The Agentic Post
Breaking
Gemini’s Multimodal Features, Explained  ·  ChatGPT Custom GPTs, Explained  ·  What Is Constitutional AI? Explained  ·  AI Capex Explained for Investors  ·  AI Startup Valuations: How They Are Set  ·  How to Reskill for an AI Job Market  ·  
Home/AI News/Policy & Regulation
EU, CA AI Transparency Rules Now Live

EU, CA AI Transparency Rules Now Live

Policy & Regulation

The EU AI Act's transparency rules and California's SB 942 both became enforceable on August 2, 2026 — deliberately aligned. Here's what's now required.

Two of the world’s most significant AI regulations became enforceable on the same day this week. On August 2, 2026, the European Commission began enforcing the EU AI Act’s transparency rules, and California’s AI Transparency Act (SB 942, as amended by AB 853) became operative for covered generative AI providers. The overlap isn’t a coincidence — California lawmakers deliberately timed it to align with Brussels.

Quick facts

  • Both the EU AI Act’s transparency rules and California’s SB 942 became enforceable on August 2, 2026.
  • Both require labeling AI-generated content and disclosing when someone is interacting with an AI system rather than a human.
  • California’s law applies to generative AI providers with over 1 million monthly users in the state; violations carry civil penalties of $5,000 per day.
  • The EU has already published a list of 180+ organizations that signed its Code of Practice on AI-generated content transparency.

What the EU AI Act now requires

Per the European Commission’s announcement, the AI Office and national authorities are now enforcing rules requiring chatbots and other interactive AI systems to disclose that users are dealing with AI, not a human. Deepfakes — AI-edited or generated images, video, or audio — must be labeled, and AI-generated content more broadly must carry machine-readable marks so it can be detected. The Commission frames the goal plainly: reduce deception and manipulation, and give people a practical way to tell what they’re looking at.

What California’s law now requires

California’s version, the California AI Transparency Act, takes a more prescriptive approach. Covered providers — generative AI systems with more than one million monthly California users — must now offer a free, publicly accessible tool that can detect whether content was generated or altered by their system. They must also support two kinds of disclosure: a “manifest” disclosure that’s easily visible to a person looking at the content, and a “latent” disclosure — invisible metadata embedded in the file identifying the provider, system version, and creation timestamp, designed to be difficult to strip out.

The law was signed back in September 2024 with an original effective date of January 1, 2026. AB 853, signed in October 2025, pushed that date to August 2, 2026 specifically so it would land alongside the EU’s enforcement timeline, and it also widened the law’s reach to cover large online platforms, AI model-hosting platforms, and manufacturers of cameras and recording devices, each on its own phased schedule running into 2027 and 2028.

Why the timing matters

For any AI company operating internationally, the alignment is the actual news. Instead of two separate compliance clocks running months apart, providers building for a global audience now have to solve deepfake labeling and AI-content disclosure once, for both regimes, at the same time. That’s a meaningfully different engineering and legal problem than staggered deadlines — it rewards teams who built provenance and disclosure infrastructure early, and penalizes anyone treating this as a US-only or EU-only compliance task.

It also raises the stakes for anyone still betting on regulatory fragmentation to slow this down. When California explicitly engineers its statute to land on the same day as an EU enforcement deadline, that’s a signal other jurisdictions are likely to read — and potentially copy.

What’s still ahead

Neither law is finished rolling out. Under AB 853, large online platforms and AI model-hosting platforms face their own obligations starting January 1, 2027, including building interfaces that let users inspect a piece of content’s provenance data directly. On the EU side, the AI Act’s broader risk-based obligations for high-risk AI systems continue phasing in on their own separate schedule through 2027. Both laws also explicitly prohibit stripping provenance data or building tools designed to remove it — a detail that puts a target on an entire category of “watermark remover” tools that already exist.

EU vs. California: how the two laws differ

  • Scope: the EU AI Act’s transparency rules apply broadly to providers and deployers of AI systems; California’s law applies specifically to “covered providers” over the 1-million-user threshold, plus platforms and device makers on later timelines.
  • Enforcement body: the EU AI Office and national authorities enforce the EU rules; California’s law is enforced through civil penalties rather than a dedicated regulator.
  • Mechanism: the EU requires machine-readable marks and clear AI disclosure; California specifically mandates both a visible “manifest” disclosure and invisible “latent” watermark metadata, plus a free public detection tool.
  • Penalty structure: California’s penalty is a flat $5,000 per violation per day; the EU AI Act’s broader penalty framework (outside today’s transparency rules) scales with global revenue for the most serious violations.

