Forced to Pick, AIs Split by Party; Data Center Poll; Character.AI

“Last October, it promised to end open-ended chats for under-eighteens. Futurism, a news site, tested that. Using accounts registered as minors, it easily found bots about suicide, self-harm, eating disorders, child abuse and school shootings.”

The chronology is reversed. The Oct 24, 2026 article recaps earlier investigations, then describes restrictions introduced in response. Its linked suicide investigation is dated Oct 29, 2024; eating-disorder reporting Nov 25, 2024; school-shooting reporting Dec 17, 2024; and notification follow-up Feb 3, 2025. These precede the Oct 29, 2025 ban announcement. The source does not establish a new test of that ban. The beat dated Oct 24, 2026 and labeled "Futurism's teen-account test", and the on-screen description of that article as a teen-account investigation, repeat the error. The retrospective introduces the tests with "Throughout this reporting" and describes the later ban under "In response to the mounting pile of lawsuits".

Fix: Say: "Futurism's new retrospective revisits safety failures documented in 2024 and early 2025. Character.AI later announced an end to open-ended chats for under-18s." Replace the timeline with the original reporting dates and label Oct 24, 2026 as a retrospective. Remove the claim that this tested the later ban unless a documented new test is obtained.

“Parents: a news site just tested Character dot AI's teen limits.”

"Just tested" presents old investigations as a new assessment of current restrictions. The underlying tests predate the October 2025 ban announcement. Publication of the retrospective in October 2026 does not date the testing.

Fix: Say: "Parents: a new retrospective revisits Character.AI's earlier safety failures." Change the headline to "Character.AI's earlier safety failures, revisited" and identify Oct 24, 2026 as the retrospective's publication date.

“Last October, it promised to end open-ended chats for under-eighteens. But Futurism, using accounts registered as minors, easily found bots about suicide, self-harm, eating disorders and school shootings. Only warning pop-ups got in the way.”

"But" makes the older tests evidence that the later ban failed. The on-screen comparison explicitly dates those tests to October 2026. The linked original investigations instead document testing in 2024 and early 2025.

Fix: Put the earlier investigations before the 2025 policy change. Replace "Futurism's test (Oct 2026)" with "Futurism's earlier reporting (2024–25)". Explain that these sources do not establish how the current ban performs.

“Futurism used accounts registered as minors and says it easily found bots about suicide, self-harm, eating disorders, child abuse and school shootings, a year after Character.AI promised to end open-ended chats for under-18s.”

The description places the testing a year after the promise. The source is a retrospective of investigations that preceded it. This materially misrepresents the evidence against the company's current safeguards.

Fix: Replace with: "Futurism's October 24 retrospective revisits harmful-bot investigations from 2024 and early 2025, before Character.AI announced its under-18 chat ban." Adjust the introductory description of the story accordingly.

“Using accounts registered as minors, Futurism says it easily found Character.AI bots about suicide, self-harm, eating disorders and school shootings, despite the app's 2025 promise to end open-ended chats for under-18s.”

The Character.AI Short description repeats the false implication that these tests occurred after the ban announcement. Its title also frames the story as a test of the teen limits.

Fix: Describe the October 2026 article as a retrospective of 2024–25 reporting. Retitle the Short "Character.AI's earlier safety failures, revisited #Shorts".

“If one's proposed near you, watch your local planning meetings. That's where tax breaks, and who pays for power upgrades, get decided.”

The Short directs viewers only to planning meetings for decisions that generally also require different authorities. Ohio's utility commission, for example, ordered the data-center tariff protecting other customers from infrastructure cost shifting. FERC's participation guide distinguishes local siting from state retail-rate allocation and federal transmission rules. The second checklist item repeats this misleading advice. FERC specifies that the "state public utility commission determines how federally-approved transmission costs are sub-allocated".

