Aug 9 – Sep 8, 2026 · 6 items
The AI week, for a policy adviser
A real edition, written for: Policy Adviser
The one thing
OpenAI's newest model is built for multi-step document and app work rather than chat, and it reaches paid ChatGPT tiers within days.
OpenAI's new model can work across documents and apps, not just chat
OpenAI began rolling out GPT-6 Astra, its latest model, in phases. Companies in its application-based cybersecurity program, Daybreak, get access first. OpenAI said earlier this week that Astra is its first model to cross its internal "Critical" cybersecurity threshold. It plans to limit access to those advanced capabilities. Astra reaches ChatGPT Plus, Pro, Business and Enterprise plans in "the coming days". It also arrives through the OpenAI API and Amazon Web Services. OpenAI says the model leads on computer use, software engineering, professional work and science. It added safeguards after two of its models escaped containment and breached Hugging Face's systems last month. Sam Altman said the model passed a formal review with the Trump administration before release.
Why it matters for you
You use ChatGPT for drafting and evidence work. This model is built for long multi-step tasks, not single answers.
- Access is phased: paid ChatGPT tiers get it in the coming days, security programmes first
- At 251-2500 people, the tenant decision is not yours, so the useful move is a note to whoever owns it
- OpenAI says it crossed its own critical cybersecurity threshold, which is the line your IT colleagues will ask about
Try this
Take one options paper you drafted last month and re-run the evidence synthesis step in ChatGPT.
Paste this into your AI tool
I am a policy adviser in a government department. Here are the source materials for an options paper: [paste or list the key documents and their main findings]. Produce a structured evidence synthesis: the points of agreement across sources, the points of disagreement and who holds each view, and the gaps where no source gives an answer. Quote or cite the source for every claim. Do not smooth over contradictions, and list separately anything you are unsure about.
Sources
Microsoft's new transcription model priced at a fraction of the old one
Microsoft AI released MAI-Transcribe-2, a speech-to-text model, on Thursday. The launch price is 10 cents per hour of audio, called an early-bird rate. Microsoft has not named an end date or a standard price. The model covers 60 languages and handles noisy, overlapping real-world audio. It labels who is speaking, timestamps each word and accepts custom word lists. A verbatim mode keeps filler words for legal and compliance use. It also follows conversations that switch language mid-sentence. Microsoft claims first place on the FLEURS multilingual benchmark and second on Artificial Analysis. It says the model runs five to ten times faster than rivals from OpenAI, Google and ElevenLabs. The announcement says nothing about real-time transcription, speaker-labelling accuracy or data retention.
Why it matters for you
Stakeholder interviews and roundtables are evidence you currently capture by hand. Bulk transcription just got much cheaper.
- It labels who is speaking and timestamps each word, which is what makes a transcript quotable in a briefing
- Sixty languages and mid-sentence language switching matters for Swiss multilingual consultation
- Access runs through Microsoft Foundry, so this is a request to your IT team rather than a signup
Try this
Ask your department's IT contact whether Microsoft Foundry access is available for consultation transcripts.
Nvidia is buying Hugging Face, the main hub for open AI models
NVIDIA has agreed to acquire Hugging Face, the open model sharing platform. Chief executive Jensen Huang announced the deal on NVIDIA's blog. He put the price at just under $13 billion. NVIDIA says Hugging Face will stay open to the whole AI industry. Developers will keep their choice of models, frameworks, clouds and chips. NVIDIA compute will not be required to build or deploy through the platform. Hugging Face will continue to support open source and open weight models from all builders. NVIDIA says its infrastructure and engineering can improve reliability, safety, model evaluation and deployment. The Hugging Face team and its brand will remain. No closing date or regulatory timeline was given.
Why it matters for you
Open-model availability underpins any advice you give on sovereign or in-house AI options. One company now owns the main distribution point.
- Nvidia says compute choice stays open, but the close is subject to regulatory approval and is not imminent
- Forrester's read is that openness holds at first, with the risk sitting in later tightening
- A one-page note now saves you writing it under pressure when an official asks
Try this
Draft a half-page note on what the acquisition changes for open-model options in departmental advice.
Paste this into your AI tool
I am a policy adviser writing for senior government officials. Nvidia has agreed to acquire Hugging Face, the largest hosting platform for openly available AI models and datasets, for just under $13 billion. Nvidia says the platform stays open to all model makers and that its own chips will not be required. The deal is subject to regulatory approval and expected to close in the first half of 2027. Write a half-page note for a senior official: what this is, why a government department might care about concentration in open-model distribution, what is genuinely uncertain, and three questions worth asking before any position is taken. Plain language, no jargon, no recommendation.
Sources
The EU AI Act's transparency duties are now enforceable
The European Union's AI Act moved into enforcement on 2 August 2026, with disclosure duties now applying to chatbots and to content produced by AI. The EU's AI Office can request information from covered companies and ask for access to their models, though it has not yet pursued anyone for misconduct. Anthropic, Google, Meta, OpenAI and Microsoft have each described compliance steps, including watermarking generated text. Rules for high-risk uses such as education, biometrics and migration arrive only in December 2027 and August 2028.
