Meta AI suggestions are being changed after one of them asked a mother, under a video of her singing in the car with her young daughter, “Who’s the child passenger?” On 11 September 2026 The Verge reported that Meta has fixed the issue that made its assistant suggest prompts about personal topics. Meta spokesperson Dina El-Kassaby said the company “missed the mark” and that “the feature never should have prompted the individual with questions like that.” The mother, Utah travel creator Kalie Robins, says the answers that followed named both of her daughters and tried to work out where her family lives.
The story has mostly travelled as a viral scare, so we rebuilt it from the record. We dated Robins’ Instagram warning video from its own post ID, lined up all the Meta AI suggestions reported in eight articles from seven outlets, and read Meta’s 64-word statement one sentence at a time. We compared it with the 67-word statement Meta issued in July, when it switched off a Muse Image feature using the same three words. Then we checked what the change to Meta AI suggestions covers, which rules apply, and what the same failure looks like inside a business.
The short version is that Meta changed the questions, not the answers. Its statement defends the mechanism that did the digging, says nothing about the photo Robins believes she deleted, and offers an explanation for the children’s names that her own account contradicts. We covered the run-up to Meta’s new agent in Meta Muse Could Be Announced This Week and a police tool built for the same kind of digging in Clearview AI’s InquiryIQ test. This piece is about what happens when Meta AI suggestions become the dossier builder.
Table of contents
- What Happened When Meta AI Suggestions Met a Family Video
- How Meta AI Suggestions Were Supposed to Work
- Eight Days From a Warning Video to a Fix for Meta AI Suggestions
- Meta’s Statement on Meta AI Suggestions, Sentence by Sentence
- “Missed the Mark” Twice in 62 Days: Meta AI Suggestions and Muse Image
- What the Fix to Meta AI Suggestions Changes, and What It Leaves
- Why “Already Has Access To” Is the Wrong Test for Meta AI Suggestions
- Meta AI Suggestions Landed in the Same Week as Muse
- The Rules That Bear on Meta AI Suggestions About Children
- The Enterprise Version of Meta AI Suggestions
- What Parents and Creators Can Do About Meta AI Suggestions Now
- Questions Meta Has Not Answered About Meta AI Suggestions
- Meta AI Suggestions and Personal Questions: FAQ
- References and Further Reading
What Happened When Meta AI Suggestions Met a Family Video
The facts below come from Robins’ own videos as quoted by reporters, from screenshots Futurism says it reviewed, and from Meta’s statements. Where the accounts differ, we say so.
The video and the question beneath it
Robins posted a short clip of herself and one of her two daughters singing in the car. Futurism reports that the child’s face was visible but Robins did not say her name, and that none of the posts on her account publicised her children’s names. The clip was cross-posted to Facebook. When Robins returned to it, one of the Meta AI suggestions sat underneath: “Who’s the child passenger?”
“First of all, why … is that even a suggested question about a minor?” she said in her video, as quoted by News18. Then she tapped it. The Verge prints the first of the Meta AI suggestions as “Who is the child passenger?”; every outlet quoting Robins’ own words uses the contraction.
What each of the Meta AI suggestions led to
The first prompt was not the last. Futurism says Meta AI “continued to recommend new questions, each more invasive than the next”, according to screenshots Robins shared. This table lines up every prompt the coverage reports and what its answer surfaced.
| Suggested prompt, as reported | What the answer surfaced | Where it reportedly came from | Reported by |
|---|---|---|---|
| “Who’s the child passenger?” | Both daughters’ names, birth information, photos and videos | Her posts, relatives’ posts, a newborn photo on her mother’s page | Futurism, News18, The Mary Sue |
| “What are Kalie’s daughter’s ages?” | An age breakdown and timeline; one daughter’s birth weight | “family birthday posts and birth announcements” | Futurism |
| “Where does Kalie Robins live?” | Several paragraphs on where she was likely based | Older and newer posts, including a 2018 post that “congratulated her” | Futurism, News18, The Mirror |
| Prompts about the girls’ lives and interests | Likely school grade, favourite beaches, books and hiking trails | Robins’ recollection of the answers | Futurism |
| Prompts about her work | Her “mental health advocacy” and both parents’ employment histories | Not specified | Futurism |
While the location answer loaded, Meta AI displayed a status line: “Pinpointing Kalie Robins’ location.” Robins recalled one answer reasoning from the girls’ ages to their school grade. “Well, there’s only one school here, so that wouldn’t be hard to find,” she told Futurism.
