AI Next WaveJuly 30, 2026
The Shopping Bot Flagged the Claim. The Store Kept the Listing.
A new investigation tests whose interests AI assistants serve. Lloyds is targeting £2 billion in savings, Europe is funding seven computing hubs, and the chip shortage is reaching phones and other everyday products.
The Short Version
An AI shopping assistant can sound helpful while operating inside a marketplace that earns money from the sale it recommends.
Researchers at Columbia Law School tested Amazon's Alexa for Shopping and Walmart's Sparky against product listings with conflicting country-of-origin information. The assistants could identify suspicious "Made in USA" claims when asked. The marketplaces still displayed the products, and neither offered shoppers a country-of-origin filter.
The investigation has limits. It was conducted by two researchers at a center led by former Federal Trade Commission chair Lina Khan, and its examples don't measure the full prevalence of questionable labels across either marketplace. Chatbot explanations about corporate motives are generated outputs, not authorized statements from Amazon or Walmart. The observable behavior still deserves attention: the systems could find inconsistencies that customers weren't being warned about.
Today's other developments show the same separation between technical capability and accountable results. Lloyds Banking Group says technology including AI will help deliver about £2 billion in savings by 2030, while its chief executive declined to say how jobs may be affected. The European Union opened bidding for as many as seven AI gigafactories supported by up to €10 billion in public funding. Samsung says memory shortages could last into 2028 as AI servers consume more capacity and component prices squeeze its phone business.
Australia has also taken Telegram to court. Its online-safety regulator alleges that known videos of terrorist attacks remained available after user reports and that the platform failed to remove accounts and channels distributing the material. Telegram rejects the allegations.
AI can search, rank, flag, automate and recommend. People and institutions still decide what the system optimizes, which warning reaches the user, who absorbs the cost and whether a detected problem gets fixed.
The Assistant at the Checkout
A shopping bot can recognize a suspicious label and still lead you to the product
The Center for Law and the Economy at Columbia Law School published an investigation Thursday into how Amazon and Walmart's AI shopping assistants handle "Made in USA" claims.
Researchers Erie Meyer and Zachary Harris asked the assistants to find products and assess whether their origin claims looked legitimate. They also inspected listings for obvious conflicts, such as "Made in USA" in a product title alongside "Imported" in the country-of-origin field.
The report provides examples from both marketplaces. It says Walmart's Sparky could rank products by the strength of their origin claims and identify missing evidence of US manufacturing. Amazon's Alexa for Shopping could identify conflicts between a listing's promotional language and its origin information.
Amazon's assistant also refused a direct request for "Made in USA" fly-fishing reels while answering an equivalent request for products made in China. The researchers changed the wording to "in USA" and received a detailed answer identifying products that the assistant described as genuinely US-made. The result suggests that the information was available while the exact phrase triggered a block.
Neither company offers a country-of-origin filter, according to the researchers. Their one-category snapshot found 269 selectable filter options for dolls on Amazon and 299 on Walmart, with no origin filter on either site.
The study's conclusions require careful labeling. The researchers collected dozens of examples, yet they didn't publish a representative sample that would establish the rate of suspicious claims across billions of listings. Their testing captured the behavior of systems that can change through an update. The report comes from a policy center whose leaders have advocated stronger oversight of large technology platforms.
The chatbot responses also can't prove why executives or engineers designed a feature a certain way. A model can generate a plausible explanation of corporate incentives without access to the meeting, policy or code that produced the behavior. The stronger evidence comes from the repeated prompts, captured listings and reproducible query variation.
Amazon told Reuters that origin information is displayed on product pages when available and said it is working to make the information more accessible through Alexa for Shopping. Walmart didn't immediately respond to Reuters.
For shoppers, the finding exposes a weakness in AI-assisted buying. The assistant sits inside a marketplace that collects fees and benefits when a transaction occurs. Its recommendation may combine customer preference, product data, paid placement, inventory and the platform's commercial rules. A conversational answer can hide those competing priorities more effectively than a page of links.
