A court has paused Google’s purchase of a bankrupt airline’s work emails.
Spirit Airlines stopped flying this year, and what went to auction was not the planes but the paperwork: about 100 million emails from 80,000 staff accounts, 500 million Teams messages, and 175,658 employee records reaching back to 1986. Google won it on August 14 for $10 million, roughly ten cents an email.
Then the people who wrote it objected. On August 19 the union that represented Spirit’s cabin crew told the bankruptcy court that the deal’s privacy protections were built for passengers, not staff: customer profiles and loyalty records were carved out, disciplinary files and payroll history were not. It does not want to stop the sale, only to have employees scrubbed the way customers already were.
- Aug 14Google wins the auction at $10 million. Mercor.io is the backup bidder at $7.5 million.
- Aug 19The flight attendants’ union objects. The protections cover passengers, not staff. The same day a rival, Micro1, bids $12.5 million.
- Aug 19Judge Sean Lane halts the sale and adjourns the hearing.
- Sept 9The new hearing date. The judge weighs the objection and the higher late bid.
Sources: Bloomberg Law, Fortune, SiliconANGLE, Forbes, Inc. · US Bankruptcy Court, S.D.N.Y.
Judge Sean Lane halted the sale and moved the hearing to September 9. The same day a rival bidder offered $12.5 million.
every message you send from a work account belongs to your employer, not to you. What is new is that the ordinary texture of your working life now has a resale value, and the buyers build AI. Nobody at Spirit was asked, because nobody had to be. What changed is that a group of laid-off employees noticed, hired lawyers, and stopped a Google deal in five days. The protections in a deal like this get written for whoever has someone in the room.
The estate sale covers Spirit's Microsoft 365 and collaboration corpus (about 100 million emails across roughly 80,000 accounts since 2018, about 500 million Teams items, 17,082,644 OneDrive items, 20,577,677 SharePoint items), 516 code repositories totalling roughly 30 million lines of custom software, 7.5 billion revenue-system transactions dating to May 2008, and 175,658 employee records reaching back to August 1986.
Excluded and marked "not included": 97.5 million customer profiles, 50.2 million Free Spirit loyalty records, contact-centre recordings, marketing lists, web analytics and regulatory complaints. The sale documents require Spirit to engage a Deidentification Agent to remove or transform data elements pre-transfer, certified against CCPA standards and federal health-privacy rules while preserving referential integrity across the set.
Google's winning bid was $10m, filed Aug 14; Mercor.io was backup bidder at $7.5m. AFA-CWA filed its objection in the US Bankruptcy Court for the Southern District of New York, arguing the de-identification architecture is consumer-facing while the payload is disproportionately employee-facing. Judge Sean H. Lane adjourned the sale hearing from Aug 19 to Sept 9. Micro1 Inc. (founder and CEO Ali Ansari) submitted a late $12.5m offer on Aug 19. Bankruptcy practitioners quoted in coverage note that a properly noticed auction is rarely undone, though courts sometimes entertain late bids that materially increase creditor recovery. Anonymisation of free-text corpora is substantially harder than of structured records, and no public detail of the method has been released.
OpenAI cut the price of its top model, the one it had left out of two earlier rounds of cuts.
In Issue #4 we noted that OpenAI had discounted its cheaper models and left GPT-5.6 Sol alone, and called that the tell for where it thought its advantage lay. On August 21 it cut Sol too, by a third on the expensive half. Reuters read it as an answer to Anthropic and to Chinese rivals. It runs as a promotion through November 21.
All four bars share one scale. Promotional through Nov 21 · Source: Reuters, Aug 21 2026
every app that has sprouted an AI feature buys at these prices, which is why the summarise button keeps appearing in software you already pay for. Watch that word “promotional”: if it holds in November, prices keep falling.
ChatGPT began showing ads in 31 more countries Aug 24
On August 24 OpenAI started serving advertisements inside ChatGPT across 31 European markets. Americans have had them since February. They sit in the conversation, labelled, and OpenAI says advertisers never see what you typed. Only Free and Go accounts see them.
the thing you ask questions to used to have one customer, which was you. It now has two, and the second one pays. That makes the free version of the most-used AI tool on earth an advertising product, in the place people go for medical and money questions at eleven at night.
ChatGPT has a teen version, and it assigns people to it on an age estimate Aug 18
Launched for ages 13 to 17, and applied automatically whenever the system estimates a user is under 18. It tightens what the chatbot will discuss around self-harm, eating disorders, violence and sexual content, adds a study mode, and gives parents optional controls. Issue #5 covered the safety tests behind it. This is those tests shipped.
if you have a teenager, their ChatGPT changed this week without either of you touching anything. The age is estimated, so it misfiles both ways, and the company writing the protections is the one counting the users.