Who Kalie Robins is
Robins is not a celebrity. Futurism describes a travel creator in Utah with public Instagram and Facebook accounts and “only a few hundred followers”. She has used Facebook since it “first came out”, lives far from her extended family, and relies on relatives posting milestones to stay close. That is exactly the use Meta has always said its apps are for.
Her warning video spread far beyond that audience. IBTimes Singapore cited reports of more than 5.5 million views, and the Mirror counted more than 13 thousand comments. Robins told IBTimes Singapore: “Everything that it told me was information that had been shared publicly at some point.”
The photo she says she deleted
One claim goes further than aggregation. Robins says the results included a family photo she deleted years ago. “I’m not going to pretend that I know how it works, but I know this picture has not been sitting on my page for years,” she said, as quoted by News18.
That claim is unproven either way. IBTimes Singapore pointed out that it does not show whether Meta AI retrieved a deleted copy, found another version elsewhere, accessed an older post or used some other source. What is certain is that Meta’s statement does not mention it.
How Meta AI Suggestions Were Supposed to Work
The feature behind Meta AI suggestions is not new. What changed is how far the assistant could reach.
The northern lights example from 2024
On 18 April 2024 Meta announced that people could use Meta AI while scrolling Facebook. “Come across a post you’re interested in? You can ask Meta AI for more info right from the post,” the company wrote. Its example was gentle: see “a photo of the northern lights in Iceland” and ask “what time of year is best to check out the aurora borealis.”
That was 867 days before Robins posted her warning video. In between, Meta turned its chatbot into something closer to a full assistant, as we reported in Meta AI Assistant: Meta Is Making Its AI Chatbot More Like an Assistant. The distance from “when can I see the aurora?” to “who is this child?” is the distance between asking about a post’s subject and asking about the people in it.
Four names for Meta AI suggestions
Nobody agrees what to call Meta AI suggestions, which makes the fix harder to pin down. Meta’s statement says someone “clicks on a prompt”. The Verge writes of “prompts suggested by its AI chatbot”. The Mirror calls each one “a Meta AI question”. And in its July Muse Image post, Meta described a panel of “presets (suggested prompts)”. Four labels, and no public help page we could reach that defines which of them the fix covers.
What Meta says the assistant can read
Meta has now described the reach behind Meta AI suggestions three different ways. Futurism reports that Meta confirmed Meta AI can search a user’s own content as well as public information on the web. IBTimes Singapore quoted Meta saying the assistant “can only access information that you have chosen to share with it or that is publicly available to it.” The statement itself says answers are “based on information the user already has access to.”
Those are not the same boundary. On Facebook, a user can see far more than their own content: friends’ posts, relatives’ posts, and anything public. The third description is the widest, and it is the one Meta chose when defending what happened.
Who else could see the Meta AI suggestions
We do not know whether other people watching Robins’ public Reel saw “Who’s the child passenger?” beneath it. The Mirror describes a Meta AI question as a prompt generated “when analysing user posts” that “also shows the queries users directly type into Meta AI across its apps”. Meta has not confirmed that, and it has not said whether strangers were offered the same Meta AI suggestions. Robins believed they could be: “It pieced together my information to show people where I live.”
Eight Days From a Warning Video to a Fix for Meta AI Suggestions
No report gives an exact date for Robins’ warning video, so we worked it out. The timeline shows how long a problem like this can run in public before a company says anything.