Fix: Say: "Watch local planning meetings for siting, local and state elected officials for tax incentives, and your utility regulator for who pays power-system costs." Update the checklist to include the regulator.

“Character.AI's teen safety test”

This menu beat and the matching stories[].label carry the same misleading current-test framing as the Character.AI narration. The reporting being revisited predates the later ban.

Fix: Use "Character.AI's safety history revisited" in both labels after correcting the story chronology.

“That's one outlet's test, not yet repeated by others.”

The source does not establish this universal negative. Independent safety testing exists: Common Sense Media's April 2025 assessment says harms were "easily replicated in our testing". That is not proof of an exact replication of every Futurism scenario, but it makes the blanket lack-of-repetition claim unjustified. Related assertions occur in short:characterai:characterai-explain and both Character.AI package descriptions.

Fix: Say: "These are Futurism's historical findings; we have not independently tested the current safeguards." Remove the blanket "not yet repeated" claims and the equivalent lower thirds.

“They tested the version programmers use, like a restaurant's wholesale kitchen, with fewer house rules.”

The study establishes different observed behavior across API and consumer-app checks, not a measured general reduction in safety rules. The README says the researchers supplied no custom system prompt and notes hidden provider-added input in Grok requests. The analogy and its "Fewer house rules than the ChatGPT app" beat turn an inference about implementation into fact. The researchers explicitly describe "no system prompt of ours", rather than a controlled comparison of safety policies.

Fix: Say: "They tested models through developer interfaces, which can behave differently from the consumer apps." Replace the analogy's claim about fewer rules with the observed channel difference.

“That's where tax breaks, and who pays for upgrading the power grid, get decided.”

Although this version includes the state utility regulator, it still lumps distinct decisions together. Planning bodies handle siting; tax incentives can require local or state elected bodies; utility commissions and federal transmission proceedings address different electricity costs.

Fix: Replace with: "Local planning bodies handle siting; local and state officials handle tax incentives; utility regulators address power costs." Update the checklist footer.

“Eighty percent of Democrats opposed it. Even Trump's MAGA supporters leaned against, forty to twenty-nine. Only about five hundred people got this question, so allow about four points either way.”

Placed immediately after the partisan figures, the margin-of-error sentence and unqualified lower third can be read as applying to those subgroups. The roughly 500 respondents cover the whole question; partisan subgroups are smaller and have greater sampling uncertainty. The package's data-center Short description has the same ordering and ambiguity. The reported overall ±4.38 figure is corroborated in reproductions, but the official PDF could not be retrieved during this audit.

Fix: Say: "The overall result has about a four-point margin of error; the party breakdowns are less precise." Label the lower third "Overall question: ±4.38 points" and qualify the package wording.

“Disney's twenty-million-dollar opening bid is only for shared permission to use some of its intellectual property.”

"Only" is too absolute. The company's announcement and bidding motion also include acquisition of certain claims against former employees; transition services are part of the arrangement. The central distinction that Disney is not buying the operating app is supported. The motion expressly includes the "purchase of unreleased claims".

Fix: Say: "Disney's twenty-million-dollar opening offer centers on a non-exclusive license to some technology and patents, rather than purchase of the app."

“CBS News — Character.AI files for bankruptcy, Disney license, Oct 2026”

This on-screen source has no corresponding URL in the scene's facts or package source list. The matching CBS report exists and is dated October 13, 2026. All actual facts[] entries elsewhere have nonempty source URLs.

Fix: Add the CBS URL to the scene's source mapping and package sources, or remove the unused credit. Use October 13, 2026 for its date.

“Convovo asked GPT-6.1 Sol and Grok 4.7 to name one candidate in 479 US midterm races (34 Senate, 410 House, 35 governor), each with one Democrat and one Republican”

The supporting fact overstates the panel's composition. The README's own results include a category labeled 'Independent where no Democrat runs', with two GPT verdicts. Its prose description of the panel is internally inconsistent on this point. The race total and the two headline party totals are supported; the assertion that every race had a Democrat is not.

