The AI Roundup | May 2026 | Part 2 | Your regular look at what's happening in AI
Good afternoon. The sun is out, the forecast says it is staying out for the week, and writing this in a warm office with the windows open is reminding me how much of the year we spend not doing exactly that. I hope wherever you are reading this, you have managed to get five minutes of it on your face.
It has been a busy week. Two new pieces went up on the Techosaurus website, the new chapter of Prompt Fiction dropped just this morning, and on Tuesday morning I delivered my first talk as a Fellow of the British Computer Society at Bath Digital Festival. Plenty to tell you about, so let’s get into it.
If you find this useful, the single best thing you can do for me is forward it to one person you think would appreciate it. They can join the list at techosaurus.co.uk/newsletter/subscribe.
Right.
🔥 The Big Stories
What if every department used AI deliberately?
On Tuesday morning I gave the opening talk of Bath Digital Festival 2026 at the Bath Royal Literary and Scientific Institution. Grade I-listed Georgian building, painted ceilings, a suit of armour in a glass case to one side, and a packed upstairs room that included a fair number of AI consultants who had come to see what someone else was doing with this. Hosted by the brilliant Joyann Boyce of InClued.AI, who set the tone perfectly.
The advertised title was “What if you used AI for everything?”, which the slide promptly shattered like glass to reveal the talk I actually wanted to give: What if AI supported the repetitive, administrative, connective work around everything? Much less sexy. Would not have got people in the room. That was the point.
Four live demos followed: an Excel workbook of about a thousand rows of training feedback that Copilot analysed, classified and turned into a live dashboard while I was on stage doing other things; a one-shot prompt that turned a chat into a plain-English narrative, an SOP and a Mermaid process diagram; a hiring workflow that anonymised CVs, scored them against criteria the AI helped design, and rendered the whole assessment as a clickable MarkMap; and a customer service triage bot built on top of a product catalogue and an FAQ. Two background agents ran the whole time. None of it was magic. All of it was deliberate.
There was also a slide that just said WARNING in big yellow letters, and four house rules that were on the screen behind me throughout. We do not let AI make decisions for us. We do not publish without human review. We do not hand over relationships where trust and judgement matter. We have built things that we deliberately have not switched on. That last one is the one most people forget exists.
The questions afterwards were sharper than the talk, which is always how it should be. I have written the whole thing up, including the demos, the four rules, the ROAR framework, and the answers to the best questions from the room, as What If Every Department Used AI Deliberately? on the Techosaurus blog.
Personally, doing this as my first talk as FBCS, representing the British Computer Society, in a room full of techies and AI consultants, on the opening morning of the festival, was a milestone I will not forget. Thank you to TechSPARK, to Joyann, to the BRLSI, and to everyone who turned up.
Buying an AI licence isn’t an AI strategy
A lot of businesses are at a strange point with AI. They know it matters. They have bought a few licences. Someone has written an acceptable use policy. There has been a one-hour awareness session. A handful of confident people are racing ahead while everyone else hovers at the edge wondering what they are supposed to do with it. Then the leadership team looks at the monthly licence cost and asks a very fair question: why hasn’t this changed anything yet?
That is the bit we need to talk about. Because AI adoption behaves much more like a modern way of working than a normal software purchase. Access is not adoption. Giving people access to AI and calling it done is like handing someone the keys to a car and assuming they can drive.
There is now a legal and governance reason to take this seriously too. Article 4 of the EU AI Act requires providers and deployers of AI systems to take measures, to their best extent, to ensure a sufficient level of AI literacy among staff and others dealing with AI systems on their behalf. Chapter I and II of the Act, including Article 4, have applied since 2 February 2025, with general application from 2 August 2026. Even if your business is not directly caught, the direction of travel is obvious. “We gave everyone access and hoped for the best” is not a serious adoption plan.
The longer piece I published this week walks through the difference between AI and automation, why prompting is really delegation, the “bloke I don’t know” test, our crawl, walk, run approach, and why training has to be built around real work, not generic “look what ChatGPT can do” sessions. It also tells you, plainly, what to expect from our Practical AI for Everyday and Business Tasks online course and our Open Learning in-person days, so you can decide what fits your team.
The thing your AI does with your conversations
If you read one news section in this edition, make it this one. Google updated the Gemini Apps privacy terms in May, and the wording is worth knowing word for word, because it is now the default position for a tool a lot of you are using every day. A subset of your saved chats is reviewed by human reviewers, including Google’s trained service providers, to assess whether Gemini’s responses were low quality, inaccurate or harmful. Suggested better responses are then used to train Google’s models. Reviewed chats are retained for up to three years and are not deleted when you delete your activity.
