In partnership with

The AI Field

Welcome back to The AI Field. Microsoft says its AI now beats general doctors at cracking hard medical cases, and the results just ran in one of science's top journals. The catch: it has only been tested on past cases, not on real patients.

Inside: what the study actually shows, a safer way to let AI agents into your accounts, and Canva's first image model. Plus five tools and a five-minute way to find your recurring charges without handing an AI your bank login.

In today's AI Field:

  • Microsoft's AI out-diagnoses doctors on the hardest cases

  • Banks are blocking AI agents, and 1Password has a safer way in

  • Canva launches its own AI image model

Plus: five tools, three quick hits, and a five-minute way to find your recurring charges without sharing your bank login.

Read time: 5 minutes

100+ ChatGPT Prompts to Revolutionize Your Day

Supercharge your productivity with HubSpot's comprehensive guide to Chat GPT. This free resource is your fast track to AI mastery:

  • Industry-Specific Use Cases: 15+ real-world applications across various sectors

  • Productivity Guide: 21 best practices to 10x your efficiency with AI

  • Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation

  • Challenge Buster: Overcome common AI hurdles with expert strategies

Plus, in-depth sections on email composition, content creation, customer support, and data analysis.

LATEST DEVELOPMENTS

MICROSOFT · HEALTH AI

Image: Microsoft AI.

The AI Field: Microsoft's diagnosis system, MAI-DxO, got the right answer on about 83% of hard case reports from the New England Journal of Medicine, in a study published Thursday in Science. General doctors working alone with reference materials got 36%, and 22% without them.

The details:

  • It works like a panel of doctors: several AI agents ask about symptoms, decide which tests to order and argue their way to a diagnosis. Microsoft also tracked what those tests would cost, and its setup brought the estimated cost per case down.

  • On real emergency room cases from Beth Israel Deaconess Medical Center in Boston, it scored 80% against an average of 47% for doctors. The gains held, and even improved, when Microsoft reran the test with newer models like OpenAI's GPT-5.5.

  • The fine print: cases this hard usually go to specialists or teams, who weren't part of the comparison. Much of the jump comes from stronger AI models themselves, and Microsoft's setup helped most on complex cases, not the ER ones.

Why it matters: This is a test on old cases, not a product. Microsoft plans to start trials at Beth Israel in the coming months, where patients who agree get normal care plus a parallel check from the AI, and Mustafa Suleyman says it still "needs clinical validation." The useful lesson today is the method: questions and tests step by step beat a one-shot answer. If you ask a chatbot about symptoms, tell it to question you first and list what a doctor might test, then bring that to a real doctor.

BANKS · AI AGENTS

The AI Field: Big US banks are finding AI agents inside customer accounts, usually signed in with the customer's own password, and several are now blocking them. The same day, 1Password launched 1Password for Agents, which lets an agent sign in for you without ever seeing your password.

The details:

  • Morgan Stanley has started blocking outside AI agents, and JPMorgan, Bank of America and Citi are restricting them. Most banks are moving toward read-only access, where an agent can see your balance but not move money.

  • Banks say they often can't tell whether the customer sent an agent or it acted on its own. They haven't seen customer funds lost to a hack, but a Chase spokesperson said customers "should not give their usernames and passwords to anyone."

  • With 1Password, the agent proves who it is with its own digital key. You get a notification saying which account it wants and why, much like a two-factor prompt, and once you approve it, 1Password fills in the login itself.

  • It works at launch with Claude Duet, ChatGPT Work and Dots, Meta Muse, Grok Bot and Town. Perplexity, Ramp, Browser Use and others are working on support.

Why it matters: Pasting a password into an agent chat leaves a copy you no longer control. Approving each login is the safer pattern, and it's likely where agent errands are heading. If you've ever pasted a password into a chat, change it today, and keep your bank out of agent chats until your bank offers its own read-only access.

The Agent Field

TRY THIS TODAY · 5 MINUTES

Find your recurring charges without giving an AI your bank login

The simplest way to follow the banks' advice: keep bank logins out of agent chats and work from downloaded statements instead.

1. In your bank's app or website, download your last 3 months of statements as PDF or CSV.

2. Open an AI chat that accepts file uploads, attach the files and paste this. Replace [TASK] with what you want, like listing your recurring charges:

"Help me with my bank account without logging in for me. I'll download my last [3] months of statements as PDF or CSV and attach them here. From those, [TASK, e.g. list my recurring charges]. Never ask for my bank username, password or one-time codes, and tell me if a step would need them."