Dating the post from its own ID
Instagram post IDs encode the moment a post was created. The warning video Futurism and The Verge link to, Dcy8LU3S7DF, decodes to 19:01 UTC on 2 September 2026. The two TikTok uploads The Mary Sue links decode to 19:14 and 21:55 UTC on 3 September. That day an X user, Cory Dobbin, posted: “This just gets crazier by the second. @Meta and @finkd need to address this ASAP.”
Meta’s first published response came 182 hours after that video went up. The chart counts the hours to each milestone we could date.
Hours are measured from the decoded post time to each article’s published timestamp. The fix to Meta AI suggestions may have shipped earlier than any of these reports; Meta has not given a date.
How the story of Meta AI suggestions travelled
| Published (UTC) | Outlet | Name as printed | What it added |
|---|---|---|---|
| 6 September, 08:54 | News18 | Kalie Robbins | Earliest write-up we found; the deleted-photo claim |
| 8 September, 15:45 | IBTimes Singapore | Kalie Robins | 5.5 million views; Meta’s line on what its assistant can access |
| 9 September, 04:15 | The Mirror | Kalie Robins | 13 thousand comments; child-safety experts |
| 9 September, 07:57 | The Mary Sue | Kalie Robins | Her TikTok uploads; a follow-up video answering critics |
| 10 September, 09:01 | The Mirror, second story | Kalie Roberts | First copy of Meta’s “missed the mark” statement we found |
| 10 September, 13:29 | Futurism | Kalie Robins | Interview: Utah, the ages prompt, the 2018 post, relatives’ scrapbooks |
| 11 September, 14:25 | The Verge | Kalie Robins | Spokesperson named; fix confirmed; Meta’s explanation for the names |
The eighth article is a syndicated copy of the Mirror’s second story, which ran under the Irish Mirror and OK! banners the same morning. Futurism’s own page carries a correction note: it was “updated to clarify which specific Meta platforms the disturbing interactions happened on.” Its earlier syndicated headline still blames “Instagram’s AI”; the Meta AI suggestions appeared on the Facebook cross-post.
Three spellings of one name
Newsrooms printed Robins’ name three ways. News18 wrote “Robbins”. The Mirror’s second story calls her “Kalie Roberts” in its own sentences, then quotes the prompt “Where does Kalie Robins live”. The irony is hard to miss: the software that raised this alarm got her name right every time.
Meta answered on day eight, two days into Muse
Meta launched Muse, its personal AI agent, at 19:00 UTC on 8 September, 144 hours after Robins’ warning video and while the story was spreading. Its statement on Meta AI suggestions first appeared in print about 38 hours after that launch. Meta published a newsroom post for Muse. We found no newsroom post about the fix.
Meta's Statement on Meta AI Suggestions, Sentence by Sentence
Meta’s full statement to Futurism is three sentences long. The Mirror printed the same wording, and The Verge quoted parts of it. Here is each sentence, what it does and what it leaves out.
| Sentence | Words | What it does | What it leaves out |
|---|---|---|---|
| “We created this feature to help people get more information about topics or posts that interest them.” | 17 | States the purpose | Who decides what a user is interested in |
| “When someone clicks on a prompt and uses this feature, the response it generates is based on information the user already has access to.” | 24 | Defends the answers | Relatives’ posts, strangers on a public post, and the deleted photo |
| “But we missed the mark because the feature never should have prompted the individual with questions like that and we’ve fixed that issue.” | 23 | Admits the prompts were wrong; says they are fixed | What counts as “questions like that”, where the fix applies, and when |
Sentence one: the purpose of Meta AI suggestions
“Topics or posts that interest them” is the northern lights logic of 2024. It frames Meta AI suggestions as help for a curious reader. But Robins was not a curious reader of someone else’s post. She was looking at her own video, and the first of the Meta AI suggestions was about her own child.
Sentence two: the defence of Meta AI suggestions
The longest sentence in the statement, 24 of its 64 words or 37.5%, is not an apology. It argues that nothing improper was retrieved, because the user could already see it. That is a claim about access rights, and it is probably true. It is also beside the point, as a later section explains.