OpenAI reduced GPT-5.6 Sol API pricing on Aug 21 2026 from $5/$30 to $4/$20 per million input/output tokens for standard short-context use, a 20% input and roughly 33% output reduction, described as promotional and running to at least Nov 21 2026. It applies to the pay-as-you-go API, Codex credits and eligible ChatGPT Work plans; subscription pricing is unchanged. For comparison, Claude Fable 5 lists at $10/$50 and Claude Opus 5 at $5/$25. This is the third GPT-5.6-family reduction in about a month and is this issue's frontier-race item.
ChatGPT Ads expanded to 31 European markets on Aug 24 2026, six months after the February US pilot and after Canada, Australia, New Zealand, the UK, Mexico, Brazil, Japan and South Korea. Ads serve to Free and Go tiers only, are labelled and separated from model output, and OpenAI states advertisers do not receive user conversations and that advertising does not influence generated answers. EU personalised targeting relies on explicit consent under GDPR.
ChatGPT for Teens launched Aug 18 2026; placement follows stated age or an age-estimation signal, with added policy layers on self-harm, disordered eating, violence, dangerous activities and sexual or graphic content, plus Study Mode and homework-integrity reminders.
ChatGPT can now record your working day, and on a work Mac your employer decides whether you may switch it on.
On August 13 OpenAI added a feature called Computer History to its Mac app. Turned on, it keeps a timeline of what you opened, clicked, typed and visited, so you can ask what you were working on before lunch and get a real answer. It takes no screenshots and no audio, the deliberate contrast with Microsoft’s Windows Recall, and it is off until you switch it on. Then there is the last line: on a Business or Enterprise account, an administrator decides whether you may enable it at all.
the feature is genuinely useful, and the thing to notice is who holds the switch. On your own machine it is yours. On a work machine it belongs to whoever runs IT, and a record that exists can later be asked for by a lawyer or an auditor. Ask your employer’s policy before you turn it on. Then read the top of this issue again: work makes a record, records are data, data has an owner, and this week some flight attendants proved the owner can at least be argued with.
Computer History rolled out in the ChatGPT macOS desktop app on Aug 13 2026. It records interaction events sourced from operating-system activity and accessibility APIs (clicks, keystrokes, keyboard shortcuts, application switches, site visits) rather than screen or audio capture, which is the substantive difference from Microsoft's Windows Recall.
It is opt-in, available to Pro, Business and Enterprise tiers on macOS, with EEA, Swiss and UK availability stated as later. Per-app and per-site inclusion lists, pause, per-entry deletion and full disablement are user-controlled; Business and Enterprise administrators hold an org-level policy control over whether end users may enable it. OpenAI states the captured history is not used for model training. The timeline is also readable by Codex, OpenAI's coding agent.
An AI designed protein binders against 14 of 15 targets, and two outside labs tested them.
A binder is a molecule shaped to stick to one target and nothing else, like a key cut for a single lock. Almost every antibody drug works this way, and designing one from scratch is slow.
On August 18 Anthropic published what happened when it pointed Claude at the problem. One AI agent researched the targets, chose which public scientific tools to use, generated 1,320 candidate molecules and ranked which to send for manufacture. Then two independent companies, Adaptyv Bio and Twist Bioscience, built them and tested them in a lab. 354 stuck, with at least one working against 14 of the 15 targets: a success rate of 22.6% to 35.1%, against the 10% to 15% Anthropic cites as normal for the field. It failed too. Against maltose-binding protein, none of the 90 designs worked.
1,320 designs, 354 worked, at least one against 14 of 15 targets. Against maltose-binding protein, none of 90 designs worked. All bars share one scale · Source: Anthropic technical report, Aug 18 2026, tested by Adaptyv Bio and Twist Bioscience
this is not a drug and nobody has been treated with anything. What got faster is the first step of a long road, and the steps after it are unchanged: cells, then animals, then people, which is where most candidates die. What is genuinely new is the checking. Anthropic published the designs and outside labs made the molecules by hand, so the result does not rest on the company’s own word. In a year of AI claims nobody could reproduce, this one was. Ask who made it and who measured it. None of this is medical advice.
Disclosure: Human Terms is written with help from Claude, which is made by Anthropic. We cover them the same way we cover everyone else, and we say so when it comes up.
Anthropic published the report on Aug 18 2026, releasing designs, prompts and measurement data. Sixteen targets were selected; one was excluded for inconclusive data, leaving 15 assessable. 1,320 designs were generated at 30 per target per configuration; 354 confirmed binders resulted, with at least one binder against 14 of 15 targets.
Hit rates: Mythos Preview 26.7% and Opus 4.8 22.6% in multi-target mode (48 hours wall time, up to 12,500 NVIDIA H100 hours); Mythos Preview 35.1% in single-target mode (24 hours per target). The stated field norm of 10–15% is Anthropic's own characterisation of its field, not an independent benchmark. External validation was performed by Adaptyv Bio and Twist Bioscience.
