Three Stories and The Truth
Just returned to Geneva after a few weeks down home on the farm in Canada. Think Parliaments of children, fishing, old friends and swimming in Lake Goddamn Ontario.
Miss a day in AI and you miss a lot. Miss three weeks and the AI world completely changes. If I have to update my workshop modules one more time…
But instead of trying to recap every single update and every AI model update, we're going to simply move on with news you can use this week. In fact, I'm making things shorter as we move forward.
You get three stories and the truth. Yep, blatant thievery of the "three chords and the truth" guitar mantra, but since the U.S.A. stole my lake, I'm stealing everything else.
So: three stories I think you as a communicator can use. One practical application I review. And a subjective evaluation.
Let's get to it.
For the zillionth time, your chatbot isn't a damn fact-checker.
AI is useful for research. It's not an independent fact-checker.
An organization called Full Fact spent five months throwing real misinformation claims at ChatGPT, Grok and Gemini. The scorecard: 39 major screwups. Fake images called real, old footage passed off as breaking news, garbage sources cited as gospel. (Fair caveat: OpenAI and Google both pushed back on the methodology — some tests ran on older or API-only models — so treat this as a limits-of-verification story.)
Source: Full Fact — analysis of AI chatbot errors against live fact-checking claims. Primary source, free.
Use AI to find what needs checking, never to confirm it's true. The chain ends at a real document — a report, a filing, a transcript — not another chatbot's opinion.
AI → source → authoritative source → human judgment. Not: AI → another AI → publish.
Directly relevant. A locked-down Copilot or bespoke enterprise model doesn't fact-check any better than a consumer chatbot — it just feels more official. The verification chain still has to end outside the AI, every time.
The web is messy as hell, and we're not going to take it anymore!
The web is increasingly contaminated by AI-generated material.
Pew Research Center scraped nearly half a million webpages and found AI-written material piling up fast: more than a third of everything published since ChatGPT launched shows signs of a machine behind the keyboard. On current material the rate runs highest on .com (10%), lower on .org (under 5%), and lowest on .edu and .gov (about 1%). Fair caveat: AI detectors aren't precise on any single page — this is a population-level pattern, not proof about one article.
Source: Pew Research Center — study of AI-authored content across the web.
Stop telling your AI to "search the web and cite sources" — half the web is now AI talking to itself. Tell it exactly which sources it's allowed to trust: government data, regulators, peer-reviewed research, named institutions. In that order.
There's a feedback loop worth naming too: AI writes text, that text lands on the web, another AI retrieves it, and repeats the cycle. This doesn't prove models are getting worse, but the information they're fed from is changing fast — which is exactly the kind of mess the tool in The Truth section below is built to cut through.
Relevant even behind a firewall. An enterprise assistant that browses the open web to ground its answers inherits this exact problem — the closed system doesn't close the loop.
Gasp… some actual good news about AI?
AI finds hidden heart disease from an ordinary ECG.
Researchers at Imperial College London built an AI that reads routine ECGs — the test run a billion times a year — and catches what doctors miss. In a trial of 67,000 patients: up to 81% of heart failure cases flagged, 90% of heart-valve disease, in under two seconds. It doesn't diagnose anything. It just tells the doctor who to look at first, so they can get an echocardiogram sooner instead of waiting months. (A prior meta-analysis of 700,000+ patients backs the promise, with the standard warning: pair it with clinical judgment, always.)
Source: The Guardian, 31 August 2026 — AI system detects hidden heart disease from routine ECGs.
Keep this one in your back pocket for the next person who asks what good AI actually looks like.
Doesn't apply directly — it's a clinical tool. Included because it's the cleanest real-world model of "augment, don't automate" available this week.
The Truth about Google Pinpoint.
This is where I take a closer look at specific AI tools, trying them out and evaluating whether or not they're worth your time. Disclaimer: I'm not paid to endorse any of these and they have not been recommended to me by the company itself.
A free research tool from Google's Journalist Studio that swallows a mountain of documents — PDFs, scans, handwritten notes, up to two hours of audio or video — and turns it into one searchable, fully transcribed archive. Ask it a question, it hands you the exact page or timestamp, evidence attached. (NotebookLM does something similar but caps out at 50 sources. Pinpoint handles up to 200,000 documents per collection.)
I've worked with international orgs buried under exactly this kind of chaos — UNFCCC included, drowning in Party submissions, half of it dense as hell and some of it, let's say, unverified. (See: Lake Goddamn Ontario.) A tool that can search all of that and hand back the actual source, not a chatbot's guess at one, essential.
journaliststudio.google.com/pinpoint. Anyone with a Google account can browse public collections right away, but to build your own — upload your own documents — you have to request full access. It's free either way. Google opens that up to journalists, academic researchers, and content creators; approval isn't instant, but setup is relatively straightforward.
Google says your uploads aren't used to train its models, but the newer AI-powered question-answering features are still early access and may sample interaction data for review, names stripped first. Fine for public archives. As anyone who has ever trusted any bloody tech company trading in information-sharing, keep confidential material out until your organization has signed off (or not at all).
Extremely useful for individual communicators and organizations up to their necks in data, spreadsheets, financial data. Offers a distinct advantage to UN, Int. Org, NGO, government comms teams that have just about enough budget clearance to buy a sandwich (in pesos). I'm also a big fan of Google's source verification tools for journalists in general. Worth another article on its own.
Mañana.
Pinpoint is a separate Google product, not part of a managed Microsoft 365/Copilot environment — check with IT before uploading anything beyond public material.
We're here for you.
Have a question or concern? Interested in a workshop? Need a speech written? We have what ails you.
I've built a custom assistant that helps me pull stories daily. I then read the list, hack the useless and boring bits out, add some comments (snide and otherwise), and verify whatever makes the final copy with my human brain.
I also use AI to design this newsletter and this portion of my website, because you definitely don't want me designing anything.
AI for Leadership Communications · Geneva
Further details and all weekly newsletters available at brentkerrigan.com/what-it-means-for-us.