Secret Chat Logo

Blog

Launch Private Chat

Articles

  • GPT 6 Astra on Secret Chat AI: The New Manual Mode

    Published September 5, 2026

    GPT 6 Astra is now available as a manual ChatGPT mode on Secret Chat AI. What OpenAI's most capable model costs, why Smart Agentic will never select it by accident, when it earns the 2.5-times price, and the cybersecurity restrictions to know before spending credits.

  • Gemini 3.8 Flash on Secret Chat AI: Newer Model, Same Price, Still No Pro

    Published September 3, 2026

    Google's Gemini 3.8 Flash now answers Gemini Web Search and Research on Secret Chat AI, three weeks after 3.7 Flash. What Google says changed, what we measured (it thinks a little more, not less), what it costs you, when Google's introductory price ends, and why the Pro tier is still missing.

  • Claude Fable 5.1 on Secret Chat AI: What Changed

    Published September 2, 2026

    Secret Chat AI's Claude Fable mode now runs Claude Fable 5.1, Anthropic's most capable model. What changed on the wire (almost nothing), the one price that moved and why long chats got cheaper, why Smart Agentic Claude still runs on Opus, how to pick the mode, and the 30-day retention Anthropic applies to this tier.

  • Why Do AI Models Give Different Answers to the Same Question?

    Published August 30, 2026

    Ask two chatbots one question and you get two answers. The five reasons they diverge, why disagreement is information rather than failure, and how to read it.

  • The Case Against AI Memory: Why Forgetting Is a Feature

    Published August 28, 2026

    The big assistants now want to remember you — ChatGPT, Gemini, Claude and Grok alike — and the pitch is convenience. The mechanism is a stored, identity-linked profile that accumulates on its own — and a stored record is a liability that outlives intentions: in the NYT suit a legal hold overrode ChatGPT's delete button for more than four months, a 20-million-conversation sample went to the plaintiffs with no individual notice to the people in it, and the publishers now allege billions of outputs were deleted regardless. Why context and history are not the same thing (a paragraph sent with each request does the useful part; the permanent archive serves the company), why the industry builds memory anyway (it is the moat), what an AI designed to forget looks like — and the honest limits of forgetting, which protects you from association, not from the words you type.

  • Anonymous AI Is Becoming a Right, Not a Preference

    Published August 28, 2026

    Using AI without being identified used to be filed under personal taste. In eighteen months the category changed: OpenAI's CEO publicly asked for an "AI privilege", Arizona passed a bill creating one through its House before it died in the Senate, the Supreme Court held in Chatrie that data stored on a company's servers can still carry a reasonable expectation of privacy, a New York court quashed a subpoena served on OpenAI for a litigant's prompts, the EDPB wrote the test that decides whether "anonymous" is true, and Germany's TDDDG already obliges every digital service to enable anonymous use where feasible — on top of a UN report that called anonymity a condition of free expression a decade before chatbots. What is decided, what is only demanded, what is still missing, and why the difference between a right and a preference is who carries the burden.

  • AI Superapps Want to Be Your Everything — That's Exactly the Problem

    Published August 27, 2026

    In twelve months the chat box grew checkout, in-chat apps, a browser, proactive briefings and ads (OpenAI), started replacing Google Assistant on most Android phones and watches (Google), merged with a social network outright (xAI), and banned rival chatbots from WhatsApp — until EU regulators ordered the door reopened in Europe (Meta). Why the everything-app pitch is a concentration problem — one identified account holding what you ask, read, buy and say at home; a profile that is both switching cost and targeting asset; features that die (Atlas, Pulse) while the data practices persist; walls that remove your choice of model — and how to keep the capability while declining the consolidation.