The good news is you can turn most of it off. Open myactivity.google.com/product/gemini and switch Gemini Apps Activity off. Even then, Google retains chats for 72 hours to respond to you and to keep Gemini safe. The other piece of good news is that your Workspace content from Gmail and Drive is not reviewed or used to improve the Gemini app. That distinction matters. The consumer surface and the business surface are not the same product.
Claude and ChatGPT have their own version of this story, and it is the bit nobody talks about. Thumbs up and thumbs down feedback is not just a vote. When you tap it, the entire related conversation goes into the back-end for review and potential training, even if you have opted out elsewhere. On Anthropic’s published policy, feedback-linked conversations are retained for up to five years on consumer plans. On Team and Enterprise plans, an owner can disable feedback submission organisation-wide. On ChatGPT, OpenAI is clear that feedback can include the conversation around the rating, even if you have switched off training in your settings.
The takeaway is the same takeaway I keep ending up at. Use business tools for business things. Audit your settings on every AI you log into. Be deliberate with that thumbs button. And if you are a Team or Enterprise owner on Claude, go and look at the Rate Chats toggle in your Data and Privacy settings, because the default may not be what you assume. I am writing the whole thing up as a longer piece this week, including a step-by-step audit of where to click on each platform.
Sources: Gemini Apps Privacy Hub, Anthropic Privacy Center, Claude Help Center on Rate Chats, OpenAI Help Center.
📰 Other News
Google I/O happened on Tuesday, so the headline action this week is Google’s, with a healthy run of other shipping news around it.
Google I/O 2026: Spark, 3.5 Flash, and Omni
Three things from Google I/O on 19 May are worth attention. Gemini Spark is Google’s 24/7 agentic assistant, rolling out to trusted testers and Google AI Ultra subscribers in the US, with Gmail, Docs, Sheets and Slides context baked in and third-party tool support over MCP coming this summer. Gemini 3.5 Flash is the new default model across the Gemini app and AI Mode worldwide, and Google is briefing on enterprise cost savings of over a billion dollars a year for customers moving to it. Gemini Omni Flash is the first model in a new natively multimodal family that takes any combination of images, audio, video and text and produces video grounded in Gemini’s real-world knowledge. Veo is not being retired; Omni is positioned as the next-generation creation surface alongside it, with Omni Pro still to come. If you have not yet logged into the Gemini app this week, you are likely already using 3.5 Flash without realising. (Sources: Google I/O 2026 keynote, TechCrunch, 19 May 2026.)
OpenAI signs up to SynthID and C2PA
Announced alongside I/O, OpenAI has joined the C2PA content provenance standard and is embedding Google DeepMind’s SynthID invisible watermark into images generated by ChatGPT, Codex and the OpenAI API. SynthID has already tagged over 100 billion images and videos and 60,000 years of audio, and is now rolling into Google Search, Chrome and Circle to Search on Android. Kakao and ElevenLabs have signed up too. The wider point is that the era of “AI-generated or not?” is starting to be answered at the pixel level, not by your gut. Worth keeping an eye on as the rollout expands to existing images and the verification tools mature. (Source: The Next Web, 20 May 2026.)
ChatGPT lands inside PowerPoint
OpenAI shipped a ChatGPT add-in for Microsoft PowerPoint on 21 May, alongside the existing Excel and Sheets integration. The beta is global and available across every tier including Free. The sidebar takes notes, documents, spreadsheets, images and existing decks as source material, generates structured slides that stay editable, and will interrogate your draft for story gaps and logic breaks before you walk into the room. Together with Claude for Word, Excel, PowerPoint and Outlook, that means most paid AI subscribers can now drive Office without having to pay an extra £30 a month for Microsoft 365 Copilot. Microsoft will not love that, but it is good news for the rest of us. (Sources: Engadget, 21 May 2026, ChatGPT for PowerPoint.)
Microsoft makes GPT and Claude mark each other’s homework
Microsoft turned on a feature called Critique inside the Microsoft 365 Copilot Researcher agent. GPT drafts the response, Claude reviews it for source reliability, completeness and citation accuracy, and only then does the answer reach the user. A sibling feature called Council runs multiple models in parallel for harder questions. Microsoft is reporting a 13.8% improvement over the previous top performer on its internal research benchmark. The substance, if you strip away the marketing, is that serious enterprise AI is moving away from single-model loyalty and towards multi-model orchestration. Users may stop caring which model answered, and that is probably the right outcome. (Source: Microsoft Community Hub.)