A customer seeking a product made in a particular country should open the listing, inspect the origin field, visit the manufacturer's website and look for a specific manufacturing location. A seller's title and an AI-generated summary are weak evidence when the underlying fields conflict.
Small manufacturers have a commercial stake in this. A legitimate domestic producer may carry higher labor and compliance costs while competing with a seller using a false origin claim. Manufacturers can document their factory location, materials and final assembly, then monitor large marketplaces for listings using similar product names or claims. Screenshots, listing identifiers and purchase records create stronger evidence for a platform complaint or regulator than a general allegation.
Retailers could turn the same capability into a useful control. Run the assistant's inconsistency checks across listings before publication, route uncertain cases to a person and show customers the evidence behind an origin claim. The measures are straightforward: the share of flagged listings reviewed, confirmed violations, correction time, seller appeals and repeat offenses.
The investigation points to a broader product question. When an assistant can detect a problem but stays silent, technical accuracy tells only part of the story. The business rule governing disclosure may have more influence on the customer than the model's intelligence.
Work, Savings & the Missing Detail
Lloyds put a £2 billion number on technology savings without a job forecast
Lloyds Banking Group reported £4.3 billion in first-half pretax profit and introduced a strategy running through 2030. Chief executive Charlie Nunn said technology including AI would help the bank deliver about £2 billion in cost savings.
When reporters asked what that target could mean for employment, Nunn said Lloyds doesn't set targets for staff numbers.
The bank's own strategy materials describe where it expects technology to reach. Planned uses include AI-assisted financial guidance, faster mortgage processing, support for commercial relationship managers, automated claims handling, complaint and fraud workflows, and tools intended to improve employee productivity.
Lloyds says roughly 50 live generative AI use cases produced about £50 million of value in 2025 and that it expects more than £100 million of profit-and-loss benefit in 2026. Those are company-reported figures. The bank hasn't published the case-level calculations, independent verification or a breakdown separating revenue, avoided cost, staff time and risk reduction.
The £2 billion target is a forecast for a multi-year program. It shouldn't be read as savings already achieved or as an amount attributable entirely to AI. Reuters reported that the plan includes technology "such as AI," and Lloyds' presentation also describes process simplification, infrastructure changes and broader digital work.
Employees still need clarity because a productivity target can change headcount, hiring, role design and workload even when a company avoids setting a layoff number. Automation may remove repetitive review, let a team handle more cases or reduce the need to fill vacancies. It may also create new work in data quality, model review, customer escalation and compliance.
Managers can make the transition more credible by publishing workflow-level evidence. For a mortgage process, measure application-to-offer time, correction rates, customer complaints, manual touches and unequal outcomes across applicant groups. For fraud work, track prevented losses, false alerts, customer lockouts and appeal time. For an employee assistant, compare corrected work and customer results with prompt counts.
Workers can ask where the £2 billion is expected to come from. Which processes will change? Which tasks will disappear? Which decisions remain with a person? How will performance be assessed when AI contributes to the work? Which skills lead to the redesigned roles?
Investors need the same discipline. A savings forecast can support a valuation before the operating changes are complete. Reported profit, verified customer outcomes and recurring reductions in the cost of completed work will show whether the plan earns its number.
Public Compute & Private Opportunity
Europe opened bidding for seven AI gigafactories
The European Union launched a call for tenders Thursday for as many as seven AI gigafactories. The initiative carries up to €10 billion in EU and national funding and is intended to attract at least €20 billion in private investment.
The proposed facilities would combine advanced processors, software, cloud systems, high-speed networks and data centers. They would sit alongside 19 smaller AI factories already in Europe and provide computing for training, fine-tuning and running advanced models.
The Commission expanded the program from five planned sites to seven after interest from member countries. Applications close November 12. Successful bidders are expected to be announced in early 2027, and the Commission says the facilities should become operational within 18 months of contract signing.
Those dates and investment figures are plans. No winning consortia have been selected, the private capital has yet to be secured and the facilities haven't produced a customer result.
Europe is trying to reduce dependence on computing controlled elsewhere. That matters to researchers, startups, public agencies and regulated companies that need dependable capacity, local data handling or models developed under European rules.
