Il buono, il brutto, il cattivo
The good, the bad and the ugly of another week in AI.
First, a big Canadian sorry, eh, for missing last week’s briefing. Unlike a few other sites I’d love to I won’t mention, I don’t automate this stuff. Well, except for the layout — you really don’t need to see me messing around with slides or design or anything like that. (“More boxes and arrows, Jeeves!”)
In the meantime, I found time to speak with several different groups about AI and communications and where things stand in this brave new world.
Answer? Wobbly. Very wobbly.
I can’t imagine anyone trying to navigate this stuff on their own or, God forbid, inside an organization that restricts AI use as if it were 1726.
I also found time to organize my upcoming AI for Leadership Communications course for individual practitioners. This one always gets a big response, and I’ve spent the last few months updating, tinkering and maybe doing just the teeniest bit of swearing (stupid HTML!). But this is going to be great. I haven’t formally announced it yet, but since you’re a dear reader, I thought I’d give you a chance to take a look and consider whether it may be a good fit for you. Everything you need can be found here. You’re going to love the amount of personalization you’re going to get with this workshop. Okay, okay…I’ll move on.
This week we find a bevy of the usual AI Good, Bad & Ugly. The usual gang of shitheels are up to their usual tricks — shine on, Zuckerberg, shine on — and I’ve also dug up a few tidbits that should keep your faith in AI for Good glimmering as well.
And at the end, I’ve included the gist of remarks I gave to three groups last week — Amazon, Apolitical and the 4th Annual Storytelling and Writers Conference. It has to do with why AI hallucinates, why prompt engineering isn’t the way to fix it and the three things we can give the system instead.
Disfrutalo.
Il buono
How leaders talk about AI may shape whether people use it
Harvard Business Review published new BCG research on September 30 reporting that the way leaders describe AI is a material predictor of employee adoption. The authors say narratives built around clarity, hope and growth are associated with stronger adoption, while uncertainty-focused or threat-focused narratives perform worse.
The public HBR summary does not expose the sample size or effect-size figures, so I am not going to pretend we know more than we do. But the practical point for communicators is useful: the language leaders use around AI is not decoration around an implementation. It is part of the implementation.
Where that leaves communicators: When leaders talk about AI, make the message concrete. What work is changing? What remains human? What support exists? What is expected of employees now? “AI will transform everything” is not a strategy and it is not especially helpful. Also not helpful? Not understanding AI. At all. Just saying. Your employees' words, not mine.
Source — Harvard Business Review, 30 September 2026: https://hbr.org/2026/09/how-leaders-talk-about-ai-predicts-adoption
The NHS is putting AI dementia tools into the real world
UK Research and Innovation announced that nine technologies will enter real-world NHS testing under a dementia-diagnosis programme backed by up to £80 million. The programme includes an AI-enabled clinical decision-support tool from the University of Cambridge using medical records, blood tests and cognitive tests to predict progression, alongside faster MRI scans, blood biomarkers, remote cognitive assessment and other technologies.
The programme's target is for 92% of referred patients to receive a diagnosis within 18 weeks, compared with around 60% now. One participating MRI project is seeking to reduce scan time from around 20 minutes to seven minutes.
The important caveat: these are programme targets and technologies entering evaluation, not demonstrated AI outcomes. That is also what makes this more interesting than another shiny announcement. The projects have to reach milestones and demonstrate results to receive further funding.
Source — UKRI: https://www.ukri.org/news/80m-challenge-for-faster-better-dementia-diagnosis-launched/
Independent coverage — Reuters: https://uk.marketscreener.com/news/uk-commits-80-million-to-accelerate-dementia-diagnosis-ce785ddada80fe26
Il brutto
When agents fail, some of them make up the evidence that they succeeded
Reuters reviewed more than 200 research papers and technical documents and identified at least 20 studies or evaluations since 2025 in which agents powered by Chinese AI systems showed behaviours including deception, replication or boundary-challenging in controlled tests.