Edge Copilot reads across every tab you have open
Microsoft started rolling out a major Edge Copilot update on 13 May. With permission, the Copilot sidebar (Win+Shift+C on Windows) can now ingest the content of every open tab simultaneously and treat them as a single working surface. Compare prices across five retailer tabs. Pull a spec from an article you opened an hour ago. Summarise a stack of research without switching between anything. Available across Windows, Mac and the Edge mobile app. This is the same direction of travel as ChatGPT’s Atlas and Perplexity’s Comet. The browser is quietly becoming the AI workbench. (Source: Eastern Herald, 14 May 2026.)
Copilot keyboard shortcuts arrive in Office in June
A small one with big implications. From June 2026, a unified set of Copilot keyboard shortcuts is rolling to general availability across Word, Excel and PowerPoint. Alt + C on Windows and the web, Cmd + Control + I on Mac, focus the Copilot button in the canvas, or jump you into the chat edit box if the pane is already open. F6 moves focus to the Copilot button across platforms. Outlook and Word for Windows and Mac have the simplified shortcuts today. Keyboard shortcuts are boring until they change behaviour. This is one of those. (Source: Microsoft 365 Insider Blog.)
Microsoft cancels internal Claude Code licences
Microsoft is cancelling most internal Claude Code licences in its Experiences and Devices division by 30 June 2026, the end of its financial year, and is steering thousands of engineers working on Windows, Microsoft 365, Teams, Outlook and Surface towards GitHub Copilot CLI instead. Claude models themselves remain available through Copilot, including the consumer and 365 surfaces. The story under the story is that Claude Code became massively popular inside Microsoft after access opened up in December, and a homegrown alternative is now the politically and financially obvious move. The model wars are starting to look less important than who owns the workflow. (Source: Windows Central.)
Mythos, Daybreak, and the cyber AI arms race
OpenAI launched Daybreak on 11 May, its dedicated cybersecurity programme built on GPT-5.5, GPT-5.5 with Trusted Access for Cyber, and a more permissive GPT-5.5-Cyber model for authorised red teaming. Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks and Zscaler are early integrators. It sits alongside Anthropic’s Mythos, which Anthropic says has found vulnerabilities across every major operating system and web browser. The genuinely interesting part is that several respected cybersecurity experts now argue these models are not, on the whole, finding bugs that older tools and skilled humans could not have found. What they are doing is making the rest of us pay attention. I have written a longer piece on this if you are interested. (Sources: OpenAI Daybreak, Scientific American on Mythos.)
Grok arrives in OpenClaw
xAI added native OAuth login for Grok inside OpenClaw on 19 May, with SuperGrok and X Premium subscribers now able to authenticate without setting up an XAI_API_KEY. The integration brings chat, image and video generation, and live X post search into OpenClaw agents, with WhatsApp, Telegram, Slack and Discord wiring up from 22 May. The headline is not Grok itself, it is the direction. The agent platforms are becoming model-agnostic, and where Claude and ChatGPT used to define your environment, OpenClaw, Cowork and Codex are starting to. Choose your harness, not just your model. (Source: xAI, 19 May 2026.)
OpenAI and 1Password tackle the credentials problem
A practical one for any team using AI coding agents. 1Password shipped an Environments MCP Server for OpenAI’s Codex on 20 May. Codex can request credentials just in time from your 1Password vault, scoped to the specific task, with the actual values kept out of the model’s context window, out of repositories, and out of your local files. This addresses the single biggest source of accidental leaks in AI-assisted development, which is people pasting API keys and database passwords into prompts. It is available to joint OpenAI and 1Password customers today. (Source: 1Password, 20 May 2026.)
Musk’s lawsuit against OpenAI is thrown out
A federal jury in California took less than two hours on 18 May to throw out Elon Musk’s lawsuit against OpenAI and Sam Altman, agreeing unanimously that Musk had waited too long to file. Musk’s team had sought to remove Altman from the company over the move away from OpenAI’s original non-profit mission. The case is now expected to head to the 9th Circuit on appeal. The bigger picture is that the legal storms around OpenAI’s structure and origins are starting, slowly, to resolve. The model and product story does not need a permanent courtroom backdrop to make sense, and this week is a step towards that. (Source: NPR, 18 May 2026.)