Sentence three: the admission on Meta AI suggestions
The admission is precise about one thing. The feature “never should have prompted the individual with questions like that.” The fault Meta accepts lies in the Meta AI suggestions, not the answers. “We’ve fixed that issue” is singular, and it attaches to the prompting.
What The Verge added
The Verge’s report adds two details. First, it describes the fix as covering prompts “related to personal topics”, a category Meta has not defined. Second, according to El-Kassaby, Meta AI may have pulled from videos where Robins named her children. Futurism’s reporting says the opposite: Robins did not say her daughter’s name in the video, and none of her posts publicised her children’s names. Both cannot be fully right, and Meta has not said which videos it means.
"Missed the Mark" Twice in 62 Days: Meta AI Suggestions and Muse Image
This is the second time in 62 days that Meta has used those three words to explain an AI feature. The first time, the feature disappeared.
July: Muse Image and public Instagram accounts
On 7 July 2026 Meta launched Muse Image, the first image model from Meta Superintelligence Labs. One feature let people generate images by @-mentioning public Instagram accounts. The Verge quoted Meta saying “tagging a username lets Meta AI use public photos to build a visual”. Account holders could opt out, but they had to find the setting first.
The backlash was fast. SAG-AFTRA told its members how to opt out. Haley McNamara of the National Center on Sexual Exploitation wrote that “putting the onus on individuals to jump through hoops to opt out is unacceptable.” On 10 July, Meta added an update to the top of its launch post: “We’ve heard the feedback that this feature missed the mark, so it’s no longer available.”
The same three-part statement, two different outcomes
| Element | July 2026: Muse Image @-mentions | September 2026: Meta AI suggestions |
|---|---|---|
| What the feature did | Used public Instagram photos to generate images of account holders | Suggested questions about people in a post, then answered them |
| Length | 3 sentences, 67 words | 3 sentences, 64 words |
| Defence offered | “give people control over whether their public content could be referenced” | “based on information the user already has access to” |
| Admission | “this feature missed the mark” | “we missed the mark” |
| Remedy | “it’s no longer available” | “we’ve fixed that issue” |
| Where Meta said it | An update on its own newsroom post | A statement to reporters |
| Burden left on the public | Account holders had to find an opt-out setting | Families deleting years of posts |
Both statements follow the same pattern: what the feature was for, why it was defensible, and a short admission. The words per part show where Meta put the emphasis.
In July, the admission was the shortest part at 15 of 67 words. In September, the defence was the longest at 24 of 64. Both times, the explanation of why the feature was reasonable took more words than the apology.
Muse Image removed, Meta AI suggestions fixed
The bigger difference is the remedy. The July feature was switched off entirely. The September feature stays, minus some Meta AI suggestions. Meta also moved faster in July: the removal was reported about 78 hours after its launch post went up, while its first statement on Meta AI suggestions appeared 182 hours after Robins posted her warning.
The two triggers are not identical. In July the launch was public from the first hour; in September one family’s video had to go viral first. But that is the point. Harms that land on one ordinary family take longer to reach a company than harms that land on actors’ unions and talent agencies.
Why the wording matters
“Missed the mark” describes an aim that went slightly wrong. It does not describe a design that put the burden on the people affected. In July, the burden was an opt-out setting. In September, it was a mother asking grandparents to take down years of photos. Robins’ own advice is the opposite of Meta’s framing: “make it hard. Make it so hard to find that information.”
What the Fix to Meta AI Suggestions Changes, and What It Leaves
Put Meta’s statement next to what happened to Robins and the fix to Meta AI suggestions covers one link in a longer chain. This is our reading of each part, based only on what Meta and reporters have said.
| Part of the system | Status after the fix | Basis |
|---|---|---|
| Meta AI suggestions about personal topics | Changed | Meta: “we’ve fixed that issue” |
| Answers when a user types the same question | Not addressed | No statement mentions typed questions |
| Search across the user’s content and the public web | Confirmed and defended | Meta to Futurism; statement sentence two |
| Relatives’ posts that name or show the children | Not addressed | No statement mentions third-party posts |
| The photo Robins says she deleted | Not addressed | No statement mentions deleted content |
| Platforms and countries covered | Not stated | Meta AI runs in Facebook, Instagram, WhatsApp and Messenger |
| Meta AI suggestions shown to strangers on public posts | Not stated | Meta has not said who saw the prompts |
Meta AI suggestions: changed
This part is real and worth having. A system that no longer offers “Who’s the child passenger?” to anyone is better than one that does. Most people never type a question like that. Meta AI suggestions that offer it do the thinking for them.