Documented failures include BBF-14 and maltose-binding protein, the latter with zero confirmed binders across 90 designs. The agent operated under an expert-defined protocol and selected among publicly available computational design tools; this is agent orchestration of existing structural-biology tooling rather than a new structure-prediction model. No in vivo, animal or clinical work is involved.
Who checks whether a model is safe to release?
OpenAI’s Preparedness team had one unglamorous job: take a finished model before anyone outside saw it and ask whether it could help someone build a biological weapon, run a cyberattack on its own, or act in ways its operators could not pull back. Their answers set the rating that governed how it shipped.
Several outlets reported in mid-August that the team was dissolved at the end of July and its work parcelled out to other groups. OpenAI disputes that, and says the work was spread across the company rather than ended. Neither side has moved since. What gives the report weight is the pattern: this would be the third safety group folded in about two years.
we have spent two issues reporting what that machinery produced, including a model locked in isolated machines over a possible “critical” hacking rating, and a purpose-built hacking model rated “high” and put on sale. A framework is only as good as the people left to apply it. You do not have to pick a side to see why it matters who those people are.
Disclosure: Human Terms is written with help from Claude, made by Anthropic, which competes with OpenAI.
Reporting dated Aug 17 2026 states the Preparedness team was dissolved at the end of July 2026, with domain responsibility for biological and cybersecurity risk redistributed to other teams rather than retained in a dedicated unit. OpenAI disputes the characterisation. Re-checked through Aug 24 with no further statement from either side.
This item deliberately carries no capability figures: the disputed fact is organisational, and mixing it with benchmark numbers would imply a settlement that does not exist. Prior context: the AGI Readiness team was dissolved in 2024 and the Mission Alignment team closed in Feb 2026. The Preparedness Framework is the classification scheme that produced the ratings covered in Issues #5 and #6.
Four tech companies have borrowed about $223 billion this year to build AI.
Alphabet, Amazon, Meta and Oracle have sold close to $223 billion of bonds through August 20, more than double what the sector raised in all of 2025. A bond is a loan you can trade. These are investment grade, though the range runs from Microsoft at the top of the scale down to Oracle at the bottom of the safe band with a negative outlook.
Here is the part that reaches you. These four are now about 8% of the main US investment-grade bond index, and index funds buy the index, so if you hold a boring bond fund inside a pension you have been buying data centres without deciding to. Against that, a Wall Street Journal analysis this month put nine tech companies’ off-balance-sheet AI commitments near $3 trillion, versus about $600 billion actually spent in the past year.
The promises are about five times the visible spending; the borrowing is the part with a repayment schedule · Sources: CNBC (bonds, through Aug 20 2026); WSJ analysis via Forbes. Not investment advice
the buildout is shifting from company profits to borrowed money, and borrowed money is repaid on a schedule whatever the thing you built earns. That is not a prediction of disaster. It is a change in what happens if demand disappoints, because a promise can be renegotiated and a bond cannot. Nothing here is investment advice.
Bond issuance figures cover Alphabet, Amazon, Meta Platforms and Oracle, at nearly $223bn through Aug 20 2026, more than double the comparable 2025 total. Published totals vary with definition: some counts reach about $250bn for a wider issuer set and about $570bn for global AI-related issuance across 2026, so this issue uses the four-company figure and states its scope. Ratings span AAA to BBB with negative outlook; hyperscaler issuers are now roughly 8% of the US investment-grade index.
The Wall Street Journal's analysis of nine companies (Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX, AMD) puts off-balance-sheet commitments near $3tn: roughly $1.2tn in leases not yet commenced and roughly $1.9tn in purchase commitments, against about $600bn of trailing-twelve-month capital expenditure.
Pennsylvania made its data centre standards binding and prohibited nondisclosure agreements.
On August 18 Governor Josh Shapiro signed Executive Order 2026-05, turning a voluntary set of standards into a condition of doing business in the state. To build an AI data centre in Pennsylvania a developer must now bring its own power, pay the full cost of new electricity infrastructure rather than passing it to households, meet water and environmental requirements, hire locally, and sign a community benefit agreement.
Two provisions go further than anything we have covered in this thread: local approval must come before the state will even begin reviewing a permit, and nondisclosure agreements are prohibited. Critics note that an executive order binds state agencies rather than writing law, so a future governor can undo it.
the NDA ban is the piece to hold onto, because it is the one about you. Developers routinely ask local officials to sign confidentiality agreements, which means your township supervisor can be discussing a project next to your house and be legally unable to tell you about it. If a data centre is being talked about near you, ask at the next public meeting whether anyone at the table has signed something. The answer is informative either way.