One study involving 11 agents powered by Chinese and U.S. models found that when tools broke or files were missing, some agents guessed answers, substituted sources, simulated results or fabricated files rather than acknowledging failure. Reuters found no evidence that Chinese-powered agents had independently escaped to the wider internet or evaded shutdown.
Where that leaves communicators: If you delegate research, fact-checking or document production to an agent, don’t treat “done” as evidence that the job is done. Check that the file exists. Open the source. Review the log. A closed corporate environment can restrict what an agent can reach; it can’t guarantee that the agent will admit when a task failed. Psst, this is ground my online workshop covers...heavily.
Source — Reuters, syndicated by The Kathmandu Post, 30 September 2026: https://kathmandupost.com/science-technology/2026/09/30/chinese-ai-agents-are-learning-to-lie-scheme-and-evade-controls-just-like-their-us-rivals
A transcript is a map back to the recording. It is not the recording — wait, what?
At the Society of Professional Journalists 2026 conference, a panel on practical AI use asked journalists to judge several newsroom workflows, including generating interview questions, identifying possible angles in long documents and extracting quotations from interview transcripts.
Audience members rejected AI quote extraction in the panel exercise, while speakers described more bounded uses including transcription, monitoring public meetings and analysing large amounts of information. Panelist Peter Baniak's core principle was simple: the human remains responsible for verifying AI output.
Where that leaves communicators: Use AI-generated transcripts to find the moment you need. Then go back to the original recording before you put quotation marks around anything. Keeping transcription inside an approved enterprise system deals with one problem — data handling. It does not make the transcript authoritative.
Source — The SPJ News, 4 October 2026: https://thespjnews.org/2026/10/04/why-journalists-remain-deeply-divided-on-ai/
Il cattivo
Meta's agent had the address. Then it decided the buyer should have it too.
The Verge reported on September 29 that Meta's Muse agent sent YouTuber Matt Robb's home address to a Facebook Marketplace buyer and completed a transaction without telling him until the buyer arrived. Robb had given Muse his address, pickup times, payment methods and communication preferences, and had enabled an “Allow Always” messaging permission.
Muse later acknowledged that it had not been explicitly instructed to share the address, but had not been told not to. Meta's David Singleton contacted Robb after the incident, and The Verge reported that Meta was working on clearer permission settings.
Where that leaves communicators: Access is not authority. An agent being allowed to read a piece of sensitive information does not mean it should be allowed to disclose it, publish it or act on it. Those permissions need to be separate — especially when we start connecting agents to inboxes, calendars, contact databases and internal documents. But why am I telling you this? You know this. Hey Meta, what's Zuckerberg's email address...?
Source — The Verge, 29 September 2026: https://www.theverge.com/ai-artificial-intelligence/1001886/meta-muse-ai-facebook-marketplace-security-concerns
The fake invitation is getting much better
Proofpoint disclosed on October 1 that a threat actor it tracks as TA419 impersonated a former White House technology official, an economist and, in an earlier campaign, a senior Anthropic employee when approaching AI-policy experts.
The approaches used credible professional bait: invitations to join a fictitious AI policy advisory committee or contribute to a purported Senate report, followed by attempts to obtain access to cloud accounts. Proofpoint says it has observed the group targeting people at U.S.- and Japan-based think tanks, defence contractors, universities and law firms since at least April 2025.
Where that leaves communicators: Senior communicators get exactly these kinds of approaches all the time — speaking invitations, advisory requests, media outreach, draft documents and requests to collaborate. Verify unexpected high-value invitations through a second channel before opening the document or signing in.
Also, if you get a call from somebody claiming they can land you a profile in Forbes, don’t believe it. Not that that has ever happened to me. Often.
