I run a company in Singapore that sets up and runs the tech stack for businesses, and I use AI assistants every day. I also keep coming back to one worry about them, often enough that I decided to test it instead of repeating it over dinner.

Here's the worry. Facebook bought Instagram in 2012 and WhatsApp in 2014 [1]. Today the same company runs Meta AI inside those apps, and it announced that from 16 December 2025, what people say to Meta AI would be used to personalise the ads and content they see [2]. Sensitive topics are excluded. Reuters reported that the only opt-out is not using Meta AI, and that the EU, UK and South Korea were left out at first [3]. Meta also says WhatsApp conversations with Meta AI aren't used for ads unless that WhatsApp account is linked to your other Meta accounts [2]. I haven't seen independent confirmation that the rollout happened as announced, so read all of that as Meta's stated plan.

So my question is this: if I'm working through a breakthrough idea for my business inside Meta's assistant, or inside Gemini, wouldn't they know?

Behind it sit a few beliefs I've repeated for a while. Meta took over the ecosystem by buying its competitors and has data on everything. The service is free, but they steer how we think and how we find things out. Lock-in is becoming a recognised problem, and it'll hit a wall on integration and data hosting. The answer is to own our own data, and that's where the money is.

I had our team check each of those against what's documented as of November 2026. A couple I'd have bet on came back unsupported.

The honest case for the other side

Adoption is mass-market. Pew found 49% of US adults use AI chatbots, up from 33% in 2024, and 24% use one daily [4]. Gemini passed a billion monthly users in August 2026 [5].

Where the work is structured, the gains are real. In one large study, customer-support agents with an AI assistant resolved about 14% more issues per hour, and novice and low-skilled agents improved by 34% [6]. Developers across three field experiments completed 26% more tasks [7].

The null results are as solid, and the people selling AI rarely bring them up. A large Danish study of workers using AI chatbots rules out effects on earnings or hours larger than 2% [8]. Kenyan entrepreneurs given GPT-4 on WhatsApp showed no average effect. In the working paper's figures, the strong performers improved by just over 15% and the weak ones did about 8% worse [9]. METR found experienced open-source developers were 19% slower with AI while believing they had been 20% faster [10]. METR published a follow-up in February 2026 but called its own new data an unreliable signal, so I've left it out [11].

The point that hit my argument hardest: we found no study that isolates the value of an assistant's memory on real-world outcomes. The benchmarks are mixed, and one found that irrelevant memories interfere with how well an assistant understands what you want [12]. If memory is meant to be the lock, nobody has shown it's worth much yet.

People also switch. On Apptopia's numbers, ChatGPT's app market share fell from 69.1% to 45.3% between January 2025 and early 2026, while Gemini's rose from 14.7% to 25.2% [13]. Sensor Tower, counting unique users across app and web, had ChatGPT below half by March 2026 [14]. Challengers have made leaving easier. Anthropic launched a memory import for Claude in March 2026, which it labels experimental [15][16], and Google followed on 26 March with imports of memories and chat history into Gemini [17].

People do switch: AI app market share

ChatGPT
January 202569.1%
Early 202645.3%
Gemini
January 202514.7%
Early 202625.2%
Share of AI chatbot app usage. Source: Apptopia, reported by Fortune [13].

Argued at full strength, the other side says a business owner who builds his own AI stack to dodge lock-in is paying a hobbyist's tax for a problem most people fix by opening a different app. At the chat window, I think that case mostly wins. It gets weaker further down.

Claim by claim

My seven claims, checked against the record

  • Meta bought its way into owning the ecosystemPartly right
  • Meta has data on everythingPartly right
  • They steer how we get information and how we thinkProven capability, unproven practice
  • Lock-in is increasingly seen as a problemRight
  • Lock-in will hit a wall on integration and data hostingWrong, for now
  • If I have a breakthrough idea in their assistant, they'd knowPartly right
  • That's where the money isNot supported
Each verdict is argued below, with sources.

Meta bought its way into owning the ecosystem. Partly right. The intent is on the record. In February 2012, after his CFO cautioned him about buying competitors, Zuckerberg wrote that the Instagram deal was "a combination of" neutralising a potential competitor and integrating its product [18]. Press reports say he walked it back about 45 minutes later [19]. I'm not going to quote the first email without the walk-back.