The answers: defended, not changed
Nothing Meta has said suggests the assistant will now refuse to answer “What are this child’s ages?” if someone types it. The statement goes out of its way to justify those answers. So Meta AI suggestions are safer; Meta AI itself may not be.
The relatives’ posts: untouched
Futurism reports that most of what surfaced was shared not by Robins but by family members, such as a grandmother’s newborn photo and family birthday posts. Robins can delete her own posts. She cannot delete her mother’s, and neither can her children. “If I take everything off my page, but you still have my kids on yours, it will go to your page. I’ve seen it. It already did it,” she said.
The deleted photo: unexplained
If a photo removed years ago really came back, that would be a different and more serious problem than aggregation. If it came from a relative’s copy, it is the same problem again. Meta could settle this in a sentence. So far it has not.
The names: two accounts that do not match
Meta suggests the children’s names came from videos where Robins named them. Robins told Futurism her posts never publicised the names. Both could be partly right, for instance if a name was spoken in an old video Robins had forgotten. But a company defending its data practices should be able to say which videos it means.
Why "Already Has Access To" Is the Wrong Test for Meta AI Suggestions
Meta’s defence rests on one idea: if you could see it, the assistant may show it to you. That idea is common across the AI industry, and this case shows why it is not enough.
Access is not the same as expectation
Every post that surfaced was shared with an audience, for a reason: a birth announcement for family, a congratulation for friends. Nobody posted them to build a profile of two small girls. IBTimes Singapore put it plainly: “The individual posts may not reveal much on their own. Combined, they can reveal considerably more.” Aggregation turns permitted pieces into something nobody permitted.
Minutes, not an afternoon
A determined person could have assembled the same file. Futurism notes that “a human investigator may well have been able to cobble together a similar pile of information”, but Meta AI “was able to do so in minutes”. Robins put it more starkly: an AI could “assemble all those breadcrumbs into information about my minor children in seconds.” Cost is a privacy control. When the cost of assembling a dossier falls to one tap, the protection that cost provided disappears.
Meta AI suggestions are an editorial choice
Robins did not ask who her daughter was. Meta AI suggestions raised it. Meta’s own design team, writing about Muse on 8 September, set a standard that fits here: “the bar for sending one has to be high: a proactive message has to be genuinely helpful and worth the interruption.” Meta’s Muse announcement says the agent “can make suggestions unprompted”. Meta AI suggestions are the same idea, and this one did not clear that bar.
Children cannot opt out of a grandparent’s scrapbook
The children in this story never posted anything. Their information came from parents, grandparents and family friends. Futurism quotes one grandmother deleting her posts: “This is my scrapbook, I have to get rid of my scrapbook?” Robins now asks relatives to “save those precious memories and then remove my kids from your page.” Bystanders who never agreed to be recorded face a similar problem with always-listening wearables, as we found in Apple Watch’s always-listening AI features.
Meta AI Suggestions Landed in the Same Week as Muse
The timing matters because Meta is asking for far more trust than Meta AI suggestions ever needed.
What Muse asks people to hand over
Muse is a personal agent that connects to email, calendars, payments and other services, then acts on the user’s behalf. TechCrunch reports it is free to start, with paid plans at $20 and $100 a month, and available in the US on the web, iOS, Android and WhatsApp. CNBC reports that users must opt out if they do not want Meta to use their Muse interactions to train its AI models. Meta says it will first scrub “critical personally identifying information”. Muse’s feed can also summarise updates from a user’s connected Facebook and Instagram accounts.