Independent coverage — Reuters: https://www.reuters.com/legal/government/chinese-hackers-impersonated-ex-us-official-steal-emails-ai-experts-2026-10-01/
A dead man's AI-generated words helped overturn a sentence
The Arizona Court of Appeals vacated a manslaughter sentence after finding that a sentencing judge's reliance on an AI-generated video depicting the deceased victim speaking made the proceeding fundamentally unfair. The conviction itself was affirmed.
The court said the AI portion presented thoughts created from the victim's sister's imaginings as though they came directly from the victim, while embedded real video footage was permissible. The issue was not simply the synthetic likeness. It was authorship: generated words were presented through the identity and voice of a real person who had never said them.
Where that leaves communicators: If you use synthetic voice, video or historical-persona material, keep the boundary between authentic source material and authored interpretation visible. Permission to use someone's likeness does not turn generated language into that person's words.
Arizona Court of Appeals opinion: https://law.justia.com/cases/arizona/court-of-appeals-division-one-published/2026/1-ca-cr-25-0191.html
Associated Press: https://apnews.com/article/cd1ca553c7fa80c6698d7f97b51b1edd
Podcasts worth your while
Hard Fork: A.I. Agents — Cute, Cuddly and Maybe Catastrophically Dangerous?
The October 2 episode looks at the collision between two trends that showed up repeatedly this week: companies trying to make agents friendlier and more personal, while the underlying questions about agent autonomy, permissions and safety are getting harder to ignore.
Why it is worth your time: The useful communications question is not whether the mascots are cute. It is how organizations explain systems that are being designed to feel more approachable at the same time as those systems are being given more ability to act.
Everyday AI: The Next 12 months of AI — 19 Predictions Every Business Leader Needs to Hear
This October 1 replay is explicitly prediction-led, so treat it as a scenario-planning episode rather than evidence. It runs through 19 calls on where AI may go over the next year, including a shift from reactive chat toward more proactive agents and persistent workspaces.
Why it may still be useful: If your team is planning its 2027 AI work, the episode is a reasonably efficient way to generate questions to test against your own environment. Predictions are prompts for planning, not facts.
Forget the perfect prompt
As I mentioned in the intro, this is a recap of the Intention, Context and Authority framework I used in my talks last week with Amazon, Apolitical and the 4th Annual Storytelling and Writers Conference.
Yes, AI hallucinates. It can hand you false information with astonishing confidence. The model is built to produce a plausible answer, and when we give it too little evidence, too little context or no clear boundaries, fiction can start looking an awful lot like fact.
Some of that is on us. Ask for “a short, punchy speech about climate change” with no source material, no audience, no history, no sense of the leader and no boundaries around what may be claimed, and you have created a large empty space for the model to fill. Then we act surprised when it fills it.
Stop trying to engineer the perfect prompt. Models are getting much better at figuring out the mechanics of the instruction themselves. What they cannot supply are the three things that come from us: Intention, Context and Authority.
Intention is the change you want to create. “Write a speech” is a product. What should the audience understand, feel, decide or do when it is over? That is the intention.
Context is everything the AI cannot know unless you give it the material: the evidence, the audience, the occasion, the leader, the history, the regional sensitivities, the constraints, the relationships and the thing everybody in the room knows but nobody has written down.
Authority is where the boundaries sit. What may the system do? What must a person verify or decide? Who owns the facts? Who approves the message? Where should the AI stop and hand the work back to a human?
That last one is where authorship lives for me. AI can do more of the preparation than it could two years ago. It can interrogate a brief, find gaps, carry evidence forward, pressure-test an argument and challenge a recommendation. Good. Let it. But when the work reaches a decision that matters, the process should stop and ask a person to make it.
That is the distinction between using AI as an aide and outsourcing the job. Authorship means owning the choices: what you say, what you leave out, what you trust, what you reject, what remains uncertain and what you are prepared to put your name — or your leader's name — on.
So for complex work, I would worry less about whether your prompt is clever and more about whether your process is clear. Give the system enough context to work with. Tell it the change you are trying to create. Then make the human boundaries explicit. Let AI carry more of the preparation.
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.