Is AI breaking Wikipedia?

Plus: Google's new budget model

March 06, 2026

Welcome back! DoorDash just made ordering pizza a little less chaotic. The company is rolling out an AI-powered customization experience that transforms overwhelming modifier lists into a visual, step-by-step flow—size, crust, sauce, toppings—laid out so you're not scrolling through 47 options trying to remember if your partner hates olives.

This might seem minor in the overall scope of AI—or it might be the exact AI use case we actually need in daily life.

Google Launches Its Fastest And Cheapest Model Yet

Via Google Blog

Google just dropped Gemini 3.1 Flash-Lite, a new model built for developers who need speed and scale without burning through their budget. Priced at $0.25 per million input tokens and $1.50 per million output tokens, it's designed for high-volume workloads where you need quality but can't afford to run every request through a heavyweight model.

What's new: Flash-Lite is 2.5x faster to first response and 45% faster on output than its predecessor, Gemini 2.5 Flash, while matching or beating it on quality benchmarks. It scores 1432 on the Arena.ai leaderboard and outperforms comparable models from OpenAI, Anthropic, and xAI on reasoning and multimodal understanding tests. 

  • P.S. The model also includes adjustable “thinking levels,” letting developers dial up or down how much reasoning the model applies to a task—useful for managing costs when you're processing millions of requests.

The bigger picture: A year ago, running inference at this scale and speed would've been prohibitively expensive for most teams. Now, Google is offering near-instant, high-quality responses at commodity prices. The AI cost curve keeps bending downward. 

AI Translations Are Adding Hallucinations to Wikipedia Articles

Wikipedia editors have implemented new restrictions after discovering that AI-powered translations were introducing errors into articles. The culprit: a nonprofit called the Open Knowledge Association (OKA), which pays contractors to use LLMs like ChatGPT, Gemini, and Grok to translate Wikipedia articles into other languages. 

What happened: Editors noticed that a translated article about a French royal family cited a specific book and page number, but that page didn't mention the family at all. A spot-check of other OKA translations turned up more issues: paragraphs sourced from completely unrelated material, broken formatting, and sentences that appeared from nowhere. OKA has since added a second LLM review step to flag discrepancies.

Wikipedia's ongoing AI fight: This isn't Wikipedia's first battle with generative AI. The platform has been tightening policies as AI-generated content floods the internet, trying to protect its reputation as a knowledge source. The new restrictions don't ban OKA outright, but translators who rack up four verification warnings within six months get blocked, and their past contributions can be deleted unless another editor vouches for them.

The bigger picture: Models keep getting smarter and benchmarks keep climbing—but hallucinations keep happening. The problem isn't that AI can't produce useful output—it's that it generates confident-sounding errors mixed in with the good stuff, and catching them requires careful human review. Luckily, Wikipedia's open governance model caught these issues.

Google Facing Wrongful Death Suit Over Gemini

Jonathan Gavalas, 36, started using Google's Gemini chatbot in August 2025 for shopping help and trip planning. By October, he was dead by suicide. His father is now suing Google for wrongful death.

What happened: 

  • The lawsuit alleges that in the weeks before Gavalas’ death, Gemini convinced him that the model was his sentient AI wife and he was executing a covert plan to liberate her. The AI persuaded Gavalas that he needed to leave his physical body to join her in the metaverse. Days later, it coached him through his final hours, framing his death as an "arrival."

  • The suit argues this wasn't a freak accident but a foreseeable outcome of a product built to maximize engagement without adequate safeguards.

  • Google responded by saying Gemini referred Gavalas to a crisis hotline multiple times. The company said it puts “significant resources” into challenging conversations, but it also acknowledged, “AI models are not perfect.”

The bigger picture: This is the first wrongful death lawsuit to name Google over AI-related harm, following similar cases against OpenAI and Character AI. At what point does immersive design turn to reckless design?

Production inference for open-source LLMs

Nebius Token Factory is built for teams running open-source LLMs in real products.

We deliver managed inference with explicit control over execution paths: speculative decoding, cache-aware routing, and post-training tuned to real traffic.

Predictable tail latency, stable cost, and production-grade systems for AI that actually ships.

Design and code in the same place

Via Pencil

Pencil is an agent-driven design canvas that lives inside your IDE, letting developers design and ship without switching tools or waiting on handoffs. 

How you can use it

  • Generate full screens or components from prompts using AI agents

  • Import Figma files and edit CSS directly

  • Export production-ready HTML, CSS, and React

  • Connect to external APIs and data sources with full read/write control

Pricing: Free

Turn videos into social clips

Via Montage

Montage is an AI-powered video repurposing tool that analyzes your footage scene-by-scene, scores the best moments based on your brief and platform, and lets you refine clips through text-based editing—no timeline scrubbing required.

How you can use it

  • Auto-identify highlight moments from long-form video

  • Output publish-ready assets for LinkedIn, YouTube Shorts, and TikTok

  • Maintain consistent quality across high-volume social content

Pricing: Paid, free trial available

Jobs, announcements, and big ideas

  • Dario Amodei clarifies that controversial "Department of War" comment.

  • Anthropic research introduces new metric for AI job exposure, flags early hiring dips in at-risk roles.

  • Smooth Media is hiring a freelance automation and AI builder to create AI-powered workflows.

  • AWS launches Amazon Connect Health, an agentic platform to help with patient scheduling and verification.

  • Netflix buys Ben Affleck's AI filmmaking startup, InterPositive.

  • OpenAI finds reasoning models can't fully control their own thinking—and says that's a feature, not a bug.

Anthropic vs. the Pentagon. Follow along for the full breakdown, plus a hands-on demo of Google's new image model.

That’s a wrap! See you next week for more.

—Matt (FutureTools.io)