  • Why "Free" AI Chatbots Are the New "Free" Social Media

    Published August 26, 2026

    Free chatbots are re-running the social media playbook, a decade faster: training on your conversations by default (now the norm at every major consumer chatbot), ads already live — ChatGPT's Free and Go tiers since February 2026, Meta AI chats feeding ad targeting with no opt-out since December 2025, ads in Grok announced — and the first engagement-tuning failure already shipped and rolled back. Why chat transcripts cut deeper than posts ever did (you performed for social media; you confess to a chatbot), what is genuinely different this time in your favor, and how to take an exit before the habits are a decade deep.

  • The Coming AI Privacy Reckoning: Predictions for 2027

    Published August 25, 2026

    2027 is when AI privacy law stops being written and starts being enforced. Seven grounded predictions: the AI Act's first Article 50 test cases and its December 2027 high-risk deadline, the first appellate ruling on AI chats in litigation (counsel direction, not civil-vs-criminal, is the emerging line), "which AI tools did you use?" as boilerplate discovery, the EDPB's new anonymisation test meeting "anonymized" marketing claims, the erasure-vs-weights standoff migrating to Ireland, the Latombe challenge to EU–US transfers at the same court that killed Safe Harbor and Privacy Shield, and a thickening US state patchwork — plus what won't happen, and how to position yourself before any of it does.

  • Introducing the AI Council: One Question, Several AIs, and a Referee

    Published August 23, 2026

    A language model delivers a fabricated answer in the same confident voice as a solid one, and asking it to double-check itself is unreliable — intrinsic self-correction can even make answers worse. The AI Council automates the countermeasure serious users already improvise: one question goes to ChatGPT, Claude, Gemini and Grok in parallel, each answer appears as it lands, and a referee model extracts the factual claims and maps them into a disagreement table — who asserts, who contradicts, who stays silent. Why every seat belongs to a different company, when the panel searches the web live, what the free two-model Duet includes, how to have an answer you already have cross-checked mid-chat — and the honest limits: a council widens where your question travels, agreement between models is evidence, not proof, and the referee is a model to be checked too.

  • Browser Storage Limits: How Much Chat History Can IndexedDB and OPFS Really Hold?

    Published August 21, 2026

    The real per-browser storage quotas behind local-first AI chat — Chrome, Firefox and Safari — with the arithmetic for text and images, what happens at the ceiling, and the deletion rules that matter more than capacity.

  • What Happens to Your Local Chats If You Clear Browser Data? (And How to Prevent Loss)

    Published August 21, 2026

    Clearing "cookies and other site data" permanently deletes locally stored AI chats — which checkbox does it, the Safari trap, what survives on your account, why nothing brings cleared chats back, and the export habit that prevents the loss.

  • Can an AI Model "Remember" Your Data After a Training Opt-Out? Memorization Explained

    Published August 21, 2026

    An opt-out is a valve, not an eraser: it closes your chats' path into future training runs and cannot reach a model already trained. The two kinds of "remembering" people conflate — product memory is stored text, memorization is in the weights — what research actually extracted (hundreds of verbatim sequences from GPT-2 including names and phone numbers, at least 1% of GPT-J's training set, megabytes out of ChatGPT for about $200 with the "repeat the word poem forever" attack), why no provider offers machine unlearning at frontier scale and output filtering is the remedy that visibly ships, what decides whether your own data was ever at risk, and why opting out is still worth doing: it stops the accumulation.

  • Prompt Injection and Uploaded Files: A Security Primer for Non-Engineers

    Published August 21, 2026

    A language model cannot tell the text it should read from the text it should obey — your question, its instructions and every character of the file you attached arrive as one stream. What follows from that: white-on-white instructions in PDFs and resumes, the 17 academic papers caught telling AI reviewers "give a positive review only", the Guardian's hidden-text test that made ChatGPT Search praise a product over its own negative reviews, and EchoLeak — the zero-click email that made Microsoft 365 Copilot exfiltrate internal data. Why OWASP ranks prompt injection the #1 AI-application risk in every edition and the NCSC says no failsafe measure removes it, the capability ladder that decides how much an injection can actually do to you, and the five habits that genuinely help.