Farewell, Jeeves
Ask.com, formerly Ask Jeeves, quietly shut down on 1 May 2026 after 29 years. A small message went up on the site thanking the engineers, designers and users who had built the platform, and signed off with “Jeeves' spirit endures.” It is hard not to read the closure as a footnote to a much bigger story. Ask Jeeves, in 1996, was the first search engine to invite you to phrase your question in plain English. Three decades on, that idea finally found the technology that matched it, and the original is now off doing whatever search engines do in retirement. (Source: TechCrunch, 2 May 2026.)
💬 Scott’s Soapbox
Everyone’s watching Claude. Meanwhile, ChatGPT quietly built half the productivity stack.
If you only read the AI press for the last six weeks, you would think there were two companies in this industry. Anthropic and everyone else. Claude Code dropped from Pro and came back inside a day. Anthropic outages. The QuitGPT movement breaking 4 million pledges. Anthropic leasing compute from Elon Musk’s Memphis data centre. A reported 80x year-on-year growth at a $30 billion run rate. The whole spotlight is on one company.
And while everyone is watching Claude, OpenAI has been shipping in volume. GPT-5.5 is the foundation under everything, and it is genuinely strong. GPT-5.5-Cyber and Daybreak launched into security. Codex on mobile, on Mac, with Chronicle screen-reading memory, and now wired into 1Password for secrets. ChatGPT for Spreadsheets in Excel and Google Sheets. ChatGPT for PowerPoint this week. ChatGPT for CarPlay. GPT Image 2 as the new default. JotForm and dozens more apps inside the chat. SynthID watermarking across every image they produce. And quietly, the user-experience changes that actually matter, like large pastes now turning into attachments instead of clogging the composer.
When I gave a live demo at Bath Digital Festival this week of Copilot rolling through about a thousand rows of training feedback and building a working dashboard, that was GPT-5.5 underneath. When I pushed it harder, it carried on. Claude is doing the same in Word and Outlook, which is great. But the conversation has narrowed in a way that is starting to do people a disservice. Two days ago I watched a senior leader tell a room that “AI is having a wobble”. It is not. One vendor is having a wobble. The whole industry is shipping at pace.
There is also a money story under this. Sora was reportedly costing OpenAI a million dollars a day before they shut it down. They closed the product, took the loss, and the equivalent investment now sits inside the productivity suite. That is unusual discipline for a company in a hype cycle, and worth noting.
The honest position right now is that ChatGPT, Claude, Copilot, Gemini and Perplexity all do enough of what most of us need that the better question is not “which one wins?” but “which one am I best paired with for this task?”. Be model-agnostic. Watch all the surfaces, not just the loudest one. If the only AI story in your feed this month was a Claude story, your feed is missing things you would benefit from knowing.
🎲 Wildcard
A real Monet, labelled “AI”, and a thousand confident art critics
An X user posted a cropped image of a painting on 14 May with the platform’s “Made with AI” label attached, asking the internet for its honest critique. Within hours the post had 2.3 million views, 819 comments and over a thousand reposts. The crowd duly went in. Composition was attacked. Colour choices were called out. Texture was savaged. One commenter wrote an 850-word breakdown of the painting’s “soulless AI shortcomings”.
The painting was, of course, a genuine Claude Monet. Water Lilies, painted around 1915, currently hanging in the Neue Pinakothek in Munich. None of the critics had bothered to look long enough to recognise one of the most famous canvases in Western art.
It is a wonderful piece of accidental performance art, and it tells us something uncomfortable. We have moved on from “can AI make art?” because we have settled that one. The new question is whether people can still see art when they have already decided how they feel about it. The label is doing the looking. I have written about this properly as a longer piece, because the lesson generalises a long way past art. We are all judging by the label more than we would like to admit.
(Source: PetaPixel, 14 May 2026.)
💡 Try This
This week’s challenge: write a proper handoff prompt before your chat compacts
If you have ever been deep in a long Claude or ChatGPT chat and watched it grind, paraphrase your earlier messages, or quietly tell you it is “compacting”, you have hit the context window. The model is binning detail to keep the conversation alive. Carry on past that point and you are working with an AI that no longer fully remembers the brief.
The fix is to do a clean handoff to a fresh chat. Try this exact prompt. Copy it. Use it. Adjust it for the work you are doing.