On 18 November 2025, Judge Boasberg ruled that the FTC hadn't shown Meta holds monopoly power now [1]. He held that TikTok and YouTube belong in the market and wrote that "TikTok alone defeats the FTC's case". Because the case turned on where the market's boundaries sit, he never decided whether buying Instagram and WhatsApp was illegal. The court didn't bless the deals. It found the market had moved on. The FTC appealed on 20 January 2026 [20]. Both sides had filed their briefs by October, and as of early November we could find no argument date [21].

So I can say Meta bought its way to scale. I can't say a court found that it holds the ecosystem today.

Meta has data on everything. Partly right, and my version was too broad. WhatsApp messages are end-to-end encrypted, so the content of what you send isn't what regulators have been fighting over. The fights are about metadata, account data and business interactions. The European Commission fined Facebook €110 million in 2017 for telling merger reviewers in 2014 that it couldn't reliably match WhatsApp and Facebook accounts, and then doing it [22]. The gap is closing from another side, though. Chats with Meta AI aren't encrypted messages to friends, and on Facebook and Instagram they're what Meta said it would use for ads.

Meta has also closed its flagship model. Muse Spark launched on 8 April 2026 in the Meta AI app, with a private API preview and no open weights, though Meta says it hopes to open-source future versions [23]. Llama 4, in April 2025, was its last flagship open-weights release.

They steer how we get information and how we think. This turned out to be two claims, and the evidence treats them differently.

How we get information is measured. In Pew's browsing data, people clicked a traditional search result in 8% of visits when Google showed an AI summary, against 15% when it didn't, and they clicked a link inside the summary about 1% of the time [24]. Google disputes the study's method. Ads are moving into that layer too. OpenAI announced in January 2026 that it would test ads for its Free and Go users in the US while keeping paid tiers ad-free, and the test started in February [25][26]. Every vendor says ads don't influence answers, and none of those promises has been independently audited.

How often people click through to a website

Clicked a search result, no AI summary on the page15%
Clicked a search result, AI summary shown8%
Clicked a link inside the AI summary1%
Share of Google search visits. Source: Pew Research Center browsing data, 2025 [24]. Google disputes the method.

How we think is where my instinct ran ahead of the evidence. Meta's own studies during the 2020 US election found that switching people to a chronological feed, removing reshares or reducing like-minded content produced no measurable change in polarisation or political attitudes [27]. Meta cooperated on the design. Still, in those studies the feed didn't move minds much.

Conversation does. Personalised GPT-4 dialogues cut belief in conspiracy theories by about 20%, with the effect lasting two months [28]. Chatbots arguing for a candidate shifted vote preferences by roughly 1.5 to 4 points on a 100-point scale in the US, and by about 10 points in Canada and Poland [29]. A study of 76,977 people found the persuasion came from post-training and prompting more than from model size, and that the gains came with lower factual accuracy [30]. The participants knew they were in persuasion studies, and some of these figures come from the universities' summaries rather than the papers themselves.

What a conversation can do that a feed didn't

~20%drop in belief in conspiracy theories after a personalised GPT-4 dialogue, still there two months later [28]
1.5 to 4points of shift in US vote preference, on a 100-point scale, from chatbots arguing for a candidate [29]
~10points of shift in the same kind of study in Canada and Poland [29]
Participants knew they were in persuasion studies.

So an assistant can move people more than a feed can, and the lever sits with whoever writes its instructions. We found no evidence of Meta or Google deliberately steering beliefs at scale, though we didn't search for that specifically. Proven as a capability, then, and unproven as a practice.

Lock-in is increasingly seen as a problem. Right. It's the only one of my claims that came back fully supported. The Google search remedies cover generative AI [41], and in April 2026 the European Commission said it would further assess whether some AI services should be designated under its Digital Markets Act [31].

Lock-in will hit a wall on integration and data hosting. Wrong, on the current evidence. Integration is converging. MCP, the protocol for connecting tools and data to assistants, moved to the Linux Foundation's new Agentic AI Foundation in December 2025, with founding contributions from Anthropic, Block and OpenAI, and Google, Microsoft and AWS among its members [32]. A shared standard makes models easier to swap. That's the opposite of the wall I expected.

Regulation is slow, here too. Singapore's PDPA has a data portability obligation that was legislated in 2020, but it only takes effect once regulations are issued, and we couldn't find any [33]. So treat it as not yet in force, pending a check of the statute's commencement record. Either way, nothing in Singapore law today makes an AI vendor hand your accumulated context to a competitor.

