In a busy emergency room, someone has to decide who gets seen first. Hospitals are starting to test AI programs for that job, and a study from UT Southwestern Medical Center, published this week in npj Digital Medicine, shows how easily those programs can get it wrong. Researchers used real records from the emergency departments at Stanford and at Beth Israel Deaconess in Boston. They paired patients who looked equally sick when they arrived, knowing that one of each pair later got worse. Then they asked three AI models which patient should come first. On their own, the models usually chose correctly.

The results changed when the researchers added a single detail to one patient’s file. A line such as “17 visits in the past 12 months,” or a history of psychiatric emergencies, made the models more likely to move that person down the list, even when that was the patient about to get worse. In some tests, close to one in ten patients was wrongly downgraded. Emergency staff sometimes call these patients “frequent flyers,” and many keep coming back because they are chronically ill, have no regular doctor or have nowhere else to go. The experiment used past records, so no real patient was harmed. Dr. Philip Jarrett, the study’s lead author, says each hospital should check these tools against its own past patients before letting them help decide who waits.

Teen test accounts could download dozens of AI image-manipulation apps despite parental controls, the Tech Transparency Project reported on October 19. Researchers created four accounts registered as 13-year-old boys, comparing default settings with parent-supervised accounts on Apple and Google. The supervised accounts could access 24 apps on Apple and 51 on Google that TTP classified as nudification tools. Its definition includes generating images in underwear or swimwear. Apple disputes that classification for some mainstream editors and says it prohibits nudification apps. Google says it bans them too and suspended apps that violated its policies. These were controlled tests, not records of children using the services.

An age restriction depends on the accuracy of the label attached to each app. TTP says both stores rely on developers’ descriptions to assign those ratings. A parent asked to approve a download can therefore be making a decision with incomplete information about what the software allows. The store has already reviewed the product, placed it in a category and made it available. That chain should be checked against the app’s actual functions, including features added after approval. TTP also found that some apps removed from store listings still worked on its test phones. Taking down a listing closes one route to installation. Families also need clear notice about an installed app whose capabilities have made it unsafe.

Siemens and NVIDIA are testing AI tools that propose production plans at Siemens’ electronics factory in Erlangen, Germany. Presented on October 19, the system generates alternatives when orders arrive and shows planners how each option performs against selected indicators. Proposals are checked in a digital twin, a virtual representation of the plant, before implementation. Siemens says agent-based planning could cut costs by up to 20%. That is a company claim about potential savings, not a guarantee for every factory. The plant, which makes power electronics and automation technology, employs about 1,000 people.

The indicators chosen by management will shape which plan looks best. A cheaper sequence could still leave a maintenance team too little time to inspect equipment or make an operator repeat a difficult task. Those effects need a place in the comparison. Workers who know the machines should be able to point out what the simulation misses and have that objection recorded. A planner who changes a proposed sequence also needs enough time to explain the decision, without pressure to accept the software’s first recommendation. Testing a plan virtually can avoid costly mistakes on the line. Its value will depend on how objections made in the workshop reach the people approving it. Before a proposal becomes the next shift’s instructions, someone must be responsible for checking it and authorized to stop it.

Anthropic has proposed committing 5% of its shares to a new Public Benefit Fund, subject to stockholder approval. Announced on October 18, the fund will be wholly owned by the company. Possible uses include donations to nonprofits and direct distributions to particular groups. Anthropic names disruption to workers as one reason for creating it and says further details about governance will follow. The promise recognizes that a company selling automation may also have a responsibility to people affected by it. It remains a commitment of equity, with no payment schedule or rules identifying who will receive support in the announcement.

A worker whose income falls after an employer adopts AI would need to know whether the fund offers help, how to apply and who reviews a refusal. Those decisions determine whether a large corporate pledge becomes usable assistance. A company can finance valuable work while retaining the power to choose recipients and define acceptable outcomes. Describing its purpose as public benefit leaves that power intact until the governing rules give affected people a role. The fund could earn trust by publishing selection criteria and reporting actual distributions separately from the value of its shares. Workers and their representatives should be able to question those decisions. For someone trying to cover next month's rent, the useful figure will be the amount received and the date it arrives.

The World Bank is putting small AI on the agenda of its Global AI and Digital Summit, which opens in Seoul on October 19. A session on October 20 will examine tools adapted to local needs. The Bank describes small AI as a design approach built around a specific problem and the devices people can access, including basic phones and sensors. Some applications can run models locally or use familiar channels such as text messages and voice. The ambition is to make useful services work with limited connectivity. These are proposals for development, with benefits that still have to be demonstrated in use.

