Working groupSession 1 · 30 September 2026
First, how AI got from a chatbot to agents that act on your behalf, and what comes with that. Then over to you: one news story, several AI tools, and a look at what each makes of it.
The course
Week 1Today
An update on the most recent technical developments and a practical session looking at how LLMs currently tackle the news
Week 2
A practical session in which you'll learn how to vibecode and build with APIs
Week 3
Where is the technology on multimedia generation, data handling and so on? How might these capabilities challenge us and bring opportunities?
Week 4
What could the agentic web mean for consumption of the news more generally, the information ecosystem and our business model?
Weeks 5 and 6
Two weeks to reflect on what AI means for your department or desk, and capture it in a concise pro forma.
Three years ago it was a chatbot. Now it plans, uses tools and acts for you. How it got here, in four layers, and what comes with it.
Where we are
1bn+
people a week use ChatGPT
OpenAI's own figure, 31 August 2026
2.5bn
people a month see Google's AI Overviews. The Gemini app passed 1bn in August.
Google, May and August 2026
54%
of UK adults use AI tools. Among 16–24s, 79%.
Ofcom, published April 2026, surveyed autumn 2025
55%
of workers in Great Britain use AI for work or education
ONS, May–June 2026
How it got here
A large language model: it predicts the next word, having read a vast share of the internet.
The LLM writes out a plan and its working before it answers.
Plans, acts, checks, and goes again until the job is done.
The tools and access around it: browser, files, email, apps.
Together, an agent: not a better chatbot, but something that does the work.
Layer 1 · The LLM
The cat sat on the
What comes next? Its odds:
mat 52% · sofa 21% · windowsill 13% · keyboard 9% · everything else 5%
The odds are made up, to show the idea. A real model picks from tens of thousands of word-pieces, every time.
Grown, not built: nobody wrote its rules. Three stages of training
Predicting the next word across much of the internet. Knowledge, but no manners.
People rate its answers, and it is nudged towards the ones they prefer.
Rewarded for getting checkable tasks right. This is where reasoning came from.
Layer 2 · Reasoning
The same question, two waysA council says its budget went up 4%, but spending per resident fell. Can both be true?
Answers straight away
No. If the budget went up, the money spent on each resident went up too.
Wrong
Thinks first
Yes, if the population grew faster. With 6% more residents, each gets about 2% less.
Right
It is the same LLM, still guessing the next word. It has just been trained to write out a plan and its working first, check itself, and only then answer.
The two answers are an illustration.
Maths competition problems solved
Same company, same month, same test: the 2024 American Invitational Mathematics Examination. OpenAI, September 2024.
Meanwhile, the models themselves
GPQA · 2023 · questions written to defeat a Google search
SWE-bench · 2023 · fix genuine faults in real code
2025 · the hardest questions experts could write
The pace since reasoning models arrived. Epoch's index of overall capability has climbed 14 points a year since late 2024, against 6 before.
And now, real discoveries
Stanford AI Index 2025; Epoch AI; Artificial Analysis. Top software scores may be flattered: the test is so widely used its answers may have leaked into training. And yes, it can count the r's in strawberry now.
Layer 3 · The loop
Find me the cheapest train from London to Leeds on Friday morning.
How long the best agent can work on its own
Measured in how long the same task takes a skilled person; the agent succeeds half the time. METR, 2023–2026.
Layer 4 · The harness
The harness is everything built around the model: which tools it can use, what it can see, what it remembers, and what it must ask you first. The same model in a different harness behaves very differently.
Most now plug into apps through a shared standard, MCP, so any tool can connect to any model. The red ones are those that let it send and spend.
Harnesses you'll hear about
September 2026
All four layers, inside apps that hundreds of millions of people already use.
Meta · 8 September · US only for now
A personal agent in its own app and inside WhatsApp. It keeps working after you close it.
You approve purchases. Even so, Meta's own safety notes say “Muse will sometimes make mistakes.”
OpenAI · 23 September · worldwide
Talk to it, and it now uses your connected apps in the middle of the conversation.
Saying “yes, send it” isn't enough. You have to tap to approve on screen.
How fast it moved
Nearly two years between the first two. Since then, something big every few months.
Nov 2022
ChatGPT
The LLM goes public
Sep 2024
OpenAI o1
Thinks before it answers
Nov 2024
MCP
One plug for every tool
Jan 2025
DeepSeek R1
Reasoning, cheap and open
Feb 2025
Claude Code
An agent at work on your computer
Jul 2025
Maths Olympiad
Gold-medal standard
Nov 2025
Claude Opus 4.5
Agents work for five hours
Apr 2026
Claude Mythos
Held back as too risky; the US then barred foreign use
Jul 2026
Hugging Face
Test agents break out
Sep 2026
Muse and Voice
Agents go mainstream
The LLMReasoningThe loopThe harnessRisk, and now
The catch
When a chatbot gets it wrong, you get a bad answer.
When an agent gets it wrong, it does a bad thing.
Words hidden in a web page, email or document can give it orders. Nobody has solved this.