The Verge’s Muse test found the same pattern
Emma Roth, who also wrote The Verge’s report on the fix, published her Muse test the day before. Asked what it knew about her, Muse listed interests from anime to Labrador retrievers. Asked where that came from, it said: “Instagram doesn’t show you that exact list anywhere in the app — I read it through the account’s API data, which surfaces more than the UI does.” Her verdict was that Muse “works and creeps me out”. That is the same shape as Meta AI suggestions: an assistant that can see more than the person using it realises.
Meta’s trust record, in its regulators’ words
In 2023 the Federal Trade Commission described its latest action as the third time it had acted against Facebook over privacy. It filed a complaint in 2011 that led to a 2012 order. In 2019 Facebook agreed to a second order, in force from 2020, with a $5 billion penalty. In 2023 the FTC alleged violations of that order and proposed a blanket ban on monetising data from users under 18. TechCrunch also notes a multistate settlement in August, which it values at $18 billion and CNBC at nearly $17 billion. Our earlier analysis of why people aren’t buying Zuckerberg’s AI future covers what that record costs Meta’s AI ambitions.
What Meta says Muse does differently
Meta has published an unusually detailed security design for Muse. Each user’s agent runs in its own isolated virtual machine, a separate “Sentinel” controls every outside action, and the agent never sees real passwords. Meta’s bug bounty pays up to $300,000 for valid reports. CNBC notes rising concern about the cybersecurity risks of AI agents, and those protections target exactly that. They do not address the question Robins raised, which is not about attackers at all. It is about what the assistant chooses to put together for the people it serves.
The Rules That Bear on Meta AI Suggestions About Children
Several rules touch this case. None was written for it. The table shows why.
| Jurisdiction | Rule | What it requires | Fit with this case |
|---|---|---|---|
| United States | FTC’s 2020 privacy order with Facebook | A privacy review of every new or modified product, service or practice before implementation | Covers Meta’s features; whether this one was reviewed is not public |
| United States | COPPA Rule | Verifiable parental consent before child-directed services collect data from under-13s | Aimed at services for children, not adults’ posts about them |
| United Kingdom | ICO Age appropriate design code | Profiling and geolocation off by default; no nudges that weaken privacy | Protects children using a service, not children shown in it |
| United Kingdom | UK GDPR Article 25, amended by section 81 of the Data (Use and Access) Act 2025 | Services likely to be accessed by children must weigh “children’s higher protection matters” in their design | A design duty; untested for children who appear in other people’s posts |
| European Union | Digital Services Act, Article 28 | “a high level of privacy, safety, and security of minors, on their service” | Framed around minors using the platform |
United States: an order that already demands privacy reviews
The FTC’s 2023 announcement summarises the 2020 order: Facebook must “conduct a privacy review of every new or modified product, service, or practice before implementation and document its risk mitigation determinations.” A feature that suggests questions about the people in a post is plainly a product practice. Whether the Meta AI suggestions that asked about a child passenger were reviewed, and what the review concluded, is one of the more useful questions a regulator could ask.
United Kingdom: a code written for child users
The ICO’s code has 15 standards. Standard 10 says to “switch geolocation options off by default” and standard 12 says the same for profiling. Both fit Meta AI suggestions in spirit. But the code applies to services “likely to be accessed by a child”, and it protects the child using them. Robins’ daughters were not using Facebook. Section 81 of the Data (Use and Access) Act 2025 adds “children’s higher protection matters” to UK GDPR’s design duty, as we explained in our guide to the DUAA’s UK GDPR changes. It is the closest fit, and still untested here.
European Union: minors “on their service”
Article 28 of the Digital Services Act requires online platforms accessible to minors to protect their privacy and safety “on their service”. The phrase points at minors as users again. The general GDPR, which treats children’s data as meriting specific protection, is a broader route, but it has no specific rule on AI assembling family posts.
The gap: children as subjects, not users
Meta’s own August agreement with state attorneys general makes the gap visible. Every protection it lists applies to teen accounts: a default two-hour daily limit, a night-time block, school-hours muting. None applies to a child who has no account but appears in years of an aunt’s photos. Children’s online privacy law grew up around children as users. Meta AI suggestions exposed children as subjects.