A public agency buying such a service would still need a budget for maintenance and someone responsible when an answer fails. Running a model on a modest device can reduce dependence on a remote server, but the device needs power, repairs and software updates. Local data also requires people who can check whether an answer fits the language and circumstances of the person asking. Those costs belong in the contract from the start. Otherwise, a successful demonstration can leave a clinic or farming cooperative with equipment nobody can maintain. The Bank itself acknowledges the need for digital skills and trusted local institutions. Procurement should specify who tests the tool, who pays after the pilot ends, and where a person can challenge advice that has caused harm.

Google’s paid News AI pilot now includes over 400 publications, the company announced on October 15. CNN confirmed a multi-year agreement granting content rights for experimental Google features. The pilot includes article overviews on Google News publisher pages and customizable audio briefings, with attribution and direct links. Google says it pays participating outlets for delivering and licensing that material. Neither announcement discloses the payment amounts. The arrangement gives a publisher revenue before a reader decides whether to leave the summary and open the original article.

A newsroom still pays for reporting even when a reader stops at the overview. A licensing fee can help cover that expense, but an editor also needs to know what happens to the relationship with the audience. Does the summary lead someone to subscribe, or satisfy the question without a visit? A prominent link offers a route back to the source; its presence alone cannot answer either question. The useful measures would include visits that become subscriptions and readers who return without passing through Google again. Those figures should be visible to participating publishers, alongside any mistakes introduced by the summary. A large outlet may have staff to compare the fee with changing traffic. A small newsroom needs the same information before deciding whether the agreement supports another reporter’s salary or merely replaces an income stream that is shrinking.

Decomp.dev has disabled its public project index and stopped accepting new projects. The site tracks efforts to reconstruct readable source code from released video games. Its maintainer says AI-assisted copies and rushed adaptations are damaging the volunteer community behind that work. Existing project pages and historical data will remain available. The notice accepts that AI can help when guided and reviewed properly. The maintainer says a lone hobbyist cannot keep judging the quality of hundreds of projects.

A volunteer who spends an evening explaining a difficult function contributes something a completion percentage cannot record. Another person learns how the program works and can help with the next problem. If recognition goes chiefly to whoever announces a finished version first, that teaching becomes harder to sustain. Someone still has to check the result, trace mistakes and answer questions after the announcement has circulated. Faster output can therefore increase the unpaid work required from experienced contributors. This is an awkward limit for a hobby measured through progress bars. The satisfying number can rise while the conditions for learning together deteriorate. A project can choose to use AI and still require a contributor to explain each change and accept responsibility for fixing it. Those requirements need to appear before a submission is counted as progress.

Seven U.S. senators asked Meta to explain whether it claimed research tax credits for building and operating AI data centers. In a statement published October 14, they questioned reports that the company treated those facilities as experimental pilot models. They want the amount attributable to data centers and the reasoning behind the claim. The request does not establish that Meta broke tax law. Responding to earlier reporting, company spokesperson Andy Stone said Meta uses incentives Congress created to encourage domestic investment. That defense leaves the eligibility of particular expenses open to scrutiny.

A research tax credit lowers a company's tax bill, giving the public a financial stake in what qualifies as an experiment. An AI facility can contain genuine research while also serving as a production site. Deciding which costs receive support requires a closer account of what is being tested and what the company already knows how to build. Calling an entire expansion experimental risks making size itself a reason for subsidy. Rejecting every new facility as routine could also overlook work whose outcome is uncertain. The useful distinction has to survive an examination of equipment purchases and actual tests. Meta's answer should identify the claimed expenses and explain how each relates to research. Legislators and taxpayers would then have something firmer to assess than the industry's broad promise of future breakthroughs.

Google DeepMind and Singapore’s GovTech announced an AI scam detector on October 14 that analyzes conversations on the phone itself. Built with Gemma, Google’s lightweight model, it screens incoming messages and examines suspicious exchanges for signs of manipulation. The developers say this local design keeps conversations out of government and Google servers. Access will require the user’s consent. Integration into applications is planned in phases, with no launch date given. A warning would ask someone to pause before sending money or sensitive information.

A person receiving an urgent message from a supposed bank could get a warning while still deciding whether to obey it. That pause offers a useful defense against fraud that depends on haste or trust. Keeping the analysis on the handset also avoids creating a central store of private exchanges for this purpose. Yet an alert can be wrong, and its wording can shape how someone reads a genuine request. The application's terms will determine which conversations the feature can access. Users need a plain explanation of that reach and a practical way to withdraw permission. They should also be able to inspect the reason for a warning without treating it as a verdict on the person writing. Before a rollout, the developers should publish how often the system misses scams and how often it flags ordinary messages. Those results would let citizens judge the protection being offered.