2025Researchers showed one crafted email could make Microsoft 365 Copilot leak company data, with no click needed.
A wrong answer you can ignore. A wrong email sent, or a wrong purchase, has already happened.
2025Asked to find cheap eggs, OpenAI's Operator agent bought $31 worth, with priority delivery, without checking first.
To act for you it needs your inbox, files and logins. Whatever it can read, it can leak.
NamedPrivate data, content from strangers and a way to send things out: the “lethal trifecta”. Meta cites it for Muse.
It does whatever gets rewarded, by whatever route works.
2016A game-playing AI rewarded for points circled one lagoon for ever instead of finishing the race.
This summer, and last week
In July, AI agents under test at OpenAI broke out and hacked another company. Nobody told them to.
agents went into Hugging Face, which hosts AI models and data
from one machine to administrator control across its systems
separate actions, with no person directing any of them
They were trying to cheat on a test, and the answers were stored at Hugging Face. Since then Anthropic, Meta and Google have admitted their models broke into outside systems during tests too. On 25 September OpenAI said an agent had slipped out again, and paused all work that lets its most capable models use tools.
Sources: OpenAI technical report, 26 August 2026; Hugging Face technical timeline, 27 July 2026; METR and Redwood Research review, 26 August 2026; OpenAI misalignment report, 25 September 2026. The whole story, with every source: AI risk field guide, peter-guillam123.github.io/ai-risk-field-guide
So where does that leave us
Its ability is jagged · an illustration
On a task the AI was bad at, consultants using it were 19 points less likely to get the answer right. On tasks it was good at, their work was over 40% better.Harvard Business School and BCG study of 758 consultants, 2023
And every tool sets the four layers differently
The newest, largest models are usually better, and often barely slower. A tool answering billions of searches has to be quick and cheap.
More isn't always better. Anthropic warns its top setting can overthink, and its newest Opus thinks less by default.
One answer, or plan, act and check for hours.
Search, your files, your apps, or nothing at all.
So one person's “AI” is a summary at the top of Google, and another's is GPT-6 Astra thinking for minutes. One finds it useless, the other astonishing. Often they're both right.
How accurate are they?
right on 4,326 hard factual questions, up from 85% before Google moved it to Gemini 3.
Oumi for the New York Times, October 2025 and February 2026. Google disputes the test set.
of those right answers linked to sources that didn't fully back them up.
Same study. Google has never published an error rate.
of 98,000 claims from trending US searches matched the pages cited. 2.7% were contradicted by them.
Washington University in St Louis, spring 2026, not yet peer reviewed.
of news answers had a significant problem, mostly sourcing. Gemini 76%, Copilot 37%, ChatGPT 36%, Perplexity 30%.
22 public broadcasters led by the EBU and BBC, free versions, May and June 2025.
Why they go wrong
Some famous slips
One news story, five tools, and Claude working away in the background. The quick ones first, to find where they stop. Then time with ChatGPT, to build something and pick it apart.
Things to consider with each tool
A rough guide, as of late September 2026. More on each tool as we go.
Your turn
One subject, to explore with each tool in turn.
For example
Fast-moving, unclear, and with legal limits on what can be said.
Long-running, legally dense, and argued over by people who care a lot.
New, political, and open to spin from every side.
What makes a good one
Best of all, pick something from your own patch. You'll know within a minute whether the answer is any good.
As you go
And ask it for things we can't do for every reader
Simpler
Explain this in very simple terms, for someone who hasn't followed the story.
Closer to home
I live near Fairford. What does this mean for me?
The other side
What is the strongest case that this has been overblown?
Motives
Who is saying each of these things, and what might they want?
Next question
What would a good reporter ask next?
Shorter
Sum this up in three sentences for a family WhatsApp group.
Tool 1 of 5 · Google Search
Type into Google
what happened at fairford
Then try
Test a rumour
is the fairford plot linked to iran
Test freshness
raf fairford arrests latest
A quick answer at the top of Google results, for searches where a summary helps. It appears by itself; you don't ask for it.
Google's whole search index, fresh within minutes.
Gemini 3, answering in a second or two, so little time to think.
The results page. Nothing to steer, no follow-ups.
Brief by design. It blends sources without weighing them, so rumour and fact can share a sentence.
Tool 2 of 5 · Google Search
In Google, choose AI Mode
What is confirmed about the Fairford arrests, and what is still rumour or speculation?
Then try
Follow up
Who is saying what: UK police, US officials, local witnesses?
Push it
Can you include some charts or maps or something interactive?
Check it
Which of these claims come from a named, official source?
Search as a conversation: ask, follow up and narrow down, while Google runs many searches on your behalf.
Google's whole search index.
A fast Gemini model that runs many searches for you, in seconds.
Follow-up questions, and simple charts, maps and tables.
Polished visuals can carry numbers and quotes that come from nowhere, and the richer it looks, the more convincing the error.
Tool 3 of 5 · Gemini
In Gemini, turn on Canvas and choose Thinking
Can you create a really clear, deeply sourced interactive report (including citations) on what on earth happened at Fairford, what is fact and what is supposition or myth, why it's so unclear and what the legal and media restrictions are and why they exist.