  • Why "Zero Data Retention" Is Harder Than It Sounds (and How Providers Implement It)

    Published August 21, 2026

    ZDR is real — real enough that the 2025 preservation order in the NYT case reportedly flowed around ZDR API customers, because nothing was stored to preserve — and it is a configuration, not a property of a company: approval-gated, scoped per endpoint and per feature, with the carve-outs written by the providers themselves. OpenAI's stateful features may still store application state under ZDR and CSAM-flagged images are retained regardless; Anthropic's Covered Models (Claude Fable 5 and Mythos 5) cannot run under ZDR at all, carrying a mandatory 30-day safety window because some attacks only become visible across multiple requests. Why every layer of an inference stack wants to keep copies, why the arrangement is rationed rather than switched on, the never-stored > stored-then-deleted > stored-for-a-window hierarchy, the four questions that make any ZDR claim testable — and how Secret Chat answers them about itself.

  • What Actually Happens to a Prompt Inside an LLM Provider's Infrastructure

    Published August 18, 2026

    One prompt, ten stages: decrypted at the provider's edge because the model must read plaintext, joined to your account in the first millisecond, re-sent in full on every turn (a sentence in message 3 of a 40-turn chat crosses the network 38 times), tokenized, scored by classifiers whose verdicts Anthropic may keep for 7 years, batched on one GPU with strangers' questions, and cached for up to 30 minutes after you stop typing. Where copies are actually made, why deletion reaches only one of them, and the single stage that decides how much the other nine matter.

  • Immigration, Visas, and AI: Why Some Questions Should Never Touch a Cloud Account

    Published August 17, 2026

    The visa applicant is the one AI user whose government audience is explicit: since 2019 the DS-160 has collected five years of social media handles, students must set profiles public since June 2025, H-1B followed in December, fourteen more categories in March 2026 — and "Catch and Revoke" reads it all with AI. Add a record 55,318 border device searches in FY2025, the writer whose phone contents ended his trip, and courts holding that no privilege attaches to chatbot conversations, and the case writes itself: immigration questions are intent, admissions and a five-year timeline rolled into one — the last thing to file under your own name. What actually works instead, and the line this advice will not cross.

  • Your Kids Are Talking to Chatbots: A Parent's Guide to AI Privacy

    Published August 16, 2026

    72% of US teens have used an AI companion, and the most sensitive conversations your child has may not be happening with a person. What a chatbot conversation actually records, the age gates as they really stand after Character.AI's under-18 ban and OpenAI's age prediction, the four documented failures that forced the current safety rules — Meta's leaked guidelines, two lawsuits, and a teddy bear — why COPPA stops at the thirteenth birthday, what parental controls deliberately don't show you, the house rules that transfer to every app, and the record about your kids growing quietly in your own AI account.

  • Should You Tell an AI Your Secrets? A Realistic Threat Model

    Published August 15, 2026

    Most AI privacy advice is either "never tell it anything" or "relax, it's encrypted", and both are useless. A ranked threat model instead: who realistically reads your chats, from the person with your unlocked laptop and the employer who administers your account down to the state actor almost nobody faces; the LeakyLM audit finding verbatim message text reaching TikTok; 300 million messages left publicly readable by one misconfigured app; the 20 million ChatGPT conversations pulled into a lawsuit their users were never told about; why the fear people actually have sits at the bottom of the list; and exactly which rows each defence covers.

  • OpenAI Reported a ChatGPT User to the FBI: What the Palm Beach Case Actually Proves

    Published August 15, 2026

    A Florida financial analyst spent weeks allegedly describing a murder plan to ChatGPT; OpenAI reviewed the messages and alerted the FBI, and federal agents handed two months of his chat logs to the county sheriff. He pleaded guilty in August 2026 with adjudication withheld. Nothing leaked — this is OpenAI's documented process: classifiers always running, human reviewers one score away, and a referral threshold the provider sets alone. Why the reporting was probably right, why that is exactly the uncomfortable part, and what a gateway that keeps no transcript can and cannot offer instead.

