Il buono, il brutto, il cattivo

On September 9, the U.S. Advanced Research Projects Agency for Health announced the teams selected for ADVOCATE, a four-year program intended to build a reliable, FDA-authorized clinical AI system for cardiovascular care. The patient-facing system is designed to support people with heart failure between visits, operate around the clock and escalate cases to a human care team when needed.

The program matters because access is part of the brief: ARPA-H says nearly half of U.S. counties have no cardiologist (SHAME!). The usual disclaimers apply: it’s still a development program, not a clinically-proven service. But it is an attempt to use AI to extend specialist care rather than just making existing care more convenient.

Source — ARPA-H, 9 September 2026: https://arpa-h.gov/news-and-events/arpa-h-launches-worlds-first-bid-build-fda-authorized-clinical-ai-cardiovascular

On September 12, Anthropic CEO Dario Amodei published an essay arguing that the AI industry must deliberately slow the pace at which it improves model capabilities, creating more space for safeguards and independent testing to catch up. His starkest forecast was that, within six to twelve months, a swarm of misaligned AI agents could be capable of taking over the internet through a persistent botnet. Sam Altman and Elon Musk publicly backed the call within hours. Thanks guys.

Here is what we know for sure: Amodei’s six-to-twelve-month warning is a forecast, not a guaranteed outcome — we hope. It describes what he believes could happen if capability growth continues without sufficient guardrails. We also know that leaders of three of the largest AI companies put the risk on the record rather than leaving it behind closed doors. Giving them the widest benefit of the doubt, it is useful that the billionaires controlling much of the technology on which our lives increasingly depend can agree on at least one thing. Giving them the slimmest benefit of the doubt: are we really trusting these people again?

Where that leaves communicators: As I wrote earlier, get your own house in order. Know what your organization’s AI governance says before someone poses an uncomfortable question in a public meeting. Know who is accountable, what teams are allowed to use AI for and which tools they may use. Get messages prepared today.

Read Dario Amodei’s essay: https://darioamodei.com/post/we-must-pace-the-frontier

Independent reporting — Associated Press, 12 September 2026: https://apnews.com/article/d59552edcb27892d8ee4d98a48397706

The New Mexico Supreme Court fined a defence lawyer $5,000 and held him in contempt after an appeal brief included fabricated police testimony and witnesses. The laywer in question said he'd uploaded a computer-generated trial transcript and other case materials to ChatGPT, expecting what he called a “bulletproof summary”. The court said the resulting filing contained false testimony from wholly invented witnesses.

Lawyers have already been sanctioned for AI-generated false citations. This case went further: fictional testimony entered a criminal appeal. The failure was not that … oh hell, you know the failure.

Where that leaves communicators: Know better than to treat a summary as a verified source. When a claim could affect someone’s reputation, liberty, health or livelihood, trace it to the original document and have a named human own the check before publication. Like we've always needed to do.

Source — Reuters, 11 September 2026: https://www.reuters.com/legal/government/chatgpt-invented-fake-police-testimony-murder-appeal-new-mexico-high-court-says-2026-09-11/

The newsroom that never existed published 8,913 articles

Anthropic’s latest threat-intelligence report documents nine cases of AI misuse, including influence operations built from fake reporters, fabricated spokespeople, forged government documents and coordinated inauthentic accounts. One commercial operation published 8,913 articles across roughly 70 fake news websites in 20 languages, supported by more than 250 fake commenter accounts. Other cases included the impersonation of a real human rights organization and ghostwritten testimony delivered at a UN Human Rights Council session. Most of the content drew little engagement, but low reach did not stop the sources from looking plausible to anyone giving it a quick once-over.

Do this: Before a site or social post enters a briefing note as evidence of real sentiment, check who owns it, how long it has existed and whether anyone outside its own network is engaging with it.

Anthropic’s full report: https://www.anthropic.com/threat-intelligence-report-september-2026

Independent coverage — Associated Press: https://apnews.com/article/00266dca90e4f8853f669648998d3bda

(Some) videos now have a paper trail

Sony and Reuters demonstrated a near-live newsroom workflow that carries C2PA provenance data from the camera through editing and publication. The system also uses a watermark as a backup if metadata is stripped. This was a demonstration at IBC 2026 in Amsterdam, not a newsroom-scale deployment.

What the hell is "provenance?" It can show where a video came from and how it changed - but it cannot prove that the scene itself is true. Still, it's a step in the right direction: build the paper trail at capture instead of trying to reconstruct authenticity the night before publication.

Do this: If your organization handles executive photos or video — event footage, B-roll or anything likely to be challenged later — ask whether provenance tracking begins at capture or gets bolted on at the end. Bonus: it also makes you sound super smart.

Source — Sony, 8 September 2026: https://www.sony.co.uk/presscentre/sony-and-reuters-demonstrate-world-first-near-live-newsroom-workflow-to-protect-content-authenticity

Consensus

I’m not paid to endorse it, etc. All the usual warnings apply here.

What it does: Consensus searches more than 220 million research papers and returns AI-assisted answers tied to identifiable studies. Its tools help compare methods, populations, sample sizes and findings.

Cost: The free tier includes unlimited basic paper searches, limited Pro messages, three Deep reviews per month and ten Study Snapshots. Pro costs US$20 per month or US$144 per year. Deep costs US$65 per month or US$540 per year. Team and Enterprise pricing is separate.

Link: https://consensus.app/

The 20-minute test

Take one evidence claim likely to appear in a speech or executive briefing — for example, “Does hybrid work reduce productivity?” or “Does employee monitoring affect trust?” Search it in Consensus. Open the strongest relevant studies rather than relying on the synthesis alone. See Il cattivo above. For the best three, record the population, method, sample size, result and limitation. Then compare that evidence with the unsupported summary a general chatbot gives you.

1. Create a free Consensus account. As always, use a separate, unconnected email account purely for testing purposes.
2. Enter one focused research question.
3. Open the strongest relevant studies rather than relying on the synthesis alone.
4. Check the methodology, population, sample size and publication date.
5. Save only claims you can trace to a named study.

How I turned Claude into a self-improving PM assistant

This is from a podcast I love: How I AI. I got a lot from this episode. Daniel Blum, product manager at Melio, walks through the Claude-and-Cowork system that manages his Notion board, processes Slack and email, and runs weekly self-improvement loops. The most useful point is not the tool stack; it is the operating model. Blum shows how persistent context, connected workflows and feedback from his own edits turn an assistant into a system that improves over time. Important for our comms teams: he also explains how Melio packaged the approach into a roughly 15-minute onboarding experience for colleagues. I wrote to him, and he sent it to me. Nice guy!

Why communicators should listen: It is a concrete example of moving beyond one-off prompts into a governed personal workflow — and a reminder that personalization, not access to a model, is often the real adoption problem.

Podcast: How I AI

Everyday AI — “Managing the AI Capability Gap: AI Is More than Ready. Most Companies Are Not”, 11 September 2026. The episode focuses on the difference between model capability and organizational use, with particular attention to workflow design, training, metrics and process redesign. For communicators, the useful lesson is to measure whether AI changes a defined work process rather than treating adoption or tool access as the outcome.

Exact episode — Everyday AI: https://www.buzzsprout.com/2175779/episodes/19731128-ep-860-managing-the-ai-capability-gap-ai-is-more-than-ready-most-companies-are-not-start-here-series-vol-19

That’s it for this issue. The machines have not ended the world.

Yet.

— Brent