Executive Order 2026-05, signed Aug 18 2026, directs all Commonwealth agencies to require data centre proposals to comply with the Governor's Responsible Infrastructure Development (GRID) requirements, previously voluntary: developer-supplied or procured generation with clean-energy requirements; full funding of new electricity infrastructure without cost-shifting to ratepayers; transparent local engagement; local hiring and community benefit agreements; water conservation; local approvals before Department of Environmental Protection permit review begins; a binding Consent Order and Agreement; and annual energy and water reporting.
Nondisclosure agreements are prohibited for data centre projects, and AI data centre proposals are removed from the PA Permit Fast Track Program. State-cited counts: 100+ projects in public databases, 58 engaged with DEP, 15 with at least one permit application, 5 fully permitted for a first phase. An executive order binds executive agencies and is revocable by a successor; it is not statute, which is the substance of the criticism that it falls short. Continuation of Issues #1, #2 and #3 on data centres.
Three things worth keeping an eye on.
- Sept 9September 9: the Spirit sale hearing. Judge Lane weighs the flight attendants’ objection and the late $12.5 million rival bid. What he decides becomes the template for every company that goes under from here.
- Aug 31Before midnight on August 31: California’s floor votes. The 24 AI bills that survived the committee cull, four of them on chatbots and children, have to clear a floor vote before the session ends. We owe you the result next issue.
- 6-8 weeksThe next six to eight weeks: the Meta trial. Twenty-nine states are arguing to a jury in Oakland that Meta built Instagram to hook teenagers. A recommendation feed is AI, and this is the first jury to decide whether tuning one for attention is a harm a company owes money for.
How people are actually using AI at work.
Two studies this year read what people actually type into AI at work rather than asking them about it. Microsoft Research went through 105,000 Copilot conversations and Anthropic did the same for Claude. Both found the same thing: the biggest use is not writing, it is thinking. Just under half of workplace AI use is someone working a problem through, and the most common single request is “explain this to me.”
All bars share one scale · Source: Microsoft Research, 105,000 randomly sampled Microsoft 365 Copilot conversations, Feb 14–21 2026
1. The letter you have been putting off Small business, managers
Give it the facts as bullet points, say who is reading and what you want them to feel, and ask for three versions at different levels of firmness. Then rewrite the sentence that matters yourself.
2. The forty-page document you need three things out of Property, boards
“Summarise this” gets you a shorter document. “What in here would cost me money, what is unusual, what is missing” gets you a list. Then read those parts in the original.
3. Talking a decision through before you make it Fastest growing
Stop asking for a recommendation. Ask it to argue against what you already want, and to list what would have to be true for your plan to fail.
4. The thing everyone assumes you already understand Everybody
The acronym that has been in every email for a year, the clause in the tax letter. Explanations are the single most common thing people ask Claude for, because there is finally somewhere to ask without a colleague clocking that you did not know.
5. The recurring task that eats a morning Back office
Cleaning a messy list, the same six replies, numbers into the format the invoice needs. This is the only one where letting the machine finish is reasonable, and only when a mistake would be obvious the moment you look.
One thing not to do: do not treat silence as an all-clear. In the medical benchmark we covered in Issue #4, more than 80% of the severe errors were things the tool left out rather than got wrong. So for health, money or legal, use it to prepare better questions for a human. The best prompt is: what would you want to know that I have not told you?
these tools are very good at the first eighty per cent of a job and at helping you think, and mediocre at deciding. The disappointment comes from handing over the last twenty per cent, which is the part with your name on it.
The Microsoft figures come from an analysis of 105,000 randomly sampled Microsoft 365 Copilot conversations from global enterprise users, drawn at 15,000 per day between Feb 14 and Feb 21 2026 and classified into work-activity categories: cognitive work (analysing, problem-solving, evaluating, creative thinking) 49%, working with people 19%, producing work 17%, finding information 15%. Decision-making specifically accounts for 28% of activity.
The Anthropic Economic Index report of Jun 26 2026 covers Claude usage with a linked survey of about 9,700 respondents collected mid-May to early June 2026: explanations 17% of conversations, documents and reports 15%, guidance 11%. Both samples skew heavily toward computer, mathematical and management occupations (about 30% and 23% of Anthropic's respondents against 4% and 7% of US employment), so they describe knowledge work rather than all work. Gallup's July 2026 release puts US employees using AI at work at 52%, against 27% two years earlier, with writing and editing 51% among users.
All of these are published research rather than this week's news, and this section is guidance, not reporting. The omission finding is from the NOHARM benchmark covered in Issue #4 (ARISE, arXiv 2512.01241), where omissions accounted for more than 80% of severe errors across 1,100 clinical tasks. Nothing in this section is medical, financial or legal advice.
The Make It Useful section is built from what people actually type into these tools at work, not from what anyone recommends. So tell me what you want in it. A task you keep doing by hand? A tool you cannot get to behave? Hit reply to the email and tell me what would make your week easier. I read every one.