  • Gemini 3.7 Flash Is Here — and Gemini 3.5 Pro Never Was

    Published August 14, 2026

    Google's Gemini 3.7 Flash powered Gemini Web Search and Research on Secret Chat from August 14 to September 3, 2026, when Gemini 3.8 Flash replaced it. Meanwhile no new Pro tier has shipped since January — Google's own latest-Pro alias still resolves to a January preview build.

  • Talking to AI About Money, Debt and Divorce: Keeping It Off the Record

    Published August 14, 2026

    In G v K [2025] EWHC 2961 (Fam) a husband produced screenshots of his wife's ChatGPT account in the High Court — no subpoena, no breach, just the password. Why money, debt and divorce prompts behave unlike anything else you type; the one category of distress that sits outside the sensitive-topic fence OpenAI's ad rules draw around health, mental health and politics; what privilege and work product really cover; and the six things that actually keep the conversation off the record.

  • Grok 4.6 Is Here: xAI's Newest Model, Now on Secret Chat

    Published August 13, 2026

    xAI's Grok 4.6 now powers the whole Smart Grok family on Secret Chat — Smart Agentic, Web Search and Deep Research. What changed since Grok 4.5, what didn't, and how anonymous access works.

  • AI and Your Job Search: Why You Shouldn't Draft Resignation Letters in Your Work ChatGPT

    Published August 12, 2026

    A resignation letter written in a work AI account is a dated, first-person statement of intent to leave, held by the party whose interests are opposed to yours. What each plan actually exposes — the Enterprise compliance export wired into litigation tooling, what Business genuinely does not share, Copilot prompts stored in your mailbox, Vault holds on Gemini — why a legal hold beats your delete button while showing you an empty history, why neither privilege nor work product is likely to cover the draft, and the complete move that does work.

  • Grok Imagine 2.0 Is Here: xAI's New Image Model, Now the Default in Our Image Generator

    Published August 12, 2026

    xAI shipped Grok Imagine Image 2.0 on August 7, 2026, and it is now the model the Secret Chat Image Generator opens on — ahead of Nano Banana 2. What the new model is, what xAI claims for it, why we moved it to the front, what you can control (aspect ratio, 1K/2K, up to ten variants, one-click editing), why your own saved choice is untouched, and how Smart Agentic Grok now draws with it too.

  • Using AI as a Therapist Substitute: The Data Risks of Emotional Conversations

    Published August 11, 2026

    OpenAI's own figures put more than a million people a week in conversations carrying explicit indicators of potential suicidal planning or intent, and nearly three in four US teens have used an AI companion. Therapy is confidential because of a stack of legal instruments a chat window has none of — an evidentiary privilege that needs a licensed clinician, and HIPAA's separate rule for psychotherapy notes. What a chat log records that a clinical file never does, why the safety features are also the inspection features, the three afterlives of an emotional transcript, and how to get the value without building the record.

  • Asking AI About Health Symptoms: Why Privacy Matters More Than You Think

    Published August 10, 2026

    A third of US adults now ask AI chatbots about their health, and three-quarters of them say they are worried about it — a knowing trade made at one in the morning. Why HIPAA does not cover the conversation at all, what changed when ChatGPT Health began linking patient portals to a consumer account in 2026, the four paths by which a symptom question comes back months later, why de-identification is weaker for health data than for anything else, and how to ask so you get the answer without handing over a medical record.

  • Using AI on a Shared or Work Computer: What Your IT Department Can See

    Published August 9, 2026

    The privacy of an AI conversation can never be better than the privacy of the machine you type it on. A layer-by-layer inventory of what is visible on a managed device — network destinations and TLS interception, enterprise browser reporting, endpoint policies that can be configured to log the prompt itself as searchable text, the admin console above a corporate AI account, and what stays on the disk. Plus the notice laws that mean you were probably already told, why Incognito and a VPN solve none of it, and the one move that actually works.

