AI news5
Midjourney invests $74M to launch full-body ultrasound scanner, targeting spa-based preventive health screening
Midjourney, the AI image generation company, announced its first hardware product this week: the Midjourney Scanner, a whole-body ultrasonic CT device that scans the body in under 60 seconds and produces 3D maps down to a fraction of a millimeter. The company has invested more than $74M, partnered with Butterfly Network's semiconductor ultrasound chip technology, and is building purpose-built spa facilities to house the machines rather than targeting hospital settings. Radiologists and medical professionals responded critically, noting an absence of clinical trial data, regulatory approval, or evidence base to support the claims.
engadget.com ↗Why it matters
one of the most recognisable AI-generation companies pivoting into consumer health hardware while positioning preventive scanning as a lifestyle product, not a clinical service, signals where AI companies see the next consumer market. The radiologist backlash puts the gap between consumer AI launches and clinical evidence standards in sharp relief.
UAE establishes Federal Authority for AI and Data, absorbing three fragmented bodies into a single national regulator
On 14 June, Sheikh Mohammed bin Rashid Al Maktoum announced the creation of the Federal Authority for Artificial Intelligence and Data, a unified body consolidating AI oversight, digital government, and data regulation. The Authority absorbs the UAE AI Office, the digital government arm of TDRA, and the Emirates Data Office. It will set national AI and data policy, propose legislation, drive compliance across federal entities, and manage enforcement of the UAE's Personal Data Protection Law (PDPL), a question the market had been watching since PDPL passed in 2021 with uncertain enforcement structure.
morganlewis.com ↗Why it matters
the UAE has been one of the most active governments on AI strategy, but fragmented oversight across three separate bodies left regulatory jurisdiction unclear. Consolidating into a single authority with direct Cabinet reporting resolves that ambiguity and sets a structural model other Gulf and emerging-market governments are watching.
OpenAI releases LifeSciBench, 750 expert-authored tasks to test whether AI can handle real drug discovery work
OpenAI released LifeSciBench on June 17: a benchmark built by 173 PhD-level life scientists from biotech and pharmaceutical research settings, covering 750 tasks across seven workflows and seven biological domains. Unlike existing evaluations that rely on fact recall or clean prediction problems, LifeSciBench tests what actual researchers do: interpreting incomplete evidence, reconciling conflicting results, designing experiments under uncertainty, troubleshooting assays, and evaluating translational risk. The benchmark includes 1,062 task artifacts and 19,020 rubric criteria reviewed by 453 expert scientists.
openai.com ↗Why it matters
the bottleneck in AI adoption for drug discovery has been the absence of benchmarks that test research-level judgment rather than knowledge recall. A benchmark built and reviewed entirely by working life scientists, with 19,020 distinct scoring criteria, sets a new evidentiary standard for what "useful in drug discovery" actually means.
OpenAI alignment team: reinforcement learning on realistic scenarios can entrench broadly beneficial AI behaviour
OpenAI's alignment research team published findings this week showing that reinforcement learning applied to realistic scenarios targeting beneficial traits produces measurable improvements across dozens of benchmarks for aligned and safe behaviour, and those improvements generalise beyond the training domains and persist under adversarial pressure. The team reports that beneficial "personas" may become deeply embedded in models trained this way, suggesting RL may be a concrete path to building models with durable beneficial orientations rather than surface-level compliance.
alignment.openai.com ↗Why it matters
most AI safety work focuses on constraints and refusals. This research suggests RL can instead entrench positive dispositions that generalise and hold up under deliberate attempts to break them, a fundamentally different approach to the alignment problem with practical implications for how frontier models are trained.