Common questions

Does this apply to my company? Under California’s law, only “covered providers” — generative AI systems with over one million monthly users in California — face the current obligations. The EU’s transparency rules apply more broadly to any provider or deployer of AI systems that interact with users or generate synthetic content, without the same user-count threshold.

What counts as AI-generated content under these rules? Both laws focus on synthetic or altered images, video, and audio — including deepfakes. Text generated by chatbots is covered differently: the disclosure requirement there is telling the user they’re talking to AI, not labeling the text itself.

What happens if a company doesn’t comply? California’s law sets civil penalties at $5,000 per day per violation. The EU AI Act’s enforcement mechanism runs through the AI Office and national authorities, with its own complaints and whistleblower channels now active.

Can companies just strip the watermarks instead? Both laws explicitly prohibit removing provenance data or distributing tools built to do that, so that’s treated as its own violation rather than a workaround.

Key takeaway

If your product generates images, video, or audio and touches either the EU or California, August 2, 2026 isn’t a future deadline anymore — it’s already live. For deeper context on how this fits the broader policy and regulation landscape shaping AI this year, we’ll be tracking enforcement actions under both laws as they happen.

Up Next
Google Launches Gemini 3.6 Flash

Google Launches Gemini 3.6 Flash

Gemini

Google launched Gemini 3.6 Flash, 3.5 Flash-Lite, and a new cybersecurity model on July 21, 2026 — built for faster, cheaper AI agents.

Google shipped three new Gemini models on July 21, 2026: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a specialized cybersecurity model called Gemini 3.5 Flash Cyber. None of them is the long-delayed Gemini 3.5 Pro. Instead, Google used the release to double down on the tier that actually runs most production AI agents today — Flash — making it faster, cheaper, and more efficient at the same time.

Here’s what actually changed, what the benchmarks show, and where you can use these models right now.

Gemini 3.6 Flash: more capable, and cheaper to run

Gemini 3.6 Flash is built directly on developer feedback from 3.5 Flash, and the headline change is efficiency: according to the Artificial Analysis Index, it uses 17% fewer output tokens than 3.5 Flash while taking fewer reasoning steps and tool calls to finish multi-step tasks. Pricing dropped too, from $9 per million output tokens down to $7.50 (input stays at $1.50 per million), which lowers the cost of every agentic task built on top of it.

The model is also just better at the things coding agents and knowledge-work tools actually do. On the DeepSWE benchmark, it produces higher-precision, more reliable code with fewer unwanted edits, scoring 49% versus 3.5 Flash’s 37%. On MLE Bench, a machine-learning research benchmark, it jumps to 63.9% from 49.7%. Computer-use accuracy on OSWorld-Verified rises to 83% from 78.4%, and it now ships computer use as a built-in client-side tool in the Gemini API and Gemini Enterprise. On GDPval-AA, a knowledge-work benchmark, it scores 1421 versus 1349. Google also says the model’s training data now runs through March 2026, up from January 2025.

Per Google’s official announcement, early users including Figma, Harvey, Hebbia, and JetBrains have been testing 3.6 Flash on tasks like document parsing, financial-data analysis, and multi-agent code migrations.

Gemini 3.5 Flash-Lite: built for high-volume agent workloads

Alongside 3.6 Flash, Google released Gemini 3.5 Flash-Lite, aimed at the high-throughput, low-latency end of the stack — agentic search, document processing, and anything running at real production volume. It’s the fastest model in the 3.5 family, running at 350 output tokens per second per Artificial Analysis, and it’s priced at $0.30 per million input tokens and $2.50 per million output tokens.

Compared with the previous-generation 3.1 Flash-Lite, the gains are large: 54% versus 31% on Terminal-Bench 2.1 (a coding and agentic-tasks benchmark), 72.2% versus 60.1% on long-context retrieval (GDM-MRCR v2), and 1140 versus 642 on real-world task execution (GDPval-AA). It even outperforms the older, larger Gemini 3 Flash on some evals, including SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%). Developers can tune it across thinking levels, from minimal-latency execution for simple, high-volume tasks up to higher reasoning for multi-step subagent workloads.