If I have a breakthrough idea in their assistant, wouldn't they know? Partly right. On consumer tiers, they have it, and what happens next is a matter of their policy. Anthropic moved consumer Claude chats to opt-out training in August 2025, with up to five years' retention for people who allow training and 30 days for people who opt out [34]. A New York Times lawsuit froze OpenAI's deletion promise from mid-May to 26 September 2025, so even chats users had deleted were kept. Enterprise, Edu and zero-retention API customers weren't covered [35].

We found no case of an AI company mining one user's idea to go after a market. We didn't run a dedicated search, so "none found" doesn't mean "none exists". The realistic risks are duller than theft. Your thinking can be absorbed through training into what the model offers everyone, your ad profile can be inferred from what you typed, and your chats can turn up in someone else's lawsuit.

The protection is a contract. Enterprise and API tiers exclude training by default, and OpenAI states this on its own enterprise privacy page [36]. Get it written into your agreement anyway, along with retention terms.

That's where the money is. Not supported. More on that at the end, since it's the one I'm least able to judge fairly.

Where the lock-in sits

This changed my thinking most. The lock-in that matters isn't in the model or the memory. It's in defaults, bundles and operating systems.

Microsoft folded Copilot into consumer Microsoft 365 in January 2025 and raised the US price by US$3 a month, its first increase since launch. That works out to about 30% on the Family plan and 43% on Personal [37]. Amazon gives Alexa+ free to Prime members and charges US$19.99 a month to everyone else [38]. In January 2026 Apple and Google announced a multi-year deal under which Apple's next foundation models are based on Gemini [39]. Bloomberg reported the deal at around US$1 billion a year, and neither company has confirmed a figure [40]. For a lot of iPhone users, Apple lock-in and Google lock-in are now partly the same thing.

The courts haven't stopped it. The final judgment in the Google search case, entered on 5 December 2025, bars exclusive distribution deals, including ones covering Gemini, but Google can still pay for default placement that isn't exclusive [41]. Exclusivity is restricted. Paying for the default isn't.

Exclusivity is restricted. Paying for the default isn't.

This is uncomfortable for my own position: owning your data does nothing here. If Siri runs on Gemini, if Copilot comes inside the Office licence your staff already use, if Alexa arrives with Prime, then the assistant your customers and employees meet first is chosen by someone else.

So am I a naysayer?

Partly. I had the incentives right and the timeline wrong. The consumer defaults now include training on chats, ads informed by chats, multi-year retention and litigation holds, and those are fair things to worry about. But I expected the incumbents to hit a wall, and the evidence shows them standardising while the assistants get more useful.

Where I've landed is "rent the model, own the memory", with two amendments.

Rent the model, own the memory, and know who controls the door

Distribution: defaults, bundles, operating systemsSiri on Gemini, Copilot in Microsoft 365, Alexa+ with Prime. Chosen by someone else, and owning your data doesn't change it.
The model: rent itTurning into a commodity. Use a commercial tier with no-training and retention terms in the contract, and swap it when something better comes along.
Your memory: own itClients, decisions, how your business works. The one asset nothing is making portable, so keep it in a place you control.

Rent the model, because it's turning into a commodity. The price of reaching a given benchmark score falls 5 to 10 times a year, even as the newest frontier models get more expensive [42]. Open-weight models sit about four months behind the closed frontier on Epoch's index [43]. But in April 2026 all ten of the top open models on Artificial Analysis came from Chinese labs [44], and Meta has closed its flagship, so the open option depends on decisions none of us control.

Own the memory, because your accumulated context (clients, decisions, how your business works) is the one asset no market force is currently making portable. There's no open standard for AI memory. The import tools carry across whatever the old assistant decides to say about you, which loses detail and can't be checked.

First amendment: rent on a commercial tier with no-training and retention terms written into the contract, and keep high-value thinking off ad-funded consumer tiers. Second: owned data doesn't get you out of distribution lock-in, and anyone advising you should say so.

Owning the stack carries its own risk. Simon Willison describes a "lethal trifecta": an assistant that holds your private data, reads untrusted content and can send things out [45]. Every assistant wired into an inbox fits, including one you build yourself. OpenAI wrote in December 2025 that prompt injection is "unlikely to ever be fully 'solved'" [46]. Own the system and its security is your job.

