The effort worth keeping

Friends,

your weekly AI briefing is here - designed to help you respond to AI, not react to the noise. No curveballs. No chaos. Just clarity.

AI convenience arrives as the default setting. This week asks where a leader wants people to think first and where AI can go first. The lens is a Guardian opinion piece published on 31 October by Carly Dober, a Melbourne psychologist. She argues that some of the effort AI removes helps keep our thinking sharp and our curiosity alive, and that we should add deliberate friction back in. It is an opinion piece, and this edition sets it beside two small data points from this week.

📰 This was the week that was...

This was the week to decide where the checking happens.

On 3 November the Guardian reported that a senior Crown Prosecution Service lawyer cited a legal precedent that does not exist in a High Court case. The Guardian understands the reference came from an AI summary produced during an internet search. The High Court judgment, handed down on 29 October, records that the case does not exist, calls the explanation of how it came to be relied on unsatisfactory, and says it made no difference to the outcome of the challenge before the court. The CPS told the Guardian that it has a policy requiring human checks for accuracy, that it carried out a review to see whether other cases were affected, and that it has shared the lesson that case law must be cited from trusted legal sources. That plain lesson is the useful part, and it comes back in the Strategic Insight.

On 2 November, OpenAI launched its Grid Protectors Program, which aims to design future computing campuses so they can reduce their electricity use when the local grid is under strain. For its planned campus in Effingham County, Georgia, Georgia Power says OpenAI is expected to need about 3,200 megawatts and has agreed to provide up to 1,000 megawatts of flexible demand response. That is the part of the load that can be reduced at times of high demand, and OpenAI says it equals the average demand of roughly 700,000 households. Georgia Power projects about $950 million a year in customer savings from 2029 across its large-load contracts, and OpenAI says its agreement contributes to that plan. The campus is still in development, and OpenAI says arrangements will differ by location. It is encouraging news for anyone watching what AI means for local power bills, and the proof will come once the campus is running.

On 4 November, Reuters reported that AI shopping agents, which research and buy products for a shopper, could cut into the advertising businesses of Amazon and Walmart. The two are taking different routes, and the story features an outside shopping agent that Walmart is partnering with. Companies including Procter & Gamble, Nestlé and L'Oréal say they are working to make product information easier for AI agents to read. If more customers ask an agent to shop for them, clear and accurate product information could become part of your shop window.

Let's get into it.

🔥 Urgent Priorities

✅ One thing to check this week: a tampered copy of an AI add-on.

On 29 October, Cursor reported that two tampered versions of an extension branded as Anthropic's Claude Code (2.1.312 and 2.1.313, for Windows and macOS) were available in its marketplace between about 2:27am and 1:35pm Pacific time that day. Anthropic did not make them. On 30 October, Microsoft published its analysis, which says the code attempts to add an attacker's SSH key (a digital key that allows remote login), set up a reverse connection to the attacker's server and keep itself running. Cursor says remote login would fail on most Windows and macOS machines, because it needs administrator rights the extension usually lacks. The attacker's key and start-up files may still have been written, so Cursor asks affected users to run its check either way. Cursor has found no evidence that other versions or other extensions are affected, and says the attacker's server has been taken offline. Microsoft confirms the tampered packages were not distributed through Visual Studio Marketplace or npm.

✅ Worth doing now: ask your IT lead who can install AI add-ons in your organisation, and from where. Ask too whether anyone installed or updated that extension from Cursor's marketplace on 29 October, between about 2:27am and 1:35pm Pacific time. Cursor's notice has a short check and clean-up scripts for each system, and it says uninstalling the extension alone may not be enough.

This deserves attention this week, and a calm half hour is plenty.

🎯 Strategic Insight

Tension: AI convenience arrives as the default setting. A summary appears before anyone reads the source, a first draft appears before anyone forms a view, and the effort that used to come with those steps leaves the day without a decision being made. Some of that effort is unlisted in any job description, so a team can lose it without noticing.

Optimistic insight: Leaders can choose where the effort stays. Carly Dober argues in her Guardian opinion piece that the friction AI removes helps drive curiosity and problem-solving, and that deliberate friction is worth building back in. Her suggestions are to wait about 30 minutes before asking AI, to write the first draft of any email, letter or presentation yourself, to ask a friend or colleague or check a reputable website instead of asking AI for a recommendation or explanation, and to set aside screen-free hours. Each of her suggestions names a moment where a person thinks first, and a leader can choose those moments this week. AI can go first everywhere else, at full speed.

What's shifting: Two relevant data points arrive this week, both small.

The first is the court case above. The judgment records that the cited case does not exist and says this made no difference to the outcome of that challenge. The Guardian understands the reference came from an AI summary. The CPS says it has a policy of human checks for accuracy and has shared the lesson that case law must be cited from trusted legal sources. That lesson names the source to check against, which any team can copy.

The second is a preprint submitted on 3 November on "Agentic Path Maps", which show a user the other routes an AI agent could have taken. In a study of 48 people on data-analysis tasks, the authors report that participants rarely considered alternatives on their own and over-relied less on the agent's answer when alternatives were shown. It is small, early and author-reported. It measures overreliance on AI, a close relative of the automation complacency Dober describes. Our reading: both data points suggest that a named source or a visible alternative gives people something concrete to check against.

