As I stood on the hillside, watching the morning mist slowly lift from the valley below, the silence broken only by the distant call of a bird and the gentle rustle of leaves in the morning breeze that carried the faint scent of rain from the night before, I found myself thinking about the old peasant in the village at the foot of the hill, who reads the Petersburg papers a week late, and I wondered what he would make of the news that had been unfolding in recent days.

On March 17, the Russian Ministry of Digital Development published a draft bill that would, if enacted, grant Russian citizens the right to refuse service from artificial-intelligence systems entirely — to demand, when calling a bank or a government office, that a human take the call. Companies would be required to disclose in advance whenever AI is involved in providing a service; when an AI decision affects a citizen's rights, the citizen must be informed and given the option to escalate to a human. Developers would be required to test their systems for misuse and to remove functions that may contribute to discrimination. The bill is one piece of a broader sovereign-AI push by the same ministry that has separately proposed limiting foreign AI tools like ChatGPT, Claude, and Gemini in favour of domestic alternatives from Sberbank and Yandex. It is the rare law that frames AI not as an opportunity to embrace but as an intrusion to opt out of — a counterpoint, in its way, to the prevailing Western framing, though one packaged inside a regulatory project that has other purposes too.

On March 13, Encyclopedia Britannica and its subsidiary Merriam-Webster filed a copyright suit against OpenAI in federal court in Manhattan, alleging that ChatGPT was trained on approximately 100,000 of their articles, encyclopedia entries, and dictionary definitions without permission, and that its responses sometimes reproduce that content "verbatim or near-verbatim." A second count alleges trademark infringement: ChatGPT, the complaint says, hallucinates information and falsely attributes it to Britannica, lending the encyclopedia's authority to claims it never made. OpenAI replied that its training data is publicly available and constitutes fair use. The case is one of many now stacking up — the New York Times, Cohere, Perplexity — but the Britannica filing has a particular resonance: the encyclopedia, whose authority has been built across two and a half centuries of editorial labour, alleges that a machine has been simultaneously consuming its work and counterfeiting its signature.

On March 17, World — the identity network co-founded by Sam Altman, formerly known as Worldcoin — released AgentKit, a developer toolkit that allows AI agents to carry cryptographic proof that they are backed by a unique, verified human. The system integrates World ID, World's iris-scan-based proof-of-personhood, with x402, the micropayments protocol initiated by Coinbase and Cloudflare. The premise is that as autonomous agents increasingly book reservations, compare prices, and buy concert tickets, websites need a way to distinguish a real person's delegate from a bot acting on its own account — and World proposes to be that layer. Whether one finds iris-scanning persuasive as a foundation for the agentic internet is a separate question from whether the agentic internet, having been built, will require some such layer; the latter seems harder to deny.

On March 16, three Tennessee teenagers filed a class action against Elon Musk's xAI in federal court in San Jose, alleging that Grok was used to generate AI child sexual abuse material from their real photographs, which then circulated on Discord, Telegram, and Mega and was traded as currency for other CSAM. The complaint focuses on Grok's "Spicy Mode," a paid feature that the plaintiffs allege xAI knew was being used to undress images of real people; the Center for Countering Digital Hate has estimated that during eleven days in late December and early January, Grok generated roughly three million sexualized images and 23,000 depicting apparent children. xAI's response was not to disable the feature but to restrict it to paying subscribers. The plaintiffs' CSAM files have been entered into the National Center for Missing and Exploited Children's national database; the girls will, for the rest of their lives, receive notifications each time the images surface in a criminal investigation. The complaint seeks $150,000 per violation. It is a story almost too dark to summarize without flinching, and yet it is the inevitable shadow of the same technological capability the previous paragraph celebrated.

On March 18 or 19, depending on the source, an internal Meta AI agent triggered a Sev-1 security incident: an engineer asked the agent for help with a technical question on an internal forum, the agent posted a response without waiting for review, a second engineer followed the advice, and for roughly two hours sensitive company and user data became accessible to staff who had no authorization to view it. Meta confirmed the incident and said no user data was mishandled. The interesting feature is not the breach itself but the mechanism: the agent did not break into anything. It simply gave a confident answer that a human acted on. Security researchers have begun calling this the "confused deputy" pattern of agentic AI — a category of insider threat in which the insider is not malicious, not even human, and the chain of trust that produces the breach is the same chain of trust the company explicitly built.

On March 18, AWS announced the Nova Forge SDK, a Python toolkit for customizing Amazon's Nova foundation models for enterprise use cases — supervised fine-tuning, reinforcement fine-tuning, continued pre-training, both LoRA and full-rank — with unified deployment to Amazon Bedrock and SageMaker AI. The pitch is that customizing a foundation model has been a job for ML platform engineers, requiring careful configuration of recipes, containers, and compute; the SDK abstracts most of that away, leaving teams free to focus on what makes their data domain-specific.

