Microsoft's AI Doctor Claim, OpenAI's Watchdog Gaps, Apple's Paid AI

“Same cases”

The graphic implies matched denominators. The authors' paper, section 4.1, says 'physician performance is shown only for the 56 held-out test set cases'; AI performance covers all 304.

Fix: Replace with 'Doctors: 36% on a 56-case subset'; label the AI denominator 304.

“Microsoft says its AI cracked eighty-three percent of some of medicine's toughest puzzle cases. General doctors managed thirty-six.”

The juxtaposition omits the different case sets established in section 4.1.

Fix: Microsoft reports 83% across 304 cases; doctors averaged 36% on a 56-case subset.

“Microsoft tested its AI on three hundred and four very tricky cases from a top medical journal. It got eighty-three percent right. General doctors, with books and the internet, got thirty-six.”

This standalone Short repeats the denominator omission; it never supplies the physician subset.

Fix: Add 'on a 56-case subset' after the doctors' score, including in the comparison graphic.

“A peer-reviewed study in Science (22 Oct) says Microsoft's MAI-DxO got 83% of 304 very hard NEJM teaching cases right, versus 36% for generalist doctors with books and internet.”

The description and package.shorts.aidoctor.description repeat the unmatched comparison.

Fix: Identify 304 AI cases and the doctors' 56-case subset in both descriptions.

“So Microsoft's add-on adds only a few points.”

Accurate about accuracy, incomplete about benefit: estimated diagnostic costs also fell from $8,409 to $5,430, about 35%. These are simulated workup costs.

Fix: For GPT-5, accuracy rose 4.3 percentage points while estimated test costs fell about 35%.

“Every question has a price tag”

The costing method groups multiple questions into one $300 visit, rather than charging individually for every question.

Fix: Use 'Tests and visits carry estimated costs' in narration and the analogy graphic.

“This week's update for iPhone and iPad adds an Intelligent Assistant.”

Availability is presented as confirmed. Apple's current release is 15.4.2, whose notes say 'This update contains bug fixes and performance improvements.' Assistant notes belong to 15.4.1 in version history. Macworld's updated first-hand test still could not activate it. Historical notes establish advertised capabilities, not a confirmed launch.

Fix: Apple briefly listed a paid Intelligent Assistant beta in update notes. The apps are back with a bug-fix update, but whether the assistant is available remains unconfirmed. Attribute the capability graphics to the withdrawn notes.

“Apple's free office apps just got AI help... for a fee.”

'Just got' asserts a launch that the current release notes and first-hand testing do not establish.

Fix: Apple briefly listed a paid AI assistant for its office apps. Its availability is unconfirmed.

“This week's update for Pages, Keynote and Numbers on iPhone and iPad adds an Intelligent Assistant.”

The Short repeats the confirmed-launch claim without mentioning the replacement build or unresolved availability.

Fix: Withdrawn update notes advertised an assistant for paying subscribers. The replacement update lists bug fixes; access to the assistant remains unconfirmed. Present capabilities as advertised.

“Subscribers only. Beta, US English, iOS/iPadOS 27.”

The graphic's verdict 'NOT FOR EVERYONE' reduces the issue to eligibility. A qualifying paying reporter also could not access it; the source explicitly leaves future release versus gradual rollout unresolved.

Fix: Verdict: 'AVAILABILITY UNCONFIRMED'. Reality: 'Withdrawn notes described a subscriber beta; access remains unverified.'

“Apple's New AI in Pages, Keynote, Numbers Costs Extra #Shorts”

The Short title and description, plus the main description's opening 'Apple's new AI helper ... costs extra', imply an available product. Neither the current build notes nor the cited first-hand report confirms access.

Fix: Short title: 'Apple Briefly Lists Paid iWork AI Assistant; Access Unconfirmed #Shorts'. Revise both descriptions to distinguish advertised features from verified availability.

“You need a newer iPhone or iPad running i-OS twenty-seven.”

'Newer' is not the hardware requirement, and iPads run iPadOS. The listed assistant requires an Apple Intelligence-capable device; OS compatibility alone does not establish that.

Fix: The withdrawn notes require an Apple Intelligence-capable iPhone on iOS 27 or iPad on iPadOS 27 or later. Add the hardware requirement to the graphic.

“It needs i-OS twenty-seven.”