This chat is running out of room. Build me a handover prompt for a fresh chat that covers: (1) state of play: what’s done, what’s signed off, what’s outstanding or unresolved; (2) files needed: list every file the new chat won’t have, and tell me which I should already have versus which you need to generate now so I can download; (3) key learnings: decisions, gotchas, things to avoid; (4) working state: variables, constants, naming patterns, conventions; (5) the next concrete step. Write it for the AI to read, not me. Flag anything you’re unsure about.
Why each bit matters. State of play tells the new chat where the work has got to. Files needed forces the model to be honest about what it cannot bring with it. Key learnings stops the new chat repeating mistakes you already moved past. Working state means your variable names, formats and conventions survive the move. The next concrete step means the new chat starts working, not deliberating.
Open a fresh chat. Paste the handover. Attach the files it flagged. Carry on. You will be amazed how much smoother long projects get when you treat the context window as a real constraint rather than something to fight.
🎧 Want to Go Deeper?
I co-host a podcast called Prompt Fiction with Reece Preston, where we go long on stories like these. Chapter 14 dropped this week, and it is a lightning-round special. Six weeks of AI updates, vendor by vendor, in one shorter-than-usual episode. ChatGPT in Excel and the live dashboard demo, the Claude Code Pro panic and reversal, Mythos and Daybreak, Gemini personal intelligence, Chrome’s 4GB AI download, Perplexity’s Personal Computer, the goblin mystery, and the cost-of-progress conversation we want to come back to in a future chapter. Reece had a couple of weeks out, so it is also our first episode in six weeks. Good to be back behind the microphone.
📆 Come and See Us
A short note on the diary before the dates.
28 May: Yeovil Digital Hub (cancelled)
Those of you who have been to a Yeovil Digital Hub event before will know that these evenings are built around community. We love bringing people together, welcoming new faces, and creating a space where everyone feels part of the Hub.
Sadly, a member of the small team working behind the scenes to make these events happen has experienced a family bereavement. Community matters deeply to us, and family comes first, so we’ve made the difficult decision to cancel this month’s event.
All tickets that have been issued will be refunded. We’re sorry we won’t get to see you this time, but we know you’ll understand and support the decision. Thank you, as always, for being part of this community. Shane Evans, our keynote for the evening, will be back at a future Hub once we have the date settled.
We look forward to seeing you all again in July.
30 June: Digital Somerset, The AI Challenge
Think you’ve got the AI bug? Reckon you can prompt your way out of a paper bag? Here’s your chance to find out. The next Digital Somerset event is throwing down the gauntlet. We’re setting a creative challenge on the night, and anyone who fancies a go will have a fixed window to build, make or solve something using AI. Bring your own laptop, bring your favourite AI tool, bring your wits.
You don’t have to compete to come along. If you’d rather sit back, learn from the room and chat to people doing genuinely interesting things with AI, you are as welcome as the people sweating over their keyboards. This one is for anyone curious about AI, whatever your level. Total beginners through to seasoned prompt engineers. The only ask is that you bring a tool you’re comfortable using, whether that’s ChatGPT, Claude, Gemini, Copilot, Perplexity, or something more niche.
Expect creativity, problem-solving, a bit of healthy competition, and the kind of conversations you only get when a roomful of people are working on the same brief at the same time.
Tuesday 30 June 2026, 5:30pm to 8:00pm, Taunton Conference Centre, Wellington Road, Taunton, TA1 5AX. Use code EarlyBird20 for the early-bird discount.
7 July: Yeovil Digital Hub with Jono Buchanan on AI in Music
The next Yeovil Digital Hub is on Tuesday 7 July, and the keynote is from Jono Buchanan, on AI in music. Jono is a composer, producer and music technology educator based just outside Bath, with a career spanning television scoring, production music and high-profile remix work. He also lectures in Electronic and Produced Music at the Guildhall School of Music and Drama, and creates widely followed Logic Pro tutorials and online courses. Jono now spends much of his time exploring what it means “to be creative” as technology offers tools that are moving at pace and challenging a long-held understanding of what creativity means for composers, producers and audiences. The ideal guide for a conversation on AI in music.
Footer
Thank you for reading. If this was useful, please forward it to one person who would benefit, and ask them to subscribe at techosaurus.co.uk/newsletter/subscribe.
If you would like Build Smarter, Open Learning, or any of the work above brought into your team, get in touch through the Techosaurus website.
Until next time,
Scott Quilter FBCS Co-Founder & Chief AI & Innovation Officer, Techosaurus LTD