3. Check the list against your bank app. Before you ever connect an agent to your bank, check your bank's rules and use read-only access if it offers it.

The AI newsletter curated by engineers who build it

TLDR AI is a free daily email curated by Anthropic and ex-Google engineers: the new tools worth trying and the research worth understanding, summarized for 1.1M+ readers.

CANVA · AI IMAGES

Image: Canva. AI-generated examples from Canva Craft.

The AI Field: Canva released Canva Craft, its first in-house image generator, and it's available now to all users worldwide. It plans the layout before it draws, which Canva says keeps text readable and puts objects where you asked.

The details:

  • It makes images up to 2K resolution in more than 30 aspect ratios, so one idea can become a post, a story or a banner.

  • Canva trained it on its own templates, like posters, menus and social posts, rather than photos. That means a result can be turned into a layered design you can edit piece by piece.

  • It ranks fourth on Design Arena's crowdsourced image leaderboard, behind OpenAI's GPT-Image models. Faster Craft-fast and Craft-ultrafast versions are coming, and Canva plans to let ChatGPT and Gemini use it too.

Why it matters: Most image models still mangle the words on a flyer or menu. If you make social posts or flyers, try Craft on something with text in it: open Dream Lab from the three-dot menu on the left, then fix one element in the layers without redoing the whole image. That's the part worth judging, not the first render.

👉Try now!

TRENDING AI TOOLS

🧱 Base44: Describe the app or website you want in plain words, and it builds the whole thing for you, with logins, a database and hosting already included. Free plan with up to 5 apps and 25 message credits a month; paid plans start at $16 a month, billed yearly.

⚙️ Make: Connect the apps you already use and let it handle the repetitive steps between them, like saving new form replies to a spreadsheet and sending a Slack message. It works with more than 1,000 apps. Free plan with 1,000 credits a month.

✨ Genspark: An AI workspace where you ask for slides, documents, spreadsheets or research, and its agents put them together for you using more than 70 AI models. Free plan with 100 credits a day, up to a lifetime limit; Plus is $24.99 a month.

📊 DeepCell: Give it a deadline job like a budget forecast or monthly report, and it builds the spreadsheet, write-up and slides together, with sources and a cheat sheet to explain your numbers. Change one figure and the linked parts update. Free for individuals.

🎬 CutBay: Make a video by chatting. Describe what you want or drop in a PDF, slides or a link, and it handles the script, visuals, voiceover, subtitles and edit. Free to start.

📹 Loom: The new Record for agents feature lets you record your screen, talk through a task and hand the video to an AI tool as instructions. Loom has a free plan; this feature is in Business + AI, with a 14-day free trial.

🚀 Fastlane: Add your website and its agent learns your brand, studies competitors and makes image and video ads, then keeps the ones that work. Your first 5 ads are free; paid plans start at $29 a month.

🎥 HeyGen Video: Describe a shot, pick a style, and it makes a short clip with the picture, dialogue and sound effects in one go. It's ranked first on Artificial Analysis' text-to-video leaderboard. Free for 2 clips a day; longer clips and 2K are in the full HeyGen app.

QUICK HITS

💰 OpenAI nears a raise at a $1.25 trillion valuation: The round of up to $30 billion is led by UAE investors and Coatue, and the $1.25 trillion is OpenAI's value before the new money comes in. It's set up as a bridge before a possible IPO, and OpenAI hopes to have signatures by the end of next week.

🤖 Claude's auto mode becomes the default next week: On Pro, Max and Team plans, new sessions in most Claude apps will stop asking you to approve each step. An automated check lets safe actions go ahead and blocks risky ones, and you can still switch back to approving steps yourself.

🌐 Arc browser shuts down on December 22: Arc stops working on Mac and Windows that day, and account data is deleted within 30 days. Its makers suggest moving to their newer browser, Dia, which imports your passwords, tabs and bookmarks, or you can export your data from Arc's settings first.

That's it for today!

Was this edition worth your time? Leave a rating to help me make The AI Field more useful for you.

Login or Subscribe to participate

See you soon,

Olle Hellman | Founder of The AI Field