  • The Privacy Settings Checklist: 10 Toggles to Flip in Every AI App You Use

    Published August 9, 2026

    A default is a company's preference expressed as your configuration, and in AI apps the gap is unusually wide. The ten settings that repeat across every major assistant — training, memory, the retention window, temporary chat, the new ad-personalization controls, live share links, connectors, the separate voice toggle everyone misses, phone permissions and account hygiene — with where each one hides today, and the honest limit: a configured account is a quieter account, not an anonymous one.

  • How to Export, Back Up, and Migrate Your Local AI Chat History Safely

    Published August 8, 2026

    When chat history lives in your browser, the export is not a spare copy — it is the only other instance of your data, and nobody can restore it for you. What actually destroys local history (five of the six causes are things you do deliberately), why an unencrypted export is the most dangerous file you own, how to prove a backup works before you need it, and how to merge two machines without overwriting recent work.

  • Local LLMs (Ollama, LM Studio) vs a Private Gateway: An Honest Speed, Quality and Privacy Test

    Published August 6, 2026

    An offline model transmits nothing — that part is real, and we say so. What it cannot do is match GPT, Claude or Gemini on your laptop: the open-weight models that rival the frontier run to 2.8 trillion parameters and roughly 1.5 TB of weights, and reaching that quality on your own hardware means a multi-node cluster, not a better machine. Measured numbers on both walls (a 671B model took 14.8 minutes to read an 8,000-token prompt), why 2026 made the hardware scarcer, and the local-plus-gateway split that actually works.

  • EXIF and Metadata in Uploaded Files: What AI Providers See That You Don't

    Published August 5, 2026

    Cropping a photo does not remove its GPS coordinates, and redacting a PDF does not touch the author field. What Exif, IPTC, XMP and document properties actually record, why "the model said it cannot see metadata" is not the same as "it was not uploaded", why stripping is now necessary but no longer sufficient — models read locations out of pixels — and how to clean a file before it travels.

  • Incognito Mode Won't Protect Your AI Chats: Here's What Actually Does

    Published August 3, 2026

    Private browsing is a setting on your own computer, and almost nothing that makes an AI conversation risky happens on your own computer. What Incognito genuinely covers, why signing in undoes it, the perverse case where it leaves you worse off — ChatGPT's signed-out training opt-out is a cookie, and clearing cookies is the whole point of the mode — what Google conceded in the Brown v. Google settlement, and the four layers that actually need covering.

  • VPN + AI Chatbot: What It Protects and What It Doesn't

    Published August 1, 2026

    A VPN changes the route your words take — not what they are, who is logged in when you send them, or what happens after they arrive. What the tunnel genuinely covers, why the account makes the IP the weakest identifier in the stack, what your VPN provider can still see (and what PureVPN's records and Mullvad's empty servers showed about "no logs"), why geo-switching is a terms question rather than a privacy one — and how a VPN and an anonymizing gateway stack at opposite ends of the same request.

  • Consumer ChatGPT vs API Access: Why the Same Model Has Two Privacy Regimes

    Published July 31, 2026

    The weights are identical; the wrapper is not. Consumer ChatGPT trains on your chats by default, remembers you across conversations and hangs everything on a named account — while the same model on the API has trained on nothing since March 2023, is stateless unless a developer asks otherwise, and comes with a processor contract. Where the line really falls, why "just use the API" overstates it, and what the 2025 New York Times preservation order proved when it reached consumer chats and standard API traffic alike.

  • AI Data Residency: Where Your Prompts Physically Live and Why It Matters

    Published July 29, 2026

    Residency, sovereignty and localisation are three different things, and "EU data residency" usually means storage at rest — not where the model ran. OpenAI has ten residency regions but only three where inference runs in-region (the US and Europe from January 2026, the UAE from August 2026), Anthropic stores in the US while routing globally, and the CLOUD Act reaches a US provider regardless of the server's country. What residency covers, what it never answers, and why consumer plans get no region choice at all.

