Welcome to Unprompted: The AI Marketing Brief, where I cut through the noise in AI news and research to show marketers what’s happening — and why it matters for your work, your team and your career. 

The days of quietly passing off AI-generated content as human-made are over. Not "ending soon." Over. Recent AI-related announcements made that official on two fronts at once. 

Amazon now requires sellers to tag product photos that feature AI-generated people — not as a best practice, but because New York made it the law. Substack shipped a tool that scores your posts and comments on how likely they are to be AI-written. Different platforms, different mediums, same message: The internet just decided it wants receipts. 

And look, the systems enforcing this are rough. Amazon won't say what triggers its shopper-facing label. Substack's detector is explicitly probabilistic, which is a nice way of saying it guesses. Nobody's actually agreed on what "disclosure" should look like. But that's not going to slow this down. It's only going to get more common, more required and more your problem to manage — not the platform's. 

And here's the kicker: While everyone scrambles to label synthetic content, the tools making it just got a lot better at not looking synthetic. Black Forest Labs opened early access to FLUX 3, which generates video with native audio in a single pass. Google dropped a new tier of Gemini Flash models, faster and cheaper than the last round. The thing you'd need to disclose is evolving faster than the disclosure rules can keep up. 

So, pretending isn't an option anymore. Being ready to prove you weren't pretending — that's the job now. 

FLUX 3 - Real World Models: Towards Multimodal Flow Models as the Backbone of Visual Intelligence 

Website: Black Forest Labs 

Just the Facts: Black Forest Labs announced FLUX 3, a new multimodal foundation model that jointly learns from images, video and audio within a unified architecture rather than treating each modality separately, and the model is now available in Early Access. FLUX 3 can generate diverse videos with native audio up to 20 seconds in a single generation, supports capabilities including text-to-video, image-to-video, video-to-video, keyframe-to-video, multilingual dialogue and agentic chaining of clips into longer sequences, and in preliminary evaluations was preferred over competing video models including Grok Imagine Video, Kling v3 Pro, Runway Gen-4.5 and Luma Ray 3.2 at varying rates. The company also outlined a phased launch plan for video/audio generation, action prediction through partners like Mimic Robotics, image synthesis and an open-weight multimodal backbone, describing FLUX 3 as a step toward unified models for content creation and physical AI. 

Why It Matters to Marketers: 

• The move toward unified multimodal models jointly trained on image, video and audio reflects a broader industry shift from single-modality generative tools toward integrated "world models," a trend also visible in competitors' recent multimodal releases referenced in the article's comparison set.
• Marketers running video-heavy campaigns can request early access now to test typography rendering, multilingual dialogue and multi-shot sequence consistency firsthand — capabilities directly relevant to localized or template-driven ad production — before committing budget to a new tool. 

Amazon Cracks Down on Use of AI Images by Sellers After New York Law 

Author: Annie Palmer

Website: CNBC 

Just the Facts: Amazon is requiring third-party sellers to label any product images or videos containing AI-generated people, in response to a New York law that mandates disclosure when advertisements use a "synthetic performer" in place of a human actor. Amazon notified sellers of the policy Wednesday, directing them to tag qualifying images and "A+ content" — videos and other graphics on listing pages — with specific metadata keywords before uploading, while clarifying the rule does not apply to TV, video game or movie characters, or to real people who have been altered using AI. The New York law, which Governor Kathy Hochul's office called "first-in-the-nation," took effect last month, and Amazon said it will add a shopper-facing indicator to qualifying listings "where applicable," though the company has not disclosed the exact criteria it will use to decide when that label appears. 

Why It Matters to Marketers: 

• Sellers and marketers producing Amazon listing imagery with AI-generated people must now embed compliance metadata before upload — an added production step that takes effect now, since Amazon notified sellers this week per the article.
• This extends a broader platform-level AI-labeling trend from social media into e-commerce; Meta, TikTok, Pinterest and YouTube have each added AI-content labels, and California now requires large AI providers to embed watermarks in AI-generated content, reflecting a fragmented but accelerating state-by-state disclosure landscape in the absence of federal AI-disclosure law.
• Marketers running multi-state e-commerce or ad campaigns should prepare for inconsistent compliance requirements rather than one national standard, since New York's law is the first of its kind and other states (following California's watermarking rule) may adopt different thresholds or definitions. 

How Can I Detect AI on Substack?  

Website: Substack 

Just the Facts: Substack introduced a "Scan for AI text" tool that estimates whether a post, note, comment or reply was written by a human or is AI-assisted, using a third-party detection service called Pangram to assign a percentage score, and the tool works on content published on or after July 21, 2026. The feature is available in the Substack Reader on the web, the Substack iOS app and on notes and comments, but it does not work on video or audio posts, standalone Substack sites or custom domains or emails. Creators can add a "How I make this" transparency statement to their publication settings, disable AI detection on individual posts or notes and report detection errors, while Substack states that neither it nor Pangram uses publisher content to train generative AI models. 

