TechLook
A German court just told Google that AI Overviews can be sued for what they say. Estonia, on the same day, started handing out ID cards to the agents themselves. And one of America's largest banks told its engineers to budget their tokens like cash. The throughline today isn't model launches — it's the rules, permissions, and accounting that turn AI into infrastructure a company can actually defend.
A German Court Just Made AI Speech Actionable — Maybe for the First Time
A German court issued a preliminary ruling holding Google liable for false statements inside its AI Overviews after two publishers said the tool branded them scams and Google ignored their cease-and-desist letter. The court's reasoning: AI Overviews are not a list of links — they're "independent, new, substantive statements" — and that strips the immunity traditional search has enjoyed.
Here's everything you need to know:
- Two publishers were falsely labeled as scams in AI Overviews; Google ignored a cease-and-desist before the ruling.
- The court explicitly held that AI summaries are not a passive relay — they make new claims and can be sued for them.
- The court barred Google from repeating the specific false statements going forward.
- Google's own data, surfaced in coverage, shows Gemini Overviews carry bad source links roughly 56% of the time.
- This is plausibly the first ruling anywhere that holds an AI vendor liable for what its model said.
- The case is preliminary; appeals will follow, and the doctrine will be tested in higher courts.
The legal frame is more important than the verdict. Search earned Section 230-style immunity by being a dumb pipe — links out, no claims made. AI summaries forfeit that the moment they synthesize an answer that didn't exist in any source. For anyone shipping a product that paraphrases, summarizes, or generates content for users, the exposure profile just changed: a confident sentence in your UI is now a potential liability line, not a convenience feature. The defensive moves are unglamorous but real — citations on every generated claim, source ranking visible to the user, and a kill-switch for hallucinated content that didn't appear in any retrieved document. If you don't have those, today is the day to start.
Estonia Starts Issuing ID Cards to AI Agents
Estonia unveiled a plan, backed by the Eesti.ai advisory board, to issue digital ID codes to AI agents so they can act on behalf of users with verifiable, auditable permissions. The goal is to make responsibility clear as agents start doing real digital work — paying bills, filing forms, calling APIs.
Here's everything you need to know:
- The system assigns each agent its own identity and a structured permission set for the tasks it's authorized to do.
- The framing is accountability: when an agent acts, you can prove who authorized it, on what terms.
- Estonia is positioning itself as the first mover in agent identity governance — a category that barely existed a year ago.
- The move follows a Google DeepMind warning earlier this week that autonomous agents need controls like employees with office keys.
- It also lines up with SK Telecom's plan to give 80,000 employees across 25 companies digital "AI colleagues" with their own departments and lifecycles.
For builders, agent identity is going from a philosophical debate to a procurement requirement inside a year. If you're selling into regulated buyers — finance, healthcare, government, anything in the EU — the next RFP is going to ask which identity system your agents run under, how you audit their actions, and how you revoke them. Bake an identity layer into your product now and you skip the most expensive retrofit in your roadmap. The same logic applies to internal tools: if your agents touch customer data, your security review is going to want a verifiable answer to "who is this, who authorized it, and what did it do."
DeepMind's "AI Control Roadmap" — The First Real Agent Rulebook
Google DeepMind published a control framework that treats internal agents like employees who could misfire, with graduated permissions, monitored reasoning, and hard blocks on risky behavior. Across one million coding tasks, most flagged issues came from overzealous execution — not malice — and DeepMind is now pushing global agent-security standards.
Here's everything you need to know:
- The framework treats agents like staff with office access: graduated permission grants, action monitoring, and the ability to block risky moves.
- Tested across one million coding tasks inside DeepMind.
- The biggest failure mode was overzealous execution — agents acting too aggressively within the scope they were given.
- DeepMind is asking the industry to align on shared agent-security standards.
- It complements SK Telecom's "digital employee" rollout and Estonia's identity plan, forming the first coherent stack for agent governance.
The interesting bit isn't the framework — it's the failure data. "Overzealous, not malicious" means the safety problem is mostly a scope and context problem, not an alignment problem. For builders, that's a solvable engineering question: tighter defaults, scoped tools, and the ability for the agent to ask before doing something irreversible. If you ship agents, the cheapest insurance you can buy is a "permission to act" gate before any write operation, side effect, or external call.
Deutsche Bank Cut Project Timelines From 2 Years to 3 Months — With a Token Budget
Denis Roux, CIO of Deutsche Bank's investment bank, said AI has collapsed some project timelines from two years to three months. The bank manages the cost by allocating tokens to engineers like a budget line — teams request more after proving value. Current work: automating financial-data analysis and portfolio exposure checks.
Here's everything you need to know:
- Project timelines on AI-assisted work dropped from ~24 months to ~3 months on multiple internal builds.