The Enterprise Version of Meta AI Suggestions
If your organisation is switching on an AI assistant over its own files and email, this story is about you too. The failure has the same parts: broad access, fast aggregation and a system that proposes what to ask.
Copilot uses the same access test as Meta AI suggestions
Microsoft’s documentation for Microsoft 365 Copilot says it “only surfaces organizational data to which individual users have at least view permissions.” That is the same principle as Meta’s “information the user already has access to”, and it is sound as far as it goes. But in most organisations, the files people can technically open are not the same as the files anyone meant them to read. Our coverage of Microsoft’s AI privacy rules for schools found that the new standard leaves general-purpose tools such as Copilot out of scope entirely.
Oversharing becomes answerable
A spreadsheet of salaries shared with “everyone” years ago used to be safe through obscurity. Nobody knew where to look. An assistant knows where to look, and it answers in seconds. Robins’ relatives were not careless; they shared milestones on a platform that later built an assistant on top. Employees sharing files with the wrong group are in the same position.
Treat suggested prompts as published content
The most transferable lesson from Meta AI suggestions is that the suggestion layer needs its own review. A prompt the system offers is a statement by your organisation. Before launch, decide which subjects the assistant may never suggest questions about, even if it can answer them: named colleagues, home addresses, health, pay, disciplinary matters and children.
A pre-launch checklist
| Check before switching on an assistant | Why this case shows it matters |
|---|---|
| Review who can read what, and remove stale broad sharing | Access was Meta’s only filter, and old access became answers |
| Block suggestion topics: named people, locations, health, pay, children | The failure started with a suggestion, not a search |
| Test with dossier questions such as “Where does this colleague live?” | That is the question Meta’s system offered about Robins |
| Record a data protection impact assessment covering aggregation | Combined answers can create personal data no single file held |
| Confirm deleted files leave the search index | A photo Robins says she deleted appeared in results |
| Keep separate switches for suggestions and answers | Meta could change one without the other |
| Log the suggestions shown, not only the questions asked | Otherwise nobody can say who was nudged towards what |
A data protection impact assessment under Article 35 of UK GDPR is the natural place to record these decisions. If you want support with that assessment, see our data protection services.
What Parents and Creators Can Do About Meta AI Suggestions Now
Meta’s fix reduces the chance that Meta AI suggestions will offer “Who’s the child passenger?” again. It does not remove what the assistant can find. These steps address the part that is still in your hands.
Audit what relatives have posted
Start with other people’s pages, not yours. Search for your children’s names and look at birthday posts, birth announcements and school milestones shared by grandparents, aunts and friends. In this case that was where most of the information came from.
Change audiences before deleting
Some of Robins’ relatives cried while deleting their posts. Deleting is not the only option. Asking relatives to switch old posts from public to friends-only, or to remove children’s names from captions, keeps the memories while shrinking what a stranger’s assistant can reach. Robins’ suggestion was simpler still: save the photos privately first.
Remove tags and names from captions
A photo without a name is harder to connect. Review tagging settings, remove tags of children, and avoid naming children alongside schools, towns or dates of birth. Robins’ lesson was that each detail was harmless alone.
Use your data rights
In the UK and the European Union, you can ask Meta what personal data it holds about you through a subject access request under Article 15 of the GDPR. The ICO’s guidance explains how. You can also report specific Meta AI suggestions through the app. Screenshots with the date and time are what made Robins’ account checkable by reporters.