Why It Matters to Marketers: 

• Content marketers publishing on Substack should audit whether AI-assisted drafts (even lightly edited ones) will now be flagged to readers by default, since the scan applies automatically to posts and notes published after July 21, 2026 unless detection is manually disabled.
• This adds Substack to a growing list of publishing and social platforms that are building AI detection or disclosure directly into the reader experience, consistent with the broader platform-level AI transparency trend seen recently at Amazon, YouTube, TikTok and Meta.
• AI detection tools, including Pangram, are probabilistic rather than definitive, so marketers relying on heavily AI-assisted drafting workflows should expect some false positives or reader skepticism even on substantially human-edited work, and should not treat a "human-written" score as a guarantee of accuracy. 

Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber  

Author: Tulsee Doshi 

Website: Google (The Keyword) 

Just the Facts: Google introduced three new Gemini models built on its Flash series: Gemini 3.6 Flash, a "workhorse" model priced at $1.50 per 1M input tokens and $7.50 per 1M output tokens that the company says reduces output token usage by 17% compared to 3.5 Flash while improving coding, knowledge work, and multimodal performance; Gemini 3.5 Flash-Lite, described as the fastest and most cost-effective model in the 3.5 class at $0.3 per 1M input tokens and $2.5 per 1M output tokens, running at 350 output tokens per second; and Gemini 3.5 Flash Cyber, a specialized model fine-tuned for cybersecurity vulnerability detection and patching that will be exclusively available to governments and trusted partners through Google's CodeMender security agent as part of a limited-access pilot program. Google states that 3.6 Flash ships with enhanced Frontier Safety safeguards against chemical, biological, radiological and nuclear (CBRN) and cyber-offense misuse, and reports performance gains across benchmarks including DeepSWE, MLE Bench, OSWorld-Verified and GDPval-AA v2 compared to prior Flash models. The company also disclosed that Gemini 3.5 Pro is currently testing with partners ahead of broader release, and that pretraining has begun on Gemini 4. 

Why It Matters to Marketers: 

• Marketing and content ops teams using Gemini-powered tools for document parsing, report drafting or agentic workflows should see faster, cheaper output as these models roll into the Gemini app, Google AI Studio and Gemini Enterprise starting now, per the article.
• The tiered release strategy — a faster/cheaper "Lite" model, a stronger "workhorse" model, and a narrowly gated specialized security model — reflects the broader industry shift toward differentiated model families optimized for cost-per-task rather than a single general-purpose model, a pattern also visible in Microsoft's MAI model-family announcement covered earlier in this roundup.
• Teams running high-volume, low-latency AI tasks — such as bulk content tagging, translation or data extraction — can test 3.5 Flash-Lite's throughput and pricing now given its stated price-to-performance advantages over prior Flash-Lite generations, before committing to a longer-term model migration. 

The State of Open Source AI — V1.0 

Website: Mozilla 

Just the Facts: Mozilla published "The State of Open Source AI — V1.0" (July 2026), a data-driven report finding that open-weight AI models have narrowed the capability gap with closed models from 8.04% to 3.3% on Chatbot Arena over 24 months, that GPT-4-class inference costs fell roughly 50-fold over 36 months, and that open-weight models now account for a majority of tokens routed through the OpenRouter platform. The report also finds that while 79% of developers adding AI functionality use open models compared to 71% using closed models, only 51% of open-model teams reach production versus 63% for closed-model teams, and it documents a multi-hundred-billion-dollar funded open-source AI business ecosystem including companies such as Databricks, Mistral and DeepSeek. The report frames open-weight AI increasingly as a "sovereignty" issue, citing the June 2026 incident in which a U.S. government export order forced a suspension of access to a closed AI model for foreign users worldwide as evidence of dependency risk, and identifies the "agentic harness" — the orchestration and tooling layer above the model — as the new competitive frontier between open and closed approaches. 

Why It Matters to Marketers: 

• Marketing ops teams evaluating AI tool costs for high-volume tasks should note the report's finding that open-weight models now carry a majority of production token volume at roughly one-sixth the per-call cost of comparable closed models, an immediate budget lever for content and data-processing workflows.
• The report's "sovereignty" framing echoes a broader enterprise shift toward vendor independence and exit rights in cloud and AI infrastructure, paralleling cloud-repatriation trends that IT leadership is increasingly weighing against convenience and lock-in risk (per the report's own citations to Linux Foundation and cloud-cost research).
• Marketers or IT-adjacent teams piloting open-model tools should expect the operational gap the report identifies — only about half of open-model projects reach production, versus nearly two-thirds for closed models — meaning lower cost doesn't guarantee an easier deployment path.
• Teams concerned about single-vendor dependency, especially after this summer's widely covered model-access suspension tied to export controls, can use the report's five-bet "opportunities" framework to evaluate where owning the harness or tooling layer — rather than the underlying model — offers more durable flexibility.