Perplexity launches Brain, a self-improving memory system that builds persistent context graphs across agent tasks
Perplexity released Brain, a memory system for AI agents that constructs a persistent context graph linking memories across tasks, projects, decisions, files, and sources, so that agents start work with accumulated relevant context rather than from scratch on every session. Brain links each memory to its original source, continuously reorganises knowledge over time, and reduces task costs by retrieving and reusing prior work. The system is designed to improve answer correctness on complex, multi-session tasks where context accumulation matters.
perplexity.ai ↗Why it matters
the stateless agent problem (every session starts with zero memory of prior work) is one of the primary practical barriers to deploying AI agents in organisations with ongoing projects and accumulated institutional knowledge. A persistent, source-linked context graph that self-improves is a meaningful step toward agents that can carry a workstream, not just a conversation.
AI in the nonprofit sector4
Kenya in talks with OpenAI to establish East Africa's first AI Academy in Nairobi following Ruto-Altman meeting
Kenyan President William Ruto met with OpenAI CEO Sam Altman in Washington this week, with discussions centred on establishing the first OpenAI Academy initiative in East Africa, based in Nairobi. The proposed Academy would expand AI education, equip learners and educators with advanced digital skills, and support development of Kenya's AI workforce at a regional scale. Kenya has been actively positioning Nairobi as a technology and AI hub, and the OpenAI talks form part of a broader diplomatic push to attract AI companies to anchor their African operations in the country.
techtrendske.co.ke ↗Why it matters
OpenAI's first African Academy partnership, if confirmed, would create significant AI education infrastructure in a region where AI literacy investment has lagged the pace of AI deployment. It also signals that OpenAI's international expansion strategy now includes dedicated capacity-building commitments, not just product access.
Microsoft documents four concrete ways AI is reducing burnout and improving outcomes for public sector social workers
Presented at the 2026 APHSA National Human Services Summit, Microsoft published a case-based breakdown of four AI deployments transforming social work practice: AI-assisted pre-visit summaries giving caseworkers household context before home visits; real-time documentation tools that capture and structure notes during visits; knowledge retrieval systems surfacing relevant policy and resource information instantly; and pattern detection flagging elevated risk cases before they escalate. The deployments draw on Azure OpenAI and Copilot, with specific state and county agencies cited as pilots.
microsoft.com ↗Why it matters
social work has one of the highest burnout rates of any profession, driven substantially by documentation burden and information overload in high-caseload environments. Concrete deployments showing time savings and earlier risk detection, presented at the sector's peak professional summit, indicate AI adoption in human services has moved from pilot discussions to operational evidence.
PATH Chief AI Officer: Africa's health AI needs data infrastructure and local governance, not more capital, to get past the pilot graveyard
In a piece published June 17, PATH Chief AI Officer Bilal Mateen argues that Africa's health AI challenge is not a lack of capital or impressive models, but a lack of the structural rails that make AI deployable at scale: standardised data interoperability, local compute infrastructure, clear data ownership frameworks, and governance models that keep institutional knowledge in-country. Mateen draws on PATH's experience across multiple African health systems to argue that most promising pilots fail at replication not because the AI is wrong but because the underlying infrastructure cannot carry it, and that investors and funders systematically underinvest in that layer.
path.org ↗Why it matters
this is one of the clearest articulations from a senior practitioner of why the "AI for global health" funding wave has produced a pilot graveyard. For nonprofits and funders working in this space, the implication is direct: tool funding without infrastructure funding produces results that cannot scale.
San Jose's bottom-up AI training model, where staff build their own tools, is drawing interest from cities nationwide
San Jose's AI Upskilling Program, which trains city workers to build their own AI tools using low-code platforms rather than procuring vendor solutions, is gaining traction and inquiries from other US municipalities, according to Route Fifty reporting on June 18. The program runs cohort-based training where staff in departments ranging from planning to public works design and deploy custom AI tools matched to their actual workflows. San Jose officials credit the bottom-up approach with higher adoption rates and more sustained use than top-down vendor procurement, because staff are solving problems they defined themselves.
route-fifty.com ↗Why it matters
most AI adoption frameworks for government and nonprofits default to procurement (identifying a vendor and deploying a tool). San Jose's model inverts this: train the people closest to the work to build what they need. The early adoption and peer-city interest suggest this is a replicable model worth watching for nonprofits facing similar adoption resistance.