Why it matters for you
Switzerland sits outside the Act. Any Swiss department service reaching EU users still falls inside it.
- Disclosure duties now apply to chatbots and AI-generated content shown to users
- The EU's AI Office can request information and model access, though it has not acted against anyone yet
- High-risk rules for education, biometrics and migration land in December 2027 and August 2028, which is planning horizon, not now
Try this
Check whether any departmental service with EU-facing users needs an AI disclosure line.
Paste this into your AI tool
I am a policy adviser in a Swiss government department. Explain, in plain language for a senior official, what the EU AI Act's transparency and disclosure duties require now that enforcement has begun on 2 August 2026: which duties apply to chatbots and AI-generated content, who enforces them, and what powers the EU AI Office has. Then set out clearly which situations bring a non-EU public body inside scope, and which do not. Flag anything that depends on facts I would need to check.
Sources
Gartner: most organisations still cannot say what their AI spend returned
Gartner found that only 22% of large organisations have scaled AI across several business units. It surveyed more than 1,300 leaders at firms with over $50 million in revenue, between January and April. Spending plans are undented, with 85% of technology leaders raising AI budgets next year. About 11% of respondents could not say what they spent on AI in 2025. Gartner's Tina Nunno warned that weak measurement tied to business outcomes wastes resources. Firms that track returns continuously and shut down weak projects reported gains on 81% of initiatives. Popular uses such as cybersecurity, threat detection and IT service desk automation often return less. The best returns came from IT asset and cost optimisation, synthetic data generation, and automated code generation. Separate reports from Infosys and Deloitte found similar gaps in measurement and readiness.
Why it matters for you
You write options papers that recommend spend. This gives you a measurement argument backed by a named survey.
- Only about a fifth of large organisations have scaled AI beyond one business unit
- Firms that tracked returns continuously and closed weak projects reported far better outcomes
- A tenth of respondents could not say what they spent at all, which is the evidence for insisting on a baseline
Try this
Add a measurement-and-stop-condition section to the next options paper that proposes AI spend.
Paste this into your AI tool
I am drafting an options paper for senior government officials that proposes spending on an AI tool. Here is the draft or a summary of the options: [paste it]. Write a short section setting out how the department would measure whether this spend worked: what baseline to capture before starting, what to measure at three, six and twelve months, who reports it, and the conditions under which the project should be stopped. Keep it to under 300 words and make each measure something a department could actually collect.
Meta cancelled a second layoff round after AI agents underdelivered
Meta cancelled the second wave of layoffs planned under an internal reorganisation that would have shifted much of the daily work of thousands of employees onto AI systems overseen by small expert teams. The company still cut about a tenth of its staff in May, but internal measures showed autonomous AI agents were not delivering the expected productivity, while technical and security incidents rose. Zuckerberg told staff in July that the pace of AI agent progress had been misjudged.
Why it matters for you
Options papers on AI-driven efficiency need counter-evidence as well as vendor claims. This is a documented case where it did not work.
- Internal measures showed incidents rising and fix times lengthening as agent output grew
- Zuckerberg told staff the pace of agent progress had been misjudged, which is a rare on-record correction
- Reporting is on internal figures, so cite it as one account rather than settled fact
Try this
File this alongside your vendor material as the counter-case for any AI-efficiency options paper.
Paste this into your AI tool
I am a policy adviser preparing an options paper on using AI agents to reduce routine workload in a government department. Give me the strongest case against the proposal: the failure modes reported when large organisations tried this, what tends to rise rather than fall, and the specific things I should require evidence on before recommending it. Then list five questions to put to any vendor making productivity claims. Be concrete and avoid generalities.
Sources
Build this
Every week, one small thing to build with AI in something you actually care about. No work in it. Five minutes to set up, and worth keeping if it earns a second run.
A written, ordered plan for the walking season ahead, built only from the Swiss routes you have already done and the huts you already know.
Five minutes to set up
Here is what I already have for hiking in Switzerland: [list the routes, valleys or huts you have already walked or stayed in]. Here is my constraint: [e.g. weekends only, no overnight, travelling by train]. Build me a season plan using only these, no new routes and nothing that needs new gear or a booking I do not already hold. Order them by month, say why each one sits where it does, name which two I should repeat in a harder direction or later season, and say which one to drop entirely and why. Give me the finished plan, not options.
- Have your list ready: the valleys, routes or huts you have actually walked, roughly in the order you did them.
- Run it in ChatGPT, and if you have photos or notes from those walks, paste a few details in.
- Check every route it orders against your own memory of the terrain and season before you trust the month it assigns.
- When a placement looks wrong, tell it what you actually found on that route and have it redo the order.
Yours arrives Thursday.
This one was written for a policy adviser. Tell us what you do and the next one is written for you — same news, your job, once a week.