AI Next WaveAugust 18, 2026
Google Bid $10 Million for the Work Spirit Airlines Left Behind
A bankruptcy sale turns employee emails and Teams chats into AI training material. Robots face an ROI test, Baidu shows why AI growth may fail to rescue a weakening core business, and an accessibility app shows what useful AI looks like.
The Short Version
Spirit Airlines stopped flying in May. The work its employees left behind still has a buyer.
Google won a bankruptcy auction for a collection that includes about 100 million employee emails, 500 million Microsoft Teams chats, calendars, spreadsheets, presentations, financial databases and custom software. The price is $10 million. Google says the material can improve its products and AI models.
The sale still needs approval from a federal bankruptcy judge at a hearing Wednesday. Google says a third party will remove personally identifiable information and that the package excludes customer and credit-card data. Those safeguards narrow the risk. They also leave a harder issue for every organization: work records created for one employer can become training material for another company after the original business disappears.
Today's other developments test AI against results people can see. China's humanoid-robot industry is moving from viral demonstrations toward factory, hotel and service work, while typical robot costs remain well above one estimate of a two-year payback price. Baidu's AI-powered business grew 25% in the second quarter, yet a 19% fall in online marketing revenue pulled total company revenue down 4%.
In Egypt, a blind 20-year-old founder built an app that describes surroundings through a phone or smart glasses. It is being used by thousands of people in 140 countries, according to Reuters. In California, 29 states are beginning a trial over allegations that Meta designed Facebook and Instagram in ways that harmed young users and improperly collected children's data. Meta denies wrongdoing.
These stories put a useful standard around the AI economy. Data has value when someone has the right to use it. Technology earns its place when it completes a task reliably, at a defensible cost, for a person who needs it.
The Data Left After a Company Closes
Spirit's employee messages have become an AI asset
Google's winning bid covers the digital record of how an airline worked. The collection includes communications, marketing material, human-resources information, project documents, financial records, audits and presentations. AI companies prize this kind of data because it contains decisions, revisions, disagreements and completed work. Public webpages rarely show that full process.
The scale is striking. Axios reported that the package includes about 100 million emails and 500 million Teams chats. Mercor, an AI data company, offered $7.5 million. Google raised the price to $10 million.
The court has yet to approve the sale. Google also says it will receive a de-identified dataset with no customer information or personally identifiable information, and it has agreed to avoid trying to identify users. Those are material protections.
De-identification cannot answer every governance question on its own. A message can reveal an operational method, a vendor relationship or a sensitive decision even after names and email addresses are removed. People who wrote to colleagues during a flight disruption or restructuring were producing company records. They probably did not write with model training in mind.
Organizations should decide how valuable work data may be used before a bankruptcy, acquisition or vendor dispute forces the issue. Retention schedules should separate records needed for law and operations from material kept because storage is cheap. Employment policies and customer contracts should describe secondary uses plainly. A company considering a license or sale needs to check ownership, confidentiality, trade-secret exposure and whether de-identification preserves enough useful context to justify the risk.
AI developers face a related test. A large archive may contain more noise, obsolete practice and internal bias than durable expertise. The useful evaluation is downstream: does training on the material improve a defined task on current data without reproducing sensitive content? The $10 million bid establishes a market price for access. It does not establish model quality or a financial return.
The Robot Has to Finish the Shift
China's humanoid makers are moving past the demonstration
More than 300 companies are expected to show over 2,000 exhibits at this week's World Robot Conference in Beijing. Organizers expect more than 150 product launches. Unitree, one of the largest humanoid makers by sales volume, is also making its Shanghai market debut after a retail offering that Reuters says was over 8,000 times oversubscribed.
The attention is running ahead of broad deployment. Humanoid robots have begun handling narrow jobs such as hotel food delivery, industrial inspection and some assembly-line tasks. Large-scale adoption beyond pilots has yet to arrive.
Cost explains part of the gap. Guotai Securities estimates that an industrial humanoid would need to cost about 160,000 yuan, including maintenance, to repay its purchase within two years against a worker earning 80,000 yuan a year. MERICS says typical robots currently cost 300,000 to 500,000 yuan.
Hardware price is only the start. A buyer also pays for integration, transport, safety controls, supervision, charging, repairs and the production time lost when the machine stops. One robotics analyst estimates that 50% to 70% of humanoids produced this year could end up in data factories, collecting training data instead of doing paid work for customers.
The industry is beginning to test the right things. Beijing's World Humanoid Robot Games will add packing, warehousing, assembly, retail, office services and electric-vehicle charging to the races and fighting that attract cameras. Those tasks expose whether a robot can identify unfamiliar objects, repeat a motion, recover from an error and continue without an engineer stepping in.
A company exploring robotics should begin with one repetitive, measurable job in a controlled space. Record how people perform it now, including errors, injuries, waiting and rework. Then count successful task cycles, human interventions, downtime and total cost. A robot that completes 95% of a demonstration can still fail the business case when a person has to rescue the remaining 5% at unpredictable moments.
Workers will shape the outcome. Technicians, process owners and frontline employees know where materials vary, where safety procedures break down and which exceptions matter. Their knowledge determines whether automation fits the operation or becomes an expensive visitor on the floor.
AI Growth Cannot Hide the Rest of the Business
Baidu's quarter shows why one rising segment may not be enough
Baidu reported second-quarter revenue of 31.33 billion yuan, down 4% from a year earlier and below the average analyst estimate compiled by LSEG.
The company's online marketing revenue fell 19% to 13.1 billion yuan as weak consumer spending and China's property downturn pressured advertising budgets. Its Core AI-powered Business, which includes cloud infrastructure and AI applications, grew 25% to 12.5 billion yuan.