  • DeepSeek Without the Chinese Servers: Why Open Weights Change the Whole Question

    Published July 29, 2026

    DeepSeek is now the free model on Secret Chat — and it never touches Chinese infrastructure. Because DeepSeek publishes its weights, the same model runs on around twenty independent providers; every request here carries a US-only provider allowlist that bounds every routing attempt, and no-logging endpoints only. What that fixes, what it honestly does not — including why open weights move the hosting but not the training, so the model's own refusals travel with it — and what the app's one DeepSeek entry can and cannot do.

  • What Your Prompts Reveal About You: The Metadata Problem Nobody Talks About

    Published July 28, 2026

    Every AI conversation leaves a second record beside it. Timestamps, cadence, model and mode choices, token counts, auto-generated chat titles, safety classification scores Anthropic keeps for up to 7 years, and the author fields hidden inside the files you upload — what the envelope around your prompts gives away, which parts deletion does not reach, and why metadata is inert until it accumulates against a name.

  • SOC 2, BAAs, and DPAs: The Compliance Alphabet for AI Tools, Explained Simply

    Published July 27, 2026

    SOC 2 is an auditor's report, a BAA is a US health-data contract, a DPA is a GDPR processor contract — and none of them means what buyers assume. What each covers, why AI coverage attaches to endpoints rather than brands, and the ten questions to ask before you send anything sensitive.

  • Attorney-Client Privilege and AI: Has Anyone Actually Waived It Yet?

    Published July 26, 2026

    In February 2026 the hypothetical got case law: United States v. Heppner found a defendant’s 31 Claude-drafted defense documents were never protected — protection never attached, which is not the same as waiver — the same day Warner v. Gilbarco protected a litigant’s ChatGPT research as work product. Four more civil courts have since sided with Warner (Morgan, Tym, Tate and Assini), making it five to one with no appellate ruling. Where the law stands, why England’s Upper Tribunal has already used the word waive, and what to do about it.

  • Claude Opus 5 Is Here: Anthropic's New Flagship Opus, Now on Secret Chat

    Published July 25, 2026

    Claude Opus 5 has replaced Opus 4.8 on Secret Chat — the tier Smart Agentic Claude escalates to and the Research mode both run on it now. What changed, what stayed fast, and how anonymous access works.

  • The EU AI Act and Your Chat Data: A Plain-English Guide

    Published July 24, 2026

    From August 2, 2026 every chatbot serving EU users must tell you it's AI, and AI-generated content must be marked — but the AI Act gives you almost no rights over your chat data. The timeline after the Digital Omnibus delay, what was already banned, the public training-data summaries, and where the GDPR takes over.

  • GDPR and AI Chatbots: What "Right to Erasure" Means When You Prompt an LLM

    Published July 23, 2026

    The GDPR right to erasure reaches your stored chats — but not a trained model's weights, which today can only be filtered, not wiped. What EDPB Opinion 28/2024 and the 2026 anonymisation guidelines say, why Italy's €15M OpenAI fine was annulled in March 2026 on jurisdiction alone, what an erasure request actually gets you, and how to protect yourself before you send.

  • AI Memory Features Are a Privacy Time Bomb: What ChatGPT and Gemini Remember About You

    Published July 22, 2026

    ChatGPT and Gemini now carry a profile of you between chats — on by default in Gemini, and in ChatGPT synthesised in the background since the June 2026 memory rebuild rather than curated by you. It feeds ad targeting on the Free and Go tiers, which reached five more countries in August 2026. What each one stores, why deleting a chat does not delete the memory, how researchers poisoned it with prompt injection (and why only one of those two attacks was fixed), and what to do about it.