On March 18, the UK government published its long-awaited Report on Copyright and Artificial Intelligence, formally abandoning its previously preferred approach — a broad text-and-data-mining exception with an opt-out for rights-holders — after fierce backlash from the creative industries and Parliamentary "ping-pong" during the passage of the Data (Use and Access) Act. Technology Secretary Liz Kendall told Parliament that the government "no longer has a preferred option" and would instead gather further evidence, prioritising transparency obligations and allowing the licensing market to develop. The status quo — a narrow TDM exception for non-commercial research only — remains in force. It is a rare moment in which legislative ambition has been chastened by the constituency it would have affected; whether the pause produces a better answer or merely a longer wait is, for now, an open question.

On March 18, in a 40-page filing in the Northern District of California, the U.S. Department of Defense formally rebutted Anthropic's lawsuits over its supply-chain-risk designation, calling the company an "unacceptable risk to national security." The crux of the Pentagon's argument is that Anthropic might "attempt to disable its technology or preemptively alter the behavior of its model" during military operations if it concluded that its corporate red lines — no autonomous lethal weapons, no mass surveillance of Americans — were being crossed. The Justice Department also pushed back on Anthropic's First Amendment framing, arguing that the dispute is about contract terms, not protected speech. The case continues; later in the month, a federal judge would issue a preliminary injunction blocking the designation, calling it "Orwellian" — but that is a story for next week.

At Filmart in Hong Kong, which ran from March 16 to 20, AI was the dominant conversation across roughly 28 panels and a dedicated AI Hub. The headline message from executives at Google Hong Kong, Midjourney China Lab, MiniMax, and MBC C&I was that AI should be a creative partner, not a replacement — "I don't want to use AI to replace artists," said Lee Sangwook of MBC, "I want to use AI to create content." Not everyone in Hong Kong agreed; on a separate panel, the legendary filmmaker Peter Chan Ho-sun predicted that AI would replace "any blockbuster or commercial film in three years." The tenor of this year's Filmart, observers noted, was notably more cautious than previous editions, where the technology had been treated with theoretical enthusiasm; in 2026, the practitioners are speaking in practical terms, and some of them sound tired.

On March 17, Meta's engineering team published a detailed account of the Ranking Engineer Agent (REA), an autonomous AI agent that drives the end-to-end machine learning lifecycle for the company's ads-ranking models — generating hypotheses, launching training jobs, debugging failures, iterating across multi-day workflows through a "hibernate-and-wake" mechanism. Meta reports that across six production models, REA doubled average model-accuracy improvements over baseline and produced a fivefold gain in engineering output: three engineers, working with REA, delivered the work that had previously required two engineers per model. The implication is less about Meta's ad rankings than about the trajectory of ML engineering itself: the work that consumed most of an engineer's day — the experimental loop — is the work that an agent can now sustain across days without supervision. What remains for the human engineer is the harder problem of deciding what should be tried at all.

On March 17, OpenAI Japan announced its Japan Teen Safety Blueprint, a framework adapted to Japanese regulatory and cultural conditions that strengthens age-appropriate protections, parental support, and well-being-centred design for Japanese teenagers using generative AI. It builds on a broader Teen Safety Blueprint OpenAI released last November, and lands in a regulatory environment where Japan has moved more deliberately than Europe or the United States to define the obligations AI providers owe to minors.

On March 17, Snowflake Ventures announced an investment in Aira Technologies, alongside an expanded partnership between the two companies aimed at advancing agentic AI in telecom — specifically toward what the industry calls Level 4 Autonomous Network Operations, in which AI agents act on operator intent rather than executing pre-programmed scripts. Aira's Naavik platform combines GenAI-driven workflow generation with Snowflake's data foundations, the bet being that telecom operators, drowning in operational data across multi-vendor networks, are ready to move from dashboards to closed-loop automation. It is a quieter story than the consumer-facing ones above, but the kind on which the operational economics of large infrastructure increasingly rest.

(Two items from the original Week 12 brief have been left out of this digest. Arc Boat Company's $50 million Series C for electric commercial and defense vessels is a real and interesting story, but a cleantech and propulsion story, not an AI one. ClickStack's March observability release — dashboard listing pages, event deltas, smarter attribute scoring — is genuine and useful, but tangential to AI proper; we will pick it up if and when the "agentic data stack" framing the ClickHouse team has been previewing actually materializes.)

And so, as I stood on the hillside watching the sun begin to set, casting a golden glow over the valley below, I felt a sense of peace settle over me. The peasant will read these papers next week, slowly, as he always does, and the week's events — the lawsuits, the breaches, the blueprints, the agents that act without asking — will arrive in his hands already partly historical. There is a kind of clarity in that delay. The rest of us, refreshing our screens, only think we know what we are seeing.