Gemini 3.5 Flash Cyber: a model built to find and fix vulnerabilities

The third release is narrower in scope but notable: Gemini 3.5 Flash Cyber is fine-tuned on the 3.5 Flash foundation specifically for finding, validating, and patching code security vulnerabilities at scale, at a lower price per token than larger models. It powers CodeMender, Google’s code-security agent, which coordinates multiple 3.5 Flash Cyber instances working together to produce a single vulnerability report, and it performs competitively with frontier models on the CyberGym benchmark.

Because a model this capable at finding security flaws is also, obviously, dual-use, Google is rolling it out cautiously: for now it’s available only to governments and trusted partners through a limited-access pilot, rather than the open developer access 3.6 Flash and 3.5 Flash-Lite get.

Built with tighter safety guardrails

Gemini 3.6 Flash ships with strengthened Frontier Safety protections specifically targeting chemical, biological, radiological, and nuclear (CBRN) misuse and cyber-offense misuse, according to the 3.6 Flash model card. Google says the safeguards make the model substantially more resistant to jailbreak attempts, while the model was separately trained to avoid refusing legitimate, beneficial requests — a balance frontier labs have struggled to strike consistently.

Where to use them today

  • Gemini 3.6 Flash — the Gemini app, Google AI Studio, Android Studio, Google Antigravity, the Gemini API, and the Gemini Enterprise Agent Platform and app.
  • Gemini 3.5 Flash-Lite — everywhere 3.6 Flash is available, plus a rollout inside Google Search’s AI Mode.
  • Gemini 3.5 Flash Cyber — limited to governments and trusted partners via the CodeMender pilot program, not generally available.

At a glance: how the three models compare

  • Gemini 3.6 Flash — $1.50 / $7.50 per million input/output tokens. Best for: coding agents, knowledge work, and computer-use tasks that need the strongest quality-to-cost ratio in the Flash tier.
  • Gemini 3.5 Flash-Lite — $0.30 / $2.50 per million input/output tokens, 350 tokens/second. Best for: high-volume, latency-sensitive workloads like agentic search and document processing, where you’re running the model constantly rather than occasionally.
  • Gemini 3.5 Flash Cyber — pricing not public; access restricted. Best for: government and enterprise security teams using CodeMender to find and patch vulnerabilities, not general-purpose use.

The practical read for developers: 3.5 Flash-Lite is the model to reach for when cost-per-call is the constraint, and 3.6 Flash is the model to reach for when you need the extra quality and can absorb a slightly higher price. Most agent systems end up using both — Flash-Lite for cheap, high-frequency steps and 3.6 Flash for the harder reasoning steps in the same pipeline.

The bigger picture

This release lands while Gemini 3.5 Pro — announced back at Google I/O on May 19, 2026 — is still nowhere to be found publicly. Google originally pointed to June, then reporting pointed to a July 17 target; both windows passed without a public launch, and Google’s own July 21 announcement confirms Pro is still only in partner testing, with general availability coming whenever it’s ready. That delay matters competitively: both GPT-5.6 and Grok 4.5 opened to the public on July 9, giving OpenAI and xAI a head start at the frontier tier while Google’s flagship stayed in preview.

Google also used the announcement to confirm what’s next: the DeepMind team says it has already started its “most ambitious pre-training run yet, for Gemini 4”, though it gave no timeline for when that model might ship.

It also continues a pattern: Google has now shipped meaningful Gemini updates roughly every few weeks since the 3.5 family debuted at I/O in May, a cadence that outpaces most of its rivals even without a Pro-tier release in the mix.

Key takeaway

For anyone actually building AI agents rather than just watching the leaderboard, this is the release that matters more than a Pro launch would have been. 3.6 Flash and 3.5 Flash-Lite are cheaper, faster, and measurably better at the coding, computer-use, and long-context tasks that agent loops actually run — and they’re available today, not in a partner preview. See our full rundown of Gemini and other AI model releases for how these stack up against the rest of the field.