  • The Hidden Ways AI Chatbots Fingerprint You Even Without an Account

    Published July 21, 2026

    Logging out is not anonymity. IP addresses, browser fingerprints, storage identifiers, analytics vendors — and, distinctively with AI, the text itself, which research shows can reveal location, age and income. With the scale caveat the scare stories omit: picking one author out of 100,000 by writing style succeeded just over 20% of the time, and stylometry corroborated the leak that exposed J.K. Rowling rather than originating it. What each signal really gives away, and what actually helps.

  • "Temporary Chat" Isn't What You Think: The Fine Print Nobody Reads

    Published July 20, 2026

    ChatGPT, Gemini, Claude, Grok and Perplexity all offer a temporary mode — and all keep the chat somewhere between 24 hours and 30 days. What the fine print says: Gemini still allows human review inside a temporary chat, feedback puts one back in training scope, Claude's incognito chats appear in an employer's data export, and the identity link no toggle removes.

  • What Happens to Your AI Chats When a Provider Gets Breached?

    Published July 19, 2026

    Real AI breaches — ChatGPT's 2023 caching bug, DeepSeek's open database, the Muah.ai dump, the Firebase misconfiguration that left about 300 million Chat & Ask AI messages from 25 million users readable without a password, and indexed share links — show exactly what happens to stored chats when security fails, and how to keep your identity out of the next leak.

  • AI Chat Logs in Divorce and Employment Lawsuits: Real Cases, Real Consequences

    Published July 15, 2026

    Real cases show AI chat logs surfacing in divorce and employment disputes. What United States v. Heppner actually held — privilege never attached, and for reasons distinct from the work-product ruling — plus the four courts that have since protected AI chats as work product, the English case where a father produced screenshots of the mother's own ChatGPT prompts, and Trinidad v. OpenAI, where developing a trade secret through ChatGPT destroyed the claim. What is genuinely exposed, what is not, and how to reduce your risk.

  • Grok 4.5 Is Here: xAI's Newest Reasoning Model, Now on Secret Chat

    Published July 15, 2026

    The Grok 4.5 release on Secret Chat — Smart Agentic, Web Search, Deep Research and image generation. Superseded by Grok 4.6 on August 13, 2026; kept for reference on how the Smart Grok modes work and why the free quick tier needs a non-reasoning model.

  • Your AI Chats Can Be Subpoenaed: What Court Cases Have Already Shown

    Published July 14, 2026

    Real cases show AI chats reaching court. OpenAI was ordered to produce a de-identified sample of 20 million consumer ChatGPT conversations from December 2022 to November 2024 — a different set from the 2025 preservation hold, which stopped applying to new data on 26 September 2025 — and in Assini v. Hayward a subpoena went to OpenAI itself and was quashed. The preservation/production distinction, the narrow work-product exception four courts have recognised, and how to reduce your exposure.

  • Secret Chat AI Now Runs GPT-5.6: Meet Sol, Terra, and Luna

    Published July 11, 2026

    We've migrated Secret Chat AI to OpenAI's new GPT-5.6 family — Sol, Terra, and Luna. What each tier does, and how you get them fully anonymized.

  • Does Grok Save Your Chats? xAI's Privacy Policy Decoded

    Published July 11, 2026

    Does Grok save your chats? xAI trains on them unless you opt out, your public X posts feed the same model through a separate buried toggle, and the August 2025 share-link episode put roughly 370,000 conversations into Google, Bing and DuckDuckGo. What xAI actually keeps, what Private Chat does and does not promise, and how the API and Zero Data Retention differ.

  • Does Perplexity Track You? Search History and AI Privacy

    Published July 11, 2026

    Does Perplexity track you? On the consumer product your searches are stored and used to improve its models by default, and Memory and Brain are built to accumulate context about you — but the advertising programme that defined its reputation was phased out, and in February 2026 Perplexity said it has no plans to resume it. What is actually kept, what Incognito's 24-hour threads and Comet's local-by-default design change, and safer options.