What I'd do

For an individual:

  1. Do your serious thinking in an assistant, or a paid tier, that isn't ad-funded.
  2. On any consumer plan, open the data controls and decide about training yourself. Don't leave it to the default.
  3. Export your chat history every few months. It's a crude ZIP file, but you'll have a copy.
  4. Keep your important ideas in your own notes. The assistant is a workbench, and your filing cabinet should live elsewhere.
  5. Be wary of devices that only work while one company keeps its servers on. Humane's AI Pin stopped working ten days after its sale to HP was announced [47].

For a small business:

  1. Move anything involving client data or unreleased plans onto enterprise or API tiers, and get no-training and retention terms into the contract in writing.
  2. Keep company knowledge (SOPs, pricing logic, client history) as structured documents you host, connected to whichever model you use through a standard like MCP, so changing models doesn't mean rebuilding.
  3. List the assistants bundled into what you already pay for (Microsoft 365, Google Workspace, your staff's phones) and decide whether you want each one.
  4. Before you connect any assistant to your inbox, find out what it can send out and to whom.
  5. Don't count on the PDPA to make your AI data portable. For now, if you want portability, you build it.

What would change the picture: whether OpenAI or Apple ships a real import path or an open memory format, whether the EU designates an AI assistant under the DMA, and whether the PDPC brings Singapore's portability obligation into force. Any one of them would make an owned memory layer worth more than it is today.

A disclosure

LOOP builds owned-stack systems for clients, so I have a commercial interest in where this article ends up, and you should read it with that in mind. The claim most convenient for my business, that owned personal AI is where the money is, is also the one with the least evidence behind it. Our research found nothing on demand, market size or willingness to pay. Half of US adults use mainstream chatbots and people switch between them cheaply, which suggests the pain an owned layer relieves is smaller than I'd like. I think the case is plausible for businesses with sensitive data, regulatory exposure or workflows they want to keep to themselves. That's a hypothesis to test with paying clients. It isn't a finding, and I won't sell it as one.