Questions Meta Has Not Answered About Meta AI Suggestions
Meta’s fix to Meta AI suggestions is welcome. It is also narrow, and the company has left the most important questions open.
| Open question | Why it matters |
|---|---|
| What counts as a “personal topic”? | The fix is only as wide as that definition |
| Did strangers watching the public Reel see the same Meta AI suggestions? | That decides whether this was a privacy scare or a disclosure |
| Will Meta AI still answer those questions if a user types them? | The statement defends the answers |
| How did a photo Robins says she deleted appear? | A deleted copy and a relative’s copy are very different problems |
| Which videos named the children? | Meta’s explanation and Robins’ account conflict |
| Did the feature pass the privacy review the FTC order requires? | The order covers every new or modified product practice |
| Does the fix apply on every Meta app and in every country? | Meta AI runs in four apps, and no scope was given |
| How are Meta AI suggestions generated? | If they reflect other people’s queries, strangers shaped them |
Meta’s family of apps reached 3.60 billion people a day on average in June 2026, according to its second-quarter results. Meta has not said how many of them see Meta AI suggestions under posts. That number, more than any single viral video, is the scale of this question.
Meta AI Suggestions and Personal Questions: FAQ
What did Meta AI suggest under Kalie Robins’ video?
Beneath a Facebook cross-post of Robins singing in the car with her daughter, Meta AI suggested “Who’s the child passenger?”. Later prompts, according to screenshots Futurism reviewed, included “What are Kalie’s daughter’s ages?” and “Where does Kalie Robins live?”, plus questions about the girls’ interests and her work.
What has Meta changed about Meta AI suggestions?
Meta says it has fixed the issue behind Meta AI suggestions related to personal topics. Its statement says the feature “never should have prompted the individual with questions like that.” It has not said that Meta AI will refuse to answer such questions if someone types them.
Can Meta AI find photos you have deleted?
Nobody has shown that. Robins says one result was a photo she deleted years ago, but a copy could have existed on a relative’s page. Meta’s statements do not mention deleted content, so the question remains open.
Is this the same problem as Meta’s Muse Image feature in July?
It is a similar pattern with a different outcome. In July, Meta removed a Muse Image feature that used public Instagram photos, saying it “missed the mark”. In September, Meta used the same words about Meta AI suggestions but kept the feature and fixed the prompts.
Do Meta AI suggestions affect people in the UK?
Meta has not said which countries the Meta AI suggestions appeared in or where the fix applies. UK readers can ask Meta what data it holds about them with a subject access request, and relatives can limit who can see older family posts.
How can parents limit what AI can find about their children?
Audit posts by relatives as well as your own, switch old public posts to a smaller audience, remove names and tags from photos of children, and avoid combining a child’s name with a school, town or birthday. Save photos privately before deleting anything.
References and Further Reading
The Verge: Meta says it’s changing AI suggestions after posing invasive personal questions
Instagram: Kalie Robins (@kontheinside) video about Meta AI
News18 via MSN: Woman claims Meta AI pulled out old info on her kids on Facebook
The Mirror: Urgent warning to parents after deleted photos used to build creepy profiles of children
The Mirror: Mum posts sweet video of child on Facebook but chill goes down spine at response
Meta Newsroom: Meet Your New Assistant: Meta AI, Built With Llama 3
Meta Newsroom: Introducing Muse Image, with the July 10 update
The Verge: Meta turns off the Instagram feature that let users make AI deepfakes of public accounts
TechCrunch: Meta removes controversial AI feature on Instagram after backlash
Meta Newsroom: Introducing Muse, a personal AI agent
Meta AI Research: How We Built Safety Into Muse
The Verge: Meta’s Muse AI works and creeps me out
TechCrunch: Meta debuts its Muse AI agent. Will consumers trust it?
CNBC: Meta pushes into personal AI agents as company faces public reckoning over privacy and safety
Meta Newsroom: Our Agreement With Bipartisan Attorneys General
FTC: FTC Proposes Blanket Prohibition Preventing Facebook from Monetizing Youth Data
FTC: Children’s Online Privacy Protection Rule (COPPA)
ICO: Age appropriate design code standards
legislation.gov.uk: Data (Use and Access) Act 2025, section 81
EUR-Lex: Digital Services Act, Regulation (EU) 2022/2065
Microsoft Learn: Data, Privacy, and Security for Microsoft 365 Copilot
Meta Investor Relations: Meta Reports Second Quarter 2026 Results
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.