Fix: Remove 'each with one Democrat and one Republican' from this supporting fact. Retain the verified 479-race total and office breakdown.

“Asked plainly, the regular ChatGPT app never picked, in eighteen tries.”

The count is correct, but 'the regular ChatGPT app' leaves out a relevant limitation of the consumer comparison. The README calls these 'One-off checks, not a panel run': six races, with ChatGPT reached logged out through DataForSEO's scraper. This is a narrower observation than testing the range of ordinary signed-in app experiences. The same qualification would improve the Short and package description.

Fix: In a small check of six races, logged-out ChatGPT gave no pick in eighteen tries. Apply the same scope qualification to short:gptgrok:gptgrok-explain and the main and gptgrok Short descriptions.

“Think of a giant factory whose raw materials are electricity and cooling water.”

This analogy makes a continuing supply of cooling water sound inherent to every AI data center. Cooling designs vary. Microsoft's primary account describes a closed loop operating 'without requiring a fresh water supply'; its earlier production immersion-cooling account also describes rejecting heat through a dry cooler. The accompanying beat, 'Raw materials: electricity and cooling water', repeats the overgeneralization. The scene's existing citations do not establish this technical definition.

Fix: Think of a giant factory that needs electricity and cooling; some designs also consume large amounts of water. Change the beat to 'Needs electricity and cooling; water use varies' and add a technical source URL to facts.

“A data center is a warehouse of computers that runs AI, fed by electricity and cooling water.”

As in the roundup, the wording and 'Fed by electricity and cooling water' beat treat cooling-water consumption as universal. Closed-loop and dry-cooling designs make that an inaccurate general definition.

Fix: An AI data center houses computers that need electricity and cooling. Some designs consume substantial water. Change the beat to 'Electricity and cooling; water use varies' and add the technical source.

“If one's proposed near you, watch your city or county planning meetings: that's where siting gets decided.”

Watching local meetings is useful advice, but the assertion about where the decision is made is too categorical. Planning commissions can recommend while elected boards approve; some projects follow administrative approval routes. Loudoun County's March 2024 official statement illustrates the distinction between by-right applications and applications reviewed by its commission and board. No scene fact currently supplies a source for the planning-meeting assertion.

Fix: If one's proposed near you, check your local planning office and public-hearing agendas to learn who approves it and when you can comment. Add a local-government source URL. Qualify the similar siting checklist in the roundup, Short and Short description.

“Users can export or delete their data, the company says.”

The company supports both options, but its linked help page says exports can take '3-4 weeks to process'. In a story about a possible shutdown, distinguishing an export request from an immediate download would make this advice more useful. The deletion link specifically describes deleting the account.

Fix: The company says users can request a data export or delete their account. Exports can take three to four weeks. Update the accompanying beat and add the help-page URLs.

“Character.AI users: the company says you can export or delete your data.”

The description omits the export delay and describes account deletion less precisely than the linked company instructions. See the corresponding roundup finding.

Fix: Character.AI users: you can request a data export, which may take three to four weeks, or delete your account. Add the two help links.

“So: chatbots aren't neutral voting guides, most voters don't want a data center next door, and teen safety promises need checking.”

This scene makes factual summary claims but has an empty facts array and no source URLs. The earlier scenes provide supporting material, so this is a scene-level traceability gap rather than a newly unsupported story. 'Next door' also changes the poll's actual distance wording.

Fix: Add facts with source URLs for the three summaries. Say 'most voters in a CNBC poll would oppose an AI data center within ten miles of home' and adjust the icon-grid label to match.

“Full edition: {{ROUNDUP_URL}}”

All three Short descriptions contain an unresolved URL placeholder. It is suitable for a template but is not a working viewer link in the package as checked.

Fix: Replace {{ROUNDUP_URL}} with the actual published roundup URL in all three Short descriptions before publication.