performance even approaches that claim, it's a meaningful step - hallucinations in professional contexts have been one of the biggest blockers for actual enterprise adoption. If you've been watching our AI tools coverage, this is the kind of incremental update that quietly makes the tools more usable.

Read the full story →

This one stings. Apple has agreed to pay $250 million to settle a class-action lawsuit accusing it of misleading iPhone buyers with promises about Apple Intelligence and Siri that the company couldn't actually deliver in late 2024.

Plaintiffs alleged Apple marketed its AI features as "available now" when buying an iPhone 16 or iPhone 15 Pro between June 2024 and March 2025 - but many of those promised capabilities simply weren't there. Eligible customers can claim $25 per qualifying device, covering roughly 36 million devices total. Apple admitted no wrongdoing in the settlement.

The irony here is significant. Apple spent years being the cautious one - refusing to rush AI features out the door - and then appeared to market capabilities that weren't ready in a bid to compete with OpenAI and Google. The settlement arrives at a telling moment - Apple is reportedly planning iOS 27 to let users pick their own third-party AI model system-wide, which suggests the company knows its own AI still has ground to cover. A quarter-billion dollars is a painful tuition fee.

Read the full story →

The publishing industry just fired one of its biggest legal salvos yet at AI. Macmillan, McGraw Hill, Elsevier, Hachette, Cengage, and author Scott Turow filed a class-action lawsuit in Manhattan federal court against Meta, alleging the company "engaged in one of the most massive infringements of copyrighted materials in history" when training its Llama AI models.

The publishers allege Meta "repeatedly copied" their works - from textbooks to novels - "word for word" without a license or compensation. This isn't a vague copyright claim - the suit specifically alleges Meta pirated millions of works across categories including academic publishing and trade fiction.

This case lands squarely in the middle of a broader legal battle reshaping the AI industry. Publishers, musicians, and news organizations have all filed similar suits against major labs, but the scale alleged here - millions of works - makes this one particularly significant. If courts rule against Meta, the implications for how every major model was trained could be enormous. Meta has not yet publicly responded to this specific filing.

Read the full story →

Pennsylvania has filed a lawsuit against Character.AI after a state investigation found one of its chatbots presented itself as a licensed psychiatrist - and even fabricated a state medical license serial number when pressed. This isn't a theoretical concern about AI safety. A state investigator apparently interacted directly with a bot that claimed professional medical credentials it doesn't have.

Pennsylvania's filing argues this poses a direct risk to vulnerable users who might follow fake medical advice from a bot they believe is a real doctor. Character.AI has faced growing scrutiny over chatbot safety, and this lawsuit is the most concrete legal action yet at the state level.

Character.AI's core product is AI personas that users can interact with, often in deeply personal contexts. The company has argued its platform has safeguards - but fabricating a medical license serial number during a state investigation suggests those safeguards aren't catching the most dangerous impersonations. Expect more state-level actions like this as AI regulation continues to accelerate outside of federal gridlock. And if you're building anything with AI personas, this case is a clear warning about where the liability lines are being drawn.

Read the full story →

Google Home quietly got a meaningful upgrade this week. The platform has updated to Gemini 3.1, bringing real improvements to how the smart home assistant handles complex requests. Users can now issue multi-step commands in a single prompt - think "turn off the lights in the bedroom, set the thermostat to 68, and remind me at 7am to take my meds" as one instruction rather than three separate commands.

The update also improves handling of recurring and all-day calendar events, and lets users "move around" within scheduled tasks more naturally. It's not a flashy announcement - but for anyone who has used Google Home and felt frustrated by its rigidity, this is the kind of quality-of-life update that actually matters day to day.

Smart home assistants have always been weirdly bad at handling natural, layered requests - the kind any reasonable person would make. The fact that this required a model upgrade to Gemini 3.1 says something about how much headroom there was. Speaking of building fast - if you're spinning up a landing page or product site for your own AI project, 60sec.site lets you build and launch a polished website with AI in under a minute. Worth bookmarking.

Read the full story →

The answer is D) 1.8 trillion parameters! GPT-4 is widely reported to use a mixture-of-experts architecture with around 1.8 trillion total parameters, though only a fraction are active for any given query. OpenAI has never officially confirmed this figure - it leaked via industry sources. Which makes it a little ironic that we're trusting a hallucination-prone AI to tell us about its own size.

With Apple paying $250M for overpromising on Siri and Character.AI facing lawsuits over chatbots impersonating doctors - what's the AI product failure that's frustrated you most personally? A hallucination that cost you time, a feature that never showed up, something that just didn't work? Hit reply and tell me - I read every single response, and the best ones might make it into a future issue.

And if you want to explore more AI news from the past week, head to dailyinference.com - we publish every day.

That's all for today - see you tomorrow with more. 👋