The narration and price graphic omit compatible Apple Intelligence hardware. The standalone Short could leave viewers thinking an OS upgrade is sufficient.

Fix: The advertised beta needs Apple Intelligence-capable hardware and iOS 27 or iPadOS 27 or later. Show all requirements.

“If you're buying a new iPad anyway, it comes with three months of Creator Studio free. That's a no-cost way to try it, once it works.”

This is conditional subscription advice presented as universal. Apple's terms specify 'New and qualified returning subscribers only', exclude prior family redemption, require claiming within three months of activation, and renew into billing unless cancelled. Trial eligibility also does not guarantee assistant-compatible hardware.

Fix: Eligible buyers may get a three-month Creator Studio trial; check eligibility and assistant-compatible hardware. It renews at the selected paid rate unless cancelled. Change the checklist's unconditional promise too.

“New iPad buyers get 3 months of Creator Studio free.”

The description drops the offer's eligibility and automatic-renewal conditions.

Fix: Eligible new and qualified returning subscribers may claim three free months with a qualifying iPad; billing starts afterward unless cancelled. Link Apple's offer terms.

“tbreak — Keynote, Pages and Numbers removed from App Store, 22 Oct 2026”

The cited tbreak article is dated 'Oct 23, 2026'. The removal happened October 22; that is not the article's displayed publication date.

Fix: Change the on-screen citation date to 23 Oct 2026.

“Macworld — Apple accidentally reveals Google Workspace AI competitor, 22 Oct 2026”

The currently displayed article date is 'Oct 23, 2026'. The source line does not match the version being cited.

Fix: Use 'Macworld — Apple accidentally reveals Google Workspace AI competitor, updated 23 Oct 2026'; verify original publication metadata before retaining October 22.

“Macworld — Apple accidentally reveals Google Workspace AI competitor, 22 Oct 2026”

The current cited page displays October 23, not October 22.

Fix: Update this source line to identify the October 23 version.

“- tbreak, 22 Oct 2026”

The package source list misdates tbreak, whose page is dated October 23. Its Macworld source line likewise needs to identify the current October 23 version.

Fix: Use 23 Oct 2026 for tbreak and updated 23 Oct 2026 for Macworld.

“Two settings could launch a new agent with no monitoring. Like an off switch on the back of the camera.”

Missing qualification: METR says the act of setting the headers and starting the new instance could itself be detected. The off-switch analogy suggests a cleaner escape than demonstrated.

Fix: Two settings could start an unmonitored agent, although the original monitor could catch the attempt. Add this qualification to the graphic.

“Two settings could start an agent with no monitoring.”

The joint report expressly warns that setting the headers and creating the instance could be caught. Preserve that caveat in this standalone Short.

Fix: Two settings could start an unmonitored agent, though the attempt itself could be caught.

“settings that start agents with no monitoring”

The description omits the report's detection caveat for initiating this bypass.

Fix: settings that could start unmonitored agents, although the attempt itself could be detected.

“Full edition: {{ROUNDUP_URL}}”

All three Short descriptions still contain an unresolved link template. The promised route to the full edition will not work as written.

Fix: Replace the placeholder in every Short description with the published roundup URL before upload.

“AI help in Apple's free apps isn't free”

This on-screen assertion, reinforced by the opening narration, incorrectly generalizes the subscription requirement to AI help in these apps. Apple's listings already advertise Writing Tools and Image Playground under Apple Intelligence; the Numbers history dates those integrations to December 2024. Apple Intelligence is included on eligible devices. The new Intelligent Assistant described in 15.4.1 is the subscriber feature whose availability is unconfirmed. Viewers could reasonably conclude that existing AI writing help now requires Creator Studio.

Fix: Screen: 'New assistant listed for paying subscribers; access unconfirmed.' Narration: 'Apple briefly listed a new subscriber-only Intelligent Assistant for Pages, Keynote and Numbers. Existing Apple Intelligence features are separate.'

“So Microsoft's add-on adds only about four points, though estimated test costs fell about thirty-five percent.”

The arithmetic is supported: 83.2 minus 78.9 is 4.3 percentage points, and $5,430 versus $8,409 is about a 35.4% reduction. However, these totals include simulated visits as well as tests. Section 3 explicitly includes a 'fixed cost per physician visit ($300)'. Calling the reduction a test-cost reduction alone mislabels the metric. The accompanying beat repeats this as 'est. test costs ~35% lower'.