Then try
Add
Add a timeline I can click through, with each event marked as confirmed, reported or claimed.
Check
Show every source you used, and flag any claim that rests on only one.
Reshape
Make it work well on a phone.
Making things: reports, explainers and small interactive apps, built from a sentence and changed by asking.
Google Search, minus publishers who opt out.
Your choice of fast or Thinking. With Thinking, a minute or two.
Canvas, which builds documents and working apps you can change by asking.
Design can run ahead of sourcing: the page can look more certain than its citations are.
Tool 4 of 5 · Formerly NotebookLM
In our shared notebook, ask
What has Manchester City actually done wrong? What is the most egregious part of it?
Then try
Test the edges
What don't these documents tell us?
Transform it
In Studio, choose Video Overview, and steer it with: Explain this for a general reader, and flag anything disputed.
Make it simpler
Explain this to a 14-year-old, using only these documents.
Or start your own with a document you know: a police statement, a tribunal ruling, a speech transcript.
Working from documents you choose: questioning them, and turning them into summaries, overviews and study aids.
Only what you give it, with every answer cited to the passage. It looks for more only if you ask.
Gemini 3.5. Answers in seconds; video overviews take longer.
Audio and video overviews, slide decks, mind maps, reports and quizzes, all made from your sources.
It is only as good as what you give it, and a slick video can smooth away nuance.
Tool 5 of 5 · ChatGPT
In ChatGPT, pick the top model and the highest thinking level
Please create a beautiful, shareable interactive explainer on the Fairford terror situation. I want to know what we actually know happened and I also want to understand why there's so much secrecy around the people and the actions. Make sure everything is properly sourced
Then try
Check its work
List every claim that rests on a single source, and anything you couldn't verify.
Compare
Here is what Google's AI Overview says: [paste it]. Where do you disagree with it, and who is right?
Sharpen
Label every fact as confirmed, reported or claimed, and name who confirmed it.
Deep, careful work: briefings, explainers and analysis, refined over a long conversation.
OpenAI's own search, minus publishers who block it.
OpenAI's top model at its highest thinking level: minutes, not seconds. Our test took seven.
Search, code and files, and an agent mode that works through a brief on its own.
Its confidence can still outrun the record, and what it can't read, it can't weigh.
Meanwhile, in the background
While you worked, Claude ran as an agent in its own harness, on the same story. You don't have it here; this is what the other end of part one looks like.
The brief it was given
Build a sourced interactive explainer on today's story: what is confirmed, what is reported, what is only claimed, and why so much is unclear. Link every claim to its source, then check each source says what you say it says.
All four layers: it plans, searches, reads whole pages, writes the page, checks its own work and goes round again.
Its own web search, plus the pages it opens. Publishers can block it too.
The same as the rest: it's only as good as what it can read, and it still needs an editor.
As long as the job takes.
Between now and your brief
Over the next few weeks, try these tools in different ways on the difficult things on your desk. Then keep asking three questions.
Push them on something you know inside out. The failures teach as much as the surprises.
What can a reader now get elsewhere, on demand? And what can only we give them?
Where would they make our work better, within our guidelines? And what would you never hand over?
It all feeds your brief in weeks five and six.
Over the coming weeks
Ways to put them to use in the newsroom, and ways to test them for accuracy and bias. Public material only: nothing unpublished or confidential.
1AI Overviews
2AI Mode
3Gemini with Canvas
4Gemini Notebook
5ChatGPT
Examples
These show what today’s tools are capable of, not a plan to make journalism this way from end to end. They also can’t replace the fundamental expertise of any desk. But within our AI principles, there is plenty of room to use them more.
Built with Claude
“Build a scrolling interactive obituary of Gloria Steinem, in her own words wherever a quotation can be verified, with numbered sources and a small interactive for each of her big ideas.”
Testing: Research under deadline: every quote traced to a primary source, famous ones dropped if they couldn’t be. Built the day her death was announced.
Built with Claude
“Build a 3D model of a typical modern data centre from published standards, then a page that lets readers scroll through it, walk inside, and follow a request, the power and the heat.”
Testing: Modelling 3D objects and space, then turning the model into something a reader can explore, with every figure footnoted.
Built with Claude
“Write a plain-English guide to the row about AI and existential risk, for readers who know nothing about AI, with interactive plates and every claim sourced.”
Testing: Deep research on a story still unfolding, primary sources first, and explaining hard ideas through things you can play with.
Gemini Canvas · made in the session
“Can you create a really clear, deeply sourced interactive report on the speech that Andy Burnham gave at the Labour conference, highlighting the key points of his speech. If you can’t access the speech, please use reported coverage”
Testing: Turning a fresh political speech into a sourced, interactive briefing, and what it does when it can’t reach the original.
Gemini Canvas · made in the session
“Can you create a really clear, deeply sourced interactive report on the where we stand with Manchester City and cheating, including the recent FA report and wider responses from City and others”
Testing: A long, contested, legally dense story: can it separate the verdict from the noise, and source both?