- Engineers get a token allocation upfront; they have to earn additional tokens by demonstrating the work paid off.
- The bank is now automating financial-data analysis and portfolio exposure checks with agents.
- This is a working playbook for a regulated, cost-disciplined organization adopting AI.
- It also lines up with the Warner Bros. Discovery agentic ad-buying stack and State Farm mandating AI for all 19,000 agents by end-2026 — large operational rollouts are no longer pilots.
The token-budget pattern is the part founders should steal. Most companies give engineers a model and no sense of cost; the bill shows up on finance's desk three months later and the program gets cut. Deutsche Bank's approach — allocate, track, earn more by proving value — turns AI spend from a platform decision into a line item engineers can defend. If you're a B2B founder with a long sales cycle, this is also what your customer will look like in 2027: per-seat tokens, usage dashboards, and procurement asking which teams got the most for theirs.
A Brain-to-Text Decoder That Runs on Off-the-Shelf Hardware
UC Davis published a Nature Medicine paper on BRAND, a machine-learning platform that decodes neural signals into fluent language. The system runs on Blackrock's off-the-shelf electrode arrays and translates attempted-speech brain activity into phonemes, then words. Trial participant Casey Harrell, who has ALS, has used BRAND to work full-time as an environmental advocate for over a year.
Here's everything you need to know:
- The platform was published in Nature Medicine by the UC Davis team behind the BrainGate consortium.
- It runs on Blackrock's standard arrays — no custom hardware required.
- The pipeline goes signal → phoneme → word, which is the structure the field had been missing.
- An ALS patient has used it continuously for more than a year in a real job.
- The breakthrough is in the decoding layer, not the capture layer — anyone with electrodes can now plug in.
For most builders this won't change your roadmap this quarter, but it's a marker for where the next decade of human-computer interaction is going. The hard part was always decoding, not capturing signal — and the decoding layer is now hardware-agnostic. If you're building voice, accessibility, or any input modality that depends on imperfect human signal, watch the BRAND papers closely. The general lesson — that the right model turns commodity hardware into a category-defining product — applies everywhere.
The Grid Can't Keep Up: 5-Year Wait to Plug In a Data Center
PJM, the largest U.S. grid operator covering roughly 13 states, warned that new power-connection reviews could stretch to five years or more as AI and data-center load surges. The operator processed a record 51 GW of new data-center requests in 2025.
Here's everything you need to know:
- PJM processes power for about 13 states and is the largest U.S. grid operator by territory.
- New connection-queue reviews are projected to take 5+ years at current pace.
- 51 GW of new data-center requests came in during 2025 alone — a record.
- The bottleneck is no longer chips or buildings; it's interconnection and generation.
- DOJ recently told a court that AI compute is national security infrastructure — the political and grid stories are now the same story.
For founders, the implication is that "we'll just rent more compute" stops being true the closer you get to frontier-scale work. If your roadmap depends on training or serving at the scale where grid delay matters, you should be doing site selection and power procurement now, not when you need it. For everyone else, watch your cloud bill: hyperscaler pass-through of grid costs is already starting, and the pricing power for compute is migrating upstream from the labs to the utilities.
⚡ Quick Hits
- Midjourney: Plans the first "Midjourney Spa" in San Francisco's Union Square for 2027 — ten scanners paired with saunas, cold plunges, and hot tubs. Confirms the medical-hardware pivot is more than a one-off press release.
- Hassabis + Amodei: Doubled down publicly on an "all diseases curable within 10–20 years" timeline. Hassabis said the window "hasn't shifted, but has hardened, has tightened."
- Warner Bros. Discovery: Rolling out an agentic ad-buying stack to unify linear TV and digital; new planning tools due Q3.
- State Farm: Mandating AI tools for all ~19,000 agents by end of 2026.
- Amazon: Now selling AWS Trainium chips directly to customers — first time outside hyperscaler infra.
- Snap spinoff "Dotmo": Building an AI-first video creation and editing app.
- Kimi Work "Goal Mode": Moonshot's desktop agent runs autonomously 24/7 toward a user-set objective, with progress tracking and redirects.
- Sequoia / Sonya Huang: Argued agentic AI now lets founders command "a virtual team of interns" — coding first, general computer work next.
- Adobe Firefly AI Assistant: Now executes multi-step tasks across Premiere, Photoshop, InDesign; expansion to ChatGPT, Claude, Gemini, and Slack is pending.
- OpenAI ChatGPT scheduled tasks: Users can now schedule prompts to run autonomously at future times with proactive notifications. (Quick mention — same item appeared in yesterday's Quick Hits; today's confirmation is full GA.)
Techlook — AI & tech signal for founders and builders.