AI News: May 27, 2026

1. OpenRouter raises $113M Series B led by CapitalG at a $1.3B valuation as monthly tokens grow 20x year-over-year

OpenRouter. AI model gateway OpenRouter closed a $113M Series B led by Alphabet’s CapitalG, with NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, a16z, and Menlo participating. The round values OpenRouter at $1.3B, up from roughly $547M a year ago, and lands as the company reports weekly token volume grew 5x in the last six months, from 5T to 100T monthly tokens routed across more than 400 models from Anthropic, Google, OpenAI, xAI, and DeepSeek. The company says it now serves 8M global users, and the deep bench of strategic investors signals that hyperscalers and data-platform vendors are treating model-routing as a critical agent infrastructure layer rather than a commodity proxy. Source

2. Human Archive raises $8.2M to pay Indian gig workers $1/hour to wear sensor rigs and generate physical-AI training data

Human Archive. Human Archive, founded by three UC Berkeley researchers and a Stanford researcher, raised $8.2M from Wing VC, NVP Capital, Y Combinator, and angels from OpenAI, Nvidia, Google, and Meta. The startup pays gig workers in India $1/hour to wear camera caps, tactile gloves, and motion-capture suits while doing home services, hospitality, and restaurant tasks, generating multi-sensor egocentric data for robotics models. More than 1,000 headsets are already deployed, with early pilots also running in Southeast Asia and the United States. The pitch frames cheap labor in services-heavy economies as the new bottleneck-breaker for robot foundation models, where bespoke teleoperation data costs orders of magnitude more per hour. Source

3. Spectral AI wins FDA De Novo clearance for DeepView, a multispectral burn-assessment system trained on 340 billion pixels

Spectral AI. Dallas-based Spectral AI received FDA De Novo classification for DeepView, a multispectral imaging device that uses proprietary AI to predict on day one whether burn areas are unlikely to heal within 21 days. Image capture takes 0.2 seconds with results returned in 20-25 seconds, with the model trained on a dataset of more than 340 billion pixels of burn imagery. Development was backed by $31.7M in BARDA funding. The company (Nasdaq: MDAI) is now authorized to begin U.S. commercial distribution to burn centers, trauma centers, and emergency departments, marking one of the first burn-specific AI imaging clearances. Source

4. China now requires top AI researchers at Alibaba, DeepSeek, and other private labs to get government clearance before leaving the country

China. Per Bloomberg, Beijing now mandates that senior AI researchers at private firms including Alibaba and DeepSeek obtain government authorization before any overseas travel, with the rule applied to staff working on strategically significant projects. The policy escalates informal 2025 advisories that had merely discouraged AI executives from visiting the United States, turning soft guidance into a formal export-control regime over human capital. Beijing frames the move as a response to IP-theft and talent-poaching concerns, and it tightens China’s grip on its domestic AI talent pool as competition with U.S. labs intensifies. The policy creates new friction for international conferences, collaborations, and dual-affiliation arrangements that until now had operated freely. Source

5. Lancet audit finds AI-fabricated citations in biomedical papers have grown 12-fold since 2023, with review articles disproportionately affected

Lancet. A Lancet study led by Columbia’s Maxim Topaz audited 2.47M biomedical papers indexed in PubMed Central from January 2023 through February 2026 and identified 4,046 fabricated references across 2,810 papers. The fabrication rate rose from about 4 per 10,000 papers in 2023 to 56.9 per 10,000 in early 2026, with review articles, which often anchor clinical guidelines, showing a 57% higher rate than primary research. As of the audit, 98.4% of affected papers had received no publisher response. The authors call for automated pre-publication reference checks and retroactive screening, and warn that hallucinated citations are now actively contaminating the evidence base on which physician-facing guidelines are built. Source

6. DuckDuckGo installs jump up to 30% in one week as users react against Google’s AI-only Search

DuckDuckGo. DuckDuckGo reported six consecutive days, May 20 through 25, of sharp install growth in the United States after Google announced at I/O 2026 that AI agents would replace traditional search results. U.S. installs averaged 18.1% week-over-week growth, peaking at 30.5% on May 25, with iOS averaging 33% and peaking at 69.9%. Traffic to the company’s noai.duckduckgo.com endpoint averaged 22.7% growth over the same period. CEO Gabriel Weinberg framed the surge as a backlash, saying Google is “force-feeding AI with no way to opt out;” DuckDuckGo still holds only about 2% of U.S. search share, so the absolute volumes remain small even at peak growth. Source

7. Universal Music Group and TikTok renew their licensing pact with new commitments to remove unauthorized AI music

UMG. Universal Music Group and TikTok renewed their multi-year licensing agreement, with new commitments to remove unauthorized AI-generated tracks from the platform, improve artist and songwriter attribution, and ensure earnings flow through to rights-holders. The deal follows UMG’s 2024 catalog pull from TikTok over AI and royalty disputes, which had stripped TikTok of access to artists including Taylor Swift, Drake, and Olivia Rodrigo for several months. TikTok also expanded its “TikTok for Artists” hub, giving labels deeper performance analytics on a per-song basis. Financial terms were not disclosed. Source

8. CiteVQA benchmark exposes “attribution hallucination”: GPT-5.4 hits 87% answer quality but drops to 59% under strict citation checks

Research. Researchers from Peking University and Shanghai AI Laboratory introduced CiteVQA, a 1,897-question benchmark across 711 PDFs that forces models to cite the exact paragraph, table, or figure backing each answer. GPT-5.4 scored 87.1% on answer quality but fell to 59% once strict citation requirements were applied, Gemini-3.1-Pro-Preview reached 76/100, and Qwen3-VL-235B-A22B managed only 22.5/100. Page-localization accuracy proved to be a bottleneck, with narrowing the search space lifting scores by more than 13 points for some models. The benchmark formalizes a failure mode that document QA pipelines have long observed informally: models get answers right while pointing at the wrong place in the source. Source

9. White House moves toward pre-release AI model vetting as Google DeepMind, Microsoft, and xAI agree to NIST safety testing

Policy. NIST announced that Google DeepMind, Microsoft, and xAI have agreed to share frontier models with the U.S. government for pre-release safety testing, as the White House weighs broader AI-vetting requirements that could become mandatory. The shift follows what Treasury Secretary Scott Bessent called a “step change in the power of one large language model,” widely read as a reference to Anthropic’s Mythos. Industry pushback has been sharp over the scope of compute disclosures and the projected review timelines, which several labs argue would slow deployment without commensurate safety gains. Notably absent from the announcement: OpenAI and Anthropic, whose participation remains under negotiation according to officials briefed on the talks. Source

10. Microsoft SkillOpt frames Markdown skill files as trainable artifacts and wins 52-of-52 head-to-head benchmark runs

Research. Microsoft researchers introduced SkillOpt, a method that treats Markdown procedural skill files for frontier agents as optimizable artifacts, applying bounded text edits with validation gates rather than touching model weights or static system prompts. In reported runs, SkillOpt won 52 of 52 head-to-head comparisons across six benchmarks. The technique offers a “third way” between fine-tuning and prompt engineering for adapting closed-weight frontier agents to specific domains, treating the agent’s instruction surface as the optimization target instead of the model. The result lands as more closed-weight agent platforms standardize on Markdown skill systems, making the optimization target widely portable. Source