That is reported revenue, not a forecast. It shows paying demand for Baidu's AI services. It also shows the limit of celebrating a fast-growing line without reading the full statement. The AI business remains smaller than the marketing segment, and higher spending on infrastructure and talent can pressure margins while the transition continues.
The same problem appears inside companies adopting AI. A successful assistant can save time in one workflow while the surrounding service loses customers, quality or revenue. Time saved has value only when the organization can use the released capacity. A stronger measure connects the tool to the business result: cases resolved, sales completed, errors avoided, waiting reduced or customers retained.
Leaders should keep the baseline visible. Record the cost and performance of the workflow before deployment, include integration and review time, then compare the full result after adoption. License activity and generated answers describe use. They do not prove that the operation improved.
Useful AI, Built From a Problem
ScribeMe gives blind users more information in the moment
Mark Morad began building ScribeMe after screen readers failed him in school. Charts, diagrams and equations could disappear into silence. He taught himself to code through YouTube tutorials and had a working app by 2023.
The app now uses a phone camera or Meta smart glasses to describe surroundings and answer questions in real time. It supports 20 languages, including Arabic. Reuters reports that thousands of people across 140 countries use it. A free version is available, with additional features priced at $20 a month or $200 a year.
For Mark's 16-year-old sister Karen, the system can help locate a classroom, recognize shops, choose clothing and study. On a safari, it described animals, their distance and their movement while she listened through the glasses. The hands-free design leaves a user's hand available for a cane.
This is a grounded use of multimodal AI because the problem comes first. The system turns visual information into a format a blind person can use during an actual task. The value lies in greater independence and access, not the novelty of generating a description.
The risk also becomes concrete. A wrong description of a shirt is inconvenient. A wrong description of traffic, medication or a platform edge can be dangerous. Product teams should test accuracy by setting, track how often users ask for clarification and make uncertainty audible. Camera data, battery life, connectivity and subscription cost can all determine whether the service works outside a demonstration.
Accessibility teams should include blind and low-vision users throughout development and pay them for that expertise. Their feedback improves the product for its intended users and often reveals interface problems that affect everyone.
A Product Design Goes on Trial
Twenty-nine states are challenging Meta over youth safety
A federal trial beginning Tuesday in Oakland will test allegations that Meta designed Facebook and Instagram in ways that harmed children and teens, misled consumers about safety and collected children's data without proper parental consent.
Four states, California, Colorado, Kentucky and New Jersey, are leading the case for a bipartisan group of 29. An eight-person jury will issue an advisory verdict. U.S. District Judge Yvonne Gonzalez Rogers will make the final decision.
The states are seeking penalties and a nationwide order that could require age restrictions, remove infinite scroll and change other product features. The amount remains unsettled. Meta says potential penalties could reach $1.4 trillion. The states have suggested a figure closer to $200 billion without making a final demand.
Meta says the claims are unsupported, argues the states have failed to show actual harm and points to protections it has built for teenagers. Mark Zuckerberg and Instagram head Adam Mosseri are expected to testify during the multiweek trial.
The case matters beyond social media. AI systems increasingly decide what a user sees next, which notification arrives and how long a feed can hold attention. A safety claim needs evidence from the product people actually use, including workarounds and behavior over time.
Teams designing for young people should test outcomes such as sleep disruption, repeated exposure to harmful material, unwanted contact, age-check errors and appeals. Engagement can remain a business measure. It should never stand in for well-being.
Opportunity Radar
Field testing for accessible AI
Companies and public agencies are adding camera-based assistants, voice interfaces and automated support without enough testing by people with disabilities in everyday settings.
An accessibility specialist could organize paid field panels for one product category, such as transit, retail or education. Buyers would receive task-completion results, safety failures, user corrections and recommendations tied to specific environments. The service must validate that testing changes the product and that participants are compensated for their expertise. A compliance checklist would miss the practical value.
What You Can Do With This
If your work lives in company systems
Treat email, chat and shared documents as durable business records. Keep sensitive reasoning in approved locations, follow retention rules and ask how records may be used after an acquisition, closure or vendor change.
If you approve AI or automation
Choose one completed outcome and establish its current cost, error rate and cycle time. Include human review, integration and downtime in the test. Expand only after the result survives ordinary operating conditions.
If you build customer technology
Recruit the people who will rely on it, pay them for testing and observe the product in context. Track corrections and abandoned tasks alongside usage.
The Bigger Picture
AI is gaining access to the records, machines and decisions that make organizations run.
Spirit's archive captures years of work after the airline has gone. Chinese robot makers are trying to turn trained movement into dependable labor. Baidu is converting AI demand into revenue while its older engine weakens. ScribeMe translates a camera view into practical independence. Meta's trial asks whether product choices carried costs for young users that the company should have addressed.
The strongest evidence appears after deployment. A dataset improves a defined task. A robot completes a shift with few interventions. A cloud service produces enough value to support the cost. An assistive app works safely in the places a person needs it. A platform's protections change what young users experience.
That standard is demanding and useful. It gives workers, customers, investors and public officials something better than a demonstration to judge.
References
Reuters: China's humanoid-robot industry faces its commercial test, August 18, 2026
Reuters: Baidu's AI-powered business grows while advertising revenue falls, August 18, 2026
ScribeMe: Official product site for the AI-assisted accessibility app
Reuters: Blind Egyptian entrepreneur Mark Morad builds ScribeMe for daily use, August 18, 2026
Reuters: Meta faces a 29-state federal trial over alleged harm to young users, August 18, 2026