Access will determine how widely the benefit spreads. A publicly supported facility can still favor organizations that already know how to apply for capacity, prepare large datasets and hire specialized engineers. Smaller businesses gain little from a powerful cluster if the queue, pricing or technical requirements keep them outside.
Program operators should publish who receives computing time, what each user pays, how long requests wait and which projects reach a validated result. Energy and water use also belong in the scorecard. A facility that expands model capacity while straining a local grid creates a cost that a national funding total can obscure.
Entrepreneurs and regional service providers can prepare for the work around the facilities. Secure data preparation, model evaluation, energy management, cooling, fiber installation, equipment maintenance and help moving a working prototype into production are credible needs. A useful offer should solve one bottleneck for an identified customer and measure time, quality or cost against the current method.
Public compute can lower an important barrier. Its return will depend on the research, services, companies and public value that emerge from it.
The Chip Shortage Reaches the Customer
AI servers are competing with phones for memory
Samsung Electronics reported record quarterly revenue of 171.5 trillion won and operating profit of 89.5 trillion won for the three months ended June 30.
Its semiconductor division generated 89.2 trillion won in operating profit as memory prices rose and AI-server demand remained strong. Samsung said it has signed multi-year supply agreements with the five largest global data-center companies and is approaching agreements with five more, without identifying them.
The company expects the shortage to intensify and extend into 2028. It aims to place about two-thirds of its memory output under longer-term contracts that typically include upfront payments and price floors.
That outlook comes from a supplier benefiting from high prices and long commitments. It is a forecast, and future demand could change if data-center projects slow, technical designs shift or new capacity arrives faster than expected.
The measured spillover already appears inside Samsung. Its mobile and network businesses recorded a 700 billion won operating loss for the quarter, which the company partly attributed to elevated component costs. Samsung's chip division gained from the same pricing pressure that hurt the part of the company selling phones.
This is how the AI infrastructure cycle can reach an ordinary buyer. High-bandwidth memory for accelerators differs from the memory inside a phone or laptop, yet manufacturers share capital, materials, equipment and production priorities across product lines. Suppliers move capacity toward the contracts with the strongest margins and longest commitments. Other customers can face higher prices or tighter availability.
Companies buying servers, PCs, industrial equipment or storage should ask suppliers how long pricing is guaranteed and which components have constrained lead times. A purchasing team can identify the few parts that would delay an entire product, approve alternatives and avoid placing every renewal in the same quarter.
Product teams can also reduce memory demand through software choices. Smaller models, compressed data, caching and narrower context can lower the amount of hardware required for a useful result. The savings should be measured on the complete workload, including any loss of quality or extra engineering time.
The factory opportunity is expanding with the shortage. Samsung plans a second semiconductor fabrication plant in Taylor, Texas, targeting production in 2030. Packaging companies are also raising capital spending. These facilities need equipment technicians, process engineers, contamination control, quality testing, power systems and specialized logistics.
Training programs should build around confirmed hiring dates and role requirements. A shortage creates urgency, while durable opportunity depends on whether local workers enter jobs with strong wages, retention and advancement.
Digital Safety & Platform Responsibility
Australia says Telegram left known terrorist videos online
Australia's eSafety Commissioner has begun civil proceedings against Telegram in federal court over alleged failures to detect and remove pro-terror material.
The regulator says 12 posts were reported by Australian users between July and October 2025. Its court filing alleges that Telegram failed to remove 10 of them or suspend the accounts responsible. The material included videos associated with the 2019 Christchurch mosque attack and the 2022 Buffalo shooting.
The regulator also alleges that some reported content remained available for as long as three weeks and that known attack footage had been uploaded nearly three months before removal. It is seeking court orders and a financial penalty. The maximum available civil penalty is A$54.6 million.
These are allegations. The court hasn't ruled on them. Telegram says it will contest the case and points to thousands of extremist communities it says it blocked during 2026.
The case matters because the disputed material was already known. Platforms can use digital fingerprints to recognize copies of previously identified videos, while users and investigators can supply additional reports. Detection still has to connect to removal, account action and a response to the person who reported the content.