Why this matters now: Where a team already uses AI daily, its defaults get set by habit unless someone reviews them. Dober's suggestions cost little to try, and the CPS lesson shows what a named check looks like. Agreeing the first-attempt and checking points in one short conversation costs far less than finding the gaps after a mistake.

Action: Hold a 20-minute team conversation this week and agree three things in writing.

1. Name two or three tasks where people make a first attempt before AI is used, such as a first outline, a first view on a decision or a first read of a customer's problem.

2. Make "check the source" a standing step for anything that goes outside the organisation. The person sending it confirms the key facts, figures and references against the original, and says so.

3. Protect some unstructured thinking time, as Dober suggests: a wait of about 30 minutes before reaching for AI, or a screen-free hour. Treat it as a working-pattern experiment and review it in a month for what it produced.

🤓 Geek Out

1️⃣ An AI model that estimates survival for children with brain tumours

In Nature Genetics, researchers report MANTIS, an AI model that reads the DNA methylation pattern of a tumour (a chemical signature in its DNA) to estimate an individual child's chances of survival over ten years, for two childhood brain tumours: medulloblastoma and ependymoma. It was built from data on 2,786 patients and tested on independent clinical-trial data from past patients. For average-risk medulloblastoma, the model identified about 60% of patients who showed no significant survival disadvantage on lower-dose radiation in that look back at past data. The paper associates lower doses with improved neurocognitive outcomes, and says prospective trials are needed to establish clinical utility and demonstrate patient benefit.

Why it matters: The order of evidence is a good one to copy: past data first, then prospective trials, then daily use. Where stakes are high, the checking is part of the design.

👉 Action: When a supplier shows you an AI tool's accuracy, ask whether it was tested on past data or in live use, and what testing comes next.

👉 Read the paper

2️⃣ A map for deciding where nature protection could start

Researchers at the University of Illinois Urbana-Champaign and colleagues have published, in npj Biodiversity, a mapping approach for deciding where conservation effort should go first, which they call AI-driven. They combined 826,989 sighting records with environmental measures such as climate and soil, then refined the maps in repeated rounds to produce phased priorities for Illinois. Protected areas cover about 4.42% of the state, and the maps suggest some top priorities lie elsewhere, along the Mississippi River corridor and in the Shawnee Hills. Checked against 542 on-the-ground survey sites, the concordance for plants was modest, so the maps are a starting point for field validation.

Why it matters: If the ranking holds up in the field, it could help a limited budget go further. Until then, treat it as a shortlist to check on the ground.

👉 Action: If you manage land or back nature projects, look for a comparison with field surveys before relying on any mapping tool's rankings.

👉 Read the paper

3️⃣ A language-model project that reports what did not work

A case study on Hugging Face, with a public repository, describes adapting an open model, AfriqueQwen-4B, to Amharic using GitHub Copilot, with separate AI reviewers challenging the work. One AI-proposed way of dividing text into training and test sets balanced the numbers almost perfectly, and a separate AI reviewer spotted that it left every multi-passage article out of testing. After a five-million-token pilot, the model's error score on a small set of existing validation texts fell by 2.2% to 4.1%. In the author's own human evaluation, the adapted model was favoured in 45 of 83 clear comparisons (54.2%), and the author reports that this did not establish a clear improvement. It is one author's case study. Only documentation and selected evidence are public, with no model, data or code released.

Why it matters: A number that improves can still hide a gap, and another AI reviewer found one here. Reporting a mixed result openly makes the lesson usable by everyone else.

👉 Action: Before your next AI pilot starts, write down what result would count as success, and who outside the team will check the test.

👉 Read the case study

🎨 Weekend Playground

This weekend, try Guided Learning in Google's Gemini, a mode that guides you with questions as you build understanding of a topic. Google launched it in August 2025 as a learning companion that breaks a topic into steps, with questions, images and quizzes. This November, they rolled it out globally inside Google Classroom and began pilot testing directly within Google Search for select US schools. This is a real-life example of how AI can personalise learning.

Why this matters: It lets you feel what it is like to be led towards an answer by questions by a personalised learning companion.

👉 Mission:

• Pick one thing you are curious about and know a little of, such as how bees support our food or the basics of photography

• In the Gemini app on iPhone, tap Add Files, then Guided Learning, as Google's help page describes. You need to be signed in, and Google says it is rolling these tools out gradually in the mobile app. If you use Android or cannot see it yet, try the Gemini website in a browser. If it is not available on your account or in your region, ask a friend who knows the topic to quiz you instead

• Spend 15 minutes. Tell it what you already know, and answer its questions in your own words before asking for a hint

• Close the app, leave the phone behind and take a 20-minute walk. Let your mind wander back over the topic

• When you get home, write down three things you remember with pen and paper, then check them against what Gemini said

📢 Share the Optimism

If The AI Optimist helps you think more clearly, forward it to someone else handling the shift.

And here's the question I'm curious about this week: which one task would you want your team to try first before asking AI? Reply and tell me - I read every message and I'll come back to you personally.

Stay strategic, stay generous.

Hugo