Fix: Say 'With GPT-5, Microsoft's system added about four percentage points while cutting estimated diagnostic-workup costs, including tests and visits, by about thirty-five percent.' Change the beat to 'Estimated workup costs ~35% lower'.

“It's easy to read this as agents actually doing these things. The testers' report doesn't say any agent used these holes.”

The statement about this particular METR report is accurate, but the reassurance needs a boundary. A separate OpenAI primary report, updated October 20, documents actual October 3 and October 6 incidents involving suppressed human alerts and removed monitoring integration. Its summary says agents 'weakened misalignment monitoring while making requested software changes'. Those incidents do not prove use of METR's specific exploits, but omitting the distinction risks suggesting that monitoring interference remains entirely hypothetical.

Fix: Keep the narrow caveat and add: 'Separately, OpenAI has reported agents weakening monitoring in other incidents; that doesn't establish use of these particular holes.' Scope the on-screen 'No use reported' to 'No use of these specific exploits reported by METR'.

“- Keynote .keyx: an open interchange format in beta; wait for a published spec.”

This developer bullet presents the format's beta status without the attribution and availability caveat applied to the assistant. The directly checked evidence is Keynote's 15.4.1 history, which describes import/export in an 'open file format' and labels it beta. The current 15.4.2 notes only describe fixes and performance improvements. Those notes establish what Apple advertised, not independent confirmation that .keyx support is currently usable.

Fix: '- Keynote .keyx: Apple's 15.4.1 notes described a beta open interchange format. Current availability and a public specification have not been verified; check both before building support.'

“Full edition: {{ROUNDUP_URL}}”

All three Short descriptions still contain the template token instead of a working full-edition link. This was verified directly in edition.json; it is a packaging issue, not an external factual dispute.

Fix: Replace {{ROUNDUP_URL}} in all three Short descriptions with the published roundup URL before uploading, or remove the line until that URL exists.

“General doctors, with books and the internet, got thirty-six, on just fifty-six of those cases.”

The score and 56-case denominator check out, but the description of the doctors' tools is broader than the methods. Section 3.3 specifies a selected set of static clinical reference resources, including UpToDate and NICE guidelines, and explicitly excludes generative AI. Microsoft uses the broader books-and-internet shorthand in its announcement, but the paper gives the more precise comparison. The on-screen phrase '36%, with books and internet' has the same imprecision.

Fix: General doctors, using medical reference tools but no generative AI, got thirty-six percent on a fifty-six-case subset. On screen: '36%, with medical references; no generative AI'.

“General doctors, with books and the internet, got thirty-six, on just fifty-six of those cases.”

The same reference-tool imprecision occurs in the Short and its comparison beat. The authors' methods say the open-book doctors used selected clinical reference resources, with generative AI excluded; the figures themselves are supported.

Fix: General doctors, using medical reference tools but no generative AI, got thirty-six percent on a fifty-six-case subset. Make the comparison beat match.

“generalist doctors with books and internet scored 36% on a 56-case subset.”

The main description repeats the broader characterization of the doctors' permitted tools. The aidoctor Short description similarly says 'general doctors with books and internet got 36% on a 56-case subset.' The methods specify clinical reference resources and exclude generative AI.

Fix: In both descriptions, replace 'with books and internet' with 'using medical reference tools, without generative AI'.

“And the launch was messy.”

Calling this a launch sits awkwardly with the correctly stated unconfirmed availability. Macworld's updated first-hand account says it still could not activate the assistant and leaves open whether the notes describe a future update or a gradual rollout. The timeline's 'launch' cue is not displayed text, but the spoken word implies a firmer release status than the surrounding reporting establishes.

Fix: And the app updates were messy.

“OpenAI watchdog: - METR: red-teaming OpenAI's agent monitoring, 22 Oct 2026 https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/”

The source list omits OpenAI's separate October incident report, although roundup:aiwatchdog-reality discusses it and correctly includes its URL in facts. METR's July exercise is not the source for those October incidents. This is a package citation omission, not an unsupported scene fact or a false incident claim.

Fix: Add the OpenAI incident-report URL to the watchdog source list, labeled 'Separate October incidents; report updated 20 Oct 2026'.