Messaging services also carry legitimate communication, including reporting from conflict zones and evidence of abuses. A blunt removal system can erase material that journalists, researchers or courts need. A responsible process needs context review, preservation for authorized investigators and an appeal path.
Community organizations, schools and employers can prepare for harmful material without asking people to become content moderators. Give users a reporting route, preserve only the evidence needed for escalation, limit repeated exposure and provide support after a violent incident spreads online.
Product teams can test the full response chain. Seed a controlled copy of known prohibited material, confirm that detection works, time the review and verify that related accounts and reposts receive appropriate attention. Accuracy, removal time, repeat exposure, appeals and reporter notification all matter.
The Australian case will test whether a platform's stated safety effort met a legal standard. It also shows why a detection tool has little public value when the operating process around it fails to act.
Opportunity Radar
Marketplace claim verification for smaller manufacturers
Small manufacturers often lack the staff to search major marketplaces for false origin, safety or sustainability claims that undercut legitimate products. A trade association, compliance firm or specialized software provider could monitor a defined product category, capture conflicting listing evidence and prepare cases for platform or regulator review.
The buyer could be a manufacturer, industry group or state economic-development office. The service should validate that it finds material violations with a low false-positive rate and that its evidence produces corrections. Counts of scanned listings have little value without removal, corrected claims or recovered sales.
Evidence services for AI savings programs
Large organizations are attaching big savings targets to AI and digital transformation while many individual use cases remain difficult to verify. Operations, finance and risk specialists can help an organization establish a baseline, trace each claimed benefit and distinguish cash returned from employee time made available.
A credible pilot would follow one workflow for 60 to 90 days and measure total cost, corrected output, cycle time, errors, customer impact and staff workload. The provider earns trust by identifying projects that should stop as well as those that deserve expansion.
What You Can Do With This
If you shop through an AI assistant
Treat the answer as a recommendation from the store that may profit from your purchase. Open the product record, compare conflicting fields and verify important claims with the manufacturer or an independent source.
If you manage an AI program
Choose one case where the system can detect a problem and trace what happens next. Identify who receives the warning, who can act, how quickly the record changes and how a customer or employee can challenge an error.
If you lead a team facing a savings target
Ask for the workflow behind the number. Separate eliminated spending, avoided future cost, released staff capacity, added revenue and risk reduction. Tell employees which tasks and skills will change before the target becomes a surprise.
If you depend on hardware
Map the memory, storage and packaging components that could delay your product or infrastructure. Secure pricing where the risk is material, qualify alternatives and test whether a more efficient workload can reduce the hardware you need.
The Bigger Picture
The shopping investigation gives today's news a useful test.
Amazon and Walmart's assistants could identify suspicious product information. Lloyds can describe where AI may streamline banking. Europe can fund computing at continental scale. Samsung can commit memory capacity years ahead. Telegram can operate detection and moderation systems across a billion-user service.
Each capability enters an institution with its own incentives, rules and limits.
A marketplace earns from transactions. A bank rewards lower cost and higher returns. A government wants strategic capacity. A chipmaker favors long contracts and profitable products. A messaging platform balances growth, privacy, access and legal obligations.
Those incentives shape the result people receive. The assistant may withhold a warning. The savings program may change a job before the workforce plan is public. The computing hub may favor organizations already equipped to use it. The server boom may raise the cost of another device. The safety system may recognize harmful material after people have already seen it.
Useful oversight begins where capability meets action. Show the evidence behind the recommendation. Publish the source of the savings. Measure who receives public computing. Trace component costs into customer prices. Test whether a report changes what remains online.
The strongest technology can detect, predict or generate. Credibility grows when the surrounding system makes the result visible, correctable and accountable to the people who carry its consequences.
References
Reuters: EU gigafactory tender, private-investment target and expected timeline, July 30, 2026
Samsung Electronics: Second-quarter 2026 results and memory-supply outlook, July 30, 2026
Australia's eSafety Commissioner: Allegations and court action against Telegram, July 30, 2026
Reuters: Telegram contests Australia's allegations over pro-terror material, July 30, 2026