Sources

  1. US District Court (D.D.C.), FTC v. Meta, memorandum opinion. https://www.ftc.gov/system/files/ftc_gov/pdf/MemorandumofOpinion.pdf
  2. Meta, Improving your recommendations with AI at Meta (1 October 2025). https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/
  3. Reuters via Inc., Meta will soon personalize ads based on what users type into its AI tools. https://www.inc.com/reuters/meta-will-soon-personalize-ads-based-on-what-users-type-into-its-ai-tools/91247113
  4. Pew Research Center, Americans and AI 2026 (17 June 2026). https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/
  5. Google, Gemini app passes 1 billion monthly users (11 August 2026). https://blog.google/innovation-and-ai/products/gemini-app/one-billion-monthly-users/
  6. Brynjolfsson, Li and Raymond, Generative AI at Work, NBER w31161. https://www.nber.org/papers/w31161
  7. Cui et al., developer field experiments, Management Science. https://pubsonline.informs.org/doi/fpi/10.1287/mnsc.2025.00535
  8. Humlum and Vestergaard, Large Language Models, Small Labor Market Effects, NBER w33777. https://www.nber.org/papers/w33777
  9. Otis et al., Kenyan entrepreneurs and GPT-4, Management Science. https://pubsonline.informs.org/doi/10.1287/mnsc.2024.06909
  10. METR, early-2025 AI and experienced open-source developer productivity (10 July 2025). https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
  11. METR, uplift update (24 February 2026). https://metr.org/blog/2026-02-24-uplift-update/
  12. RPEval (arXiv 2601.16621). https://arxiv.org/abs/2601.16621
  13. Fortune, ChatGPT app market share (5 February 2026). https://fortune.com/2026/02/05/chatgpt-openai-market-share-app-slip-google-rivals-close-the-gap/
  14. Sensor Tower, State of AI 2026 report (16 June 2026). https://sensortower.com/press/sensor-tower-state-of-ai-2026-report-global-time-spent-on-generative-ai-apps-projected-to-more-than-double-year-over-year
  15. Anthropic, Importing and exporting your memory. https://support.claude.com/en/articles/12123587-importing-and-exporting-your-memory-from-chatgpt-and-other-ai-providers
  16. SiliconANGLE, Anthropic makes switching from competitors easier (2 March 2026). https://siliconangle.com/2026/03/02/anthropic-makes-switching-competitors-easier
  17. Google, Switch to the Gemini app (26 March 2026). https://blog.google/innovation-and-ai/products/gemini-app/switch-to-gemini-app/
  18. FTC, substitute amended complaint, ECF No. 82. https://www.ftc.gov/system/files/documents/cases/2021-09-08_redacted_substitute_amended_complaint_ecf_no._82.pdf
  19. Global Competition Review, Zuckerberg saw Instagram as a competitive threat. https://globalcompetitionreview.com/gcr-usa/article/zuckerberg-saw-instagram-competitive-threat
  20. FTC, FTC appeals ruling in Meta monopolization case (20 January 2026). https://www.ftc.gov/news-events/news/press-releases/2026/01/ftc-appeals-ruling-meta-monopolization-case
  21. Justia, D.C. Circuit docket 26-5028. https://dockets.justia.com/docket/circuit-courts/cadc/26-5028
  22. European Commission, press release IP/17/1369 (18 May 2017). https://ec.europa.eu/commission/presscorner/detail/en/IP_17_1369
  23. Meta AI, Introducing Muse Spark (8 April 2026). https://ai.meta.com/blog/introducing-muse-spark-msl/
  24. Pew Research Center, clicks and AI summaries (22 July 2025). https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
  25. OpenAI, Our approach to advertising and expanding access. https://openai.com/index/our-approach-to-advertising-and-expanding-access/
  26. OpenAI, Testing ads in ChatGPT. https://openai.com/index/testing-ads-in-chatgpt/
  27. NYU CSMaP, US 2020 Facebook and Instagram election studies. https://csmapnyu.org/impact/news/first-four-papers-from-us-2020-facebook-instagram-research-election-study-published-in-science-and-nature
  28. Cornell, AI dialogues and conspiracy belief. https://news.cornell.edu/node/329310
  29. Cornell, AI chatbots can effectively sway voters in either direction (December 2025). https://www.news.cornell.edu/stories/2025/12/ai-chatbots-can-effectively-sway-voters-either-direction
  30. Hackenburg et al., The levers of political persuasion with conversational AI (arXiv 2507.13919). https://arxiv.org/abs/2507.13919
  31. European Commission, DMA review report COM(2026) 178 (28 April 2026). https://digital-markets-act.ec.europa.eu/system/files/2026-04/DMA%20Review%20Report_COM_2026_178_1_EN.pdf
  32. Linux Foundation, formation of the Agentic AI Foundation (9 December 2025). https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation
  33. Allen & Gledhill, PDPC on data portability. https://www.allenandgledhill.com/sg/publication/articles/14292/pdpc-issues-response-to-public-consultation-on-data-portability-and-data-innovation-provisions
  34. Anthropic, Updates to our consumer terms (August 2025). https://www.anthropic.com/news/updates-to-our-consumer-terms
  35. OpenAI, Response to New York Times data demands. https://openai.com/index/response-to-nyt-data-demands/
  36. OpenAI, Enterprise privacy. https://openai.com/enterprise-privacy/
  37. Microsoft, Copilot is now included in Microsoft 365 Personal and Family (16 January 2025). https://www.microsoft.com/en-us/microsoft-365/blog/2025/01/16/copilot-is-now-included-in-microsoft-365-personal-and-family/
  38. Amazon, Alexa+ available free for Prime members. https://www.aboutamazon.com/news/devices/alexa-plus-available-free-prime-members-us
  39. Google, Joint statement from Google and Apple (January 2026). https://blog.google/company-news/inside-google/company-announcements/joint-statement-google-apple/
  40. Android Authority, Apple and Google Gemini deal (reporting Bloomberg). https://www.androidauthority.com/apple-siri-google-gemini-deal-3631890
  41. US Department of Justice, United States v. Google, final judgment (5 December 2025). https://www.justice.gov/atr/media/1421546/dl?inline=
  42. Inference price trends (arXiv 2511.23455). https://arxiv.org/abs/2511.23455
  43. Epoch AI, open vs closed model gap (29 May 2026). https://epoch.ai/data-insights/open-closed-eci-gap
  44. Artificial Analysis, recent open-weights model launches (30 April 2026). https://artificialanalysis.ai/articles/recent-open-weights-model-launches
  45. Simon Willison, The lethal trifecta (16 June 2025). https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/
  46. OpenAI, Hardening Atlas against prompt injection (22 December 2025). https://openai.com/index/hardening-atlas-against-prompt-injection/
  47. Humane, Ai Pin consumers FAQ. https://support.humane.com/hc/en-us/articles/34243204841997-Ai-Pin-Consumers-FAQ