Enforcement and rollback define the weekend's AI agenda. EU transparency duties took effect, Minnesota's nudify ban survived an emergency challenge, and Google paused Earth image generation after one day. Reddit's referral exposure and a scam experiment reveal distribution and social costs, while Google's agent API makes retention an architectural choice.
1. EU begins enforcing AI Act transparency duties
From August 2, the European Commission's AI Office and national authorities begin enforcing the AI Act, and Article 50 transparency duties apply. Providers must make generated or manipulated audio, images, video, and text machine-detectable where technically feasible; interactive AI systems must disclose that users are dealing with a machine. Deployers must visibly identify deepfakes and AI-written public-interest text unless it received human review and has an accountable editor.
The scope is narrower than a universal "AI-made" badge: artistic, satirical, and fictional works receive tailored treatment, and edited public-interest text has an exception. The Commission's icons are optional, but the underlying duties are not. Compliance therefore requires both provenance metadata in generation pipelines and durable labels at distribution; a disclosure added only by the model provider will not necessarily satisfy the deployer that publishes the output.
Sources: European Commission enforcement notice · European Commission guidance on AI-generated-content labels
2. Minnesota's nudify ban takes effect after xAI loses emergency bid
US District Judge Donovan Frank denied xAI's request for a temporary restraining order against Minnesota's ban on apps that digitally remove clothing from images. The law took effect August 1. Frank emphasized that xAI sought emergency relief on July 29, nearly three months after the measure was signed and three days before its effective date.
The order leaves xAI's underlying lawsuit alive and does not resolve its claim that the first-of-its-kind state ban is overinclusive. It permits enforcement while that litigation proceeds. Timing defeated the immediate bid, shifting the next legal test toward the statute's scope and the alternatives xAI says could address nonconsensual sexual imagery with fewer speech restrictions.
Sources: TechCrunch on the denied injunction request · US District Court order
3. Google pauses Earth image generation one day after launch
Google paused an AI image-generation feature in Google Earth one day after introducing it. Powered by Nano Banana 2, the tool let users place photorealistic generated scenes over satellite, aerial, and 3D imagery for any location. Google said screenshots circulating online appeared to violate its policies and that stronger guardrails were under development.
The generated scenes were watermarked and did not enter the main Earth experience, but grounding them in recognizable real places increased their misinformation value outside the product. The one-day rollback shows a gap between embedded provenance and screenshot distribution: a label inside the tool cannot preserve context after the image moves elsewhere.
Sources: Reuters on Google's Earth image-generation rollback
4. Reddit's referral volatility exposes the bargain behind AI search
Reddit shares fell more than 20% after its quarterly report even though core financial results beat expectations, as investors focused on chief executive Steve Huffman's description of Google search referrals as "choppy." Huffman said AI Overviews had not produced the ecosystem-wide "win-win" created by conventional result links. The concern is measurable beyond Reddit: in a 2025 panel, Pew observed a traditional result click on 8% of Google visits with an AI summary versus 15% without one.
Pew's study covered 900 US adults and was observational, while Google says aggregate organic clicks remained relatively stable and the clicks it sent were higher quality; Google did not publish underlying traffic data in that response. The risk mechanism is substitution: a summary can satisfy demand inside Search while the source still pays to produce and host the material. Licensing revenue buys access to content; referral exposure determines whether users still reach its publisher.
Sources: Ars Technica on Reddit's search-referral risk · Pew Research Center click-through study · Google's response on Search traffic
5. AI outperforms a human in a small trust-building scam experiment
Researchers from four universities compared a Claude agent with a human romance-scam expert in weeklong conversations with 22 unwitting participants. At the end, 46% followed the bot's request to download an app, versus 18% who followed the human's request to download a game. Participants gave the bot a mean trust score of 3.78 against 3.31 for the human and sent it 80% of their messages.
The sample is small, the final requests differed, and the Claude version came from early 2025; the experiment measured exploitable trust, not a completed financial fraud. Its mechanism remains concerning because most relationship-building messages look benign. A human can take over only for the final solicitation, bypassing safeguards tuned to explicit scam language while automating the longest phase.
Sources: Ars Technica on the trust-building experiment · WIRED's original report
6. Google makes Interactions the default API path for Gemini agents
As inspected on August 2, Google's developer documentation presented the Interactions API as generally available and recommended for new Gemini projects. One interface handles models and specialized or managed agents, records execution steps, supports background work, and continues a task through previous_interaction_id. The separate Gemini Agents API defines reusable configurations, tools, source mounts, and network rules but remains beta under /v1beta/.
Stateful convenience changes the data and cost model. Interactions are stored by default for 55 days on paid projects and one day on free projects; store=false disables background execution and history chaining, while state reuse can improve context-cache hits. An explicit retention choice and pinned beta schemas keep those two release states from becoming an accidental data policy.
Sources: Google's Interactions API overview · Gemini Agents API reference