  • How to Opt Out of AI Training on Every Major Platform (2026 Master Guide)

    Published July 11, 2026

    Opt out of AI training on ChatGPT, Claude, Gemini, Grok, Perplexity, Meta, Copilot, LinkedIn, GitHub, Mistral Vibe (renamed from Le Chat), DeepSeek and Character.AI — the current steps for each, plus what opting out does not cover: Anthropic still uses safety-flagged and self-reported material, Google still applies human review for safety with Keep Activity off, and Meta offers a statutory objection only in the EU, UK and Brazil.

  • Does Gemini Read Your Data? What Google Actually Keeps

    Published July 10, 2026

    Does Gemini read your chats? On the consumer app a subset go to human reviewers, training is on by default, and activity is kept 18 months — with human-reviewed chats held up to three years and not deleted when you clear your activity. Plus what turning Keep Activity off does not stop (Google still uses chats for responding and safety, with human help), the EEA/Switzerland/UK exception that gives free API users paid-tier treatment, and what Personal Intelligence and default-on memory add to the question.

  • Does Claude Store Your Conversations? Anthropic's Data Retention Explained

    Published July 9, 2026

    Does Claude save your chats? How Anthropic's data retention really works — the August 2025 training change, and why the widely quoted "30 days or 5 years" is misleading: your history lasts until you delete it, the 30 days is the back-end purge after deletion, and the 5 years covers a de-identified copy in the training pipeline. Plus the opt-out's written exceptions for safety-flagged and self-reported material, memory entries that outlive the chats they came from, Covered Models that cannot use Zero Data Retention, and how the business API differs.

  • How to Use ChatGPT Anonymously: No Account, No Saved History

    Published July 6, 2026

    Five real ways to use ChatGPT (or GPT models) anonymously in 2026: no-login ChatGPT, Temporary Chat, data controls, Duck.ai, and Secret Chat AI — with what each does and does not hide.

  • Secret Chat AI Subscriptions: Save Up to 25% on AI Credits

    Published July 4, 2026

    Secret Chat AI now has subscriptions — Basic, Regular and Pro plans with monthly or annual billing, up to 25% off, credits that refill every month, upgrade-anytime proration, and no lock-in.

  • Claude Fable 5 Is Here: Anthropic's Most Capable Model, Now on Secret Chat

    Published July 3, 2026

    Claude Fable 5 is now available on Secret Chat as a manual Claude mode. What Fable is, where it came from, why it costs more, and why Smart Agentic Claude still runs on Opus.

  • Secure AI Gateway Explained: A Practical Alternative to Self-Hosting

    Published July 1, 2026

    What a secure AI gateway is, how it compares to self-hosting an AI model, and when each one fits — with the real privacy trade-offs explained in plain terms.

  • Private AI for Business Owners: Protecting Customer and Company Data

    Published June 23, 2026

    How business owners can use AI more safely with customer and company data: what private AI protects, what it cannot, and the safeguards every team should keep.

  • Is ChatGPT HIPAA Compliant? A Guide for Therapists and Coaches

    Published June 22, 2026

    Standard ChatGPT is not HIPAA compliant: the OpenAI BAA covers enterprise and clinician products, and therapists and coaches are not eligible for the clinician route. What that means for client notes, including the separate status of psychotherapy notes.

  • Private AI for Journalists: Protecting Sources and Notes

    Published June 21, 2026

    How reporters can use AI more safely with sensitive notes and source material: what private AI protects, what it cannot, and the habits that keep sources safe.

  • Can Lawyers Use ChatGPT? A Confidentiality and Privilege Guide

    Published June 20, 2026

    Can lawyers use ChatGPT with client information? What confidentiality and privilege require, where consumer AI falls short, and the safeguards that make AI use defensible.

  • Multi-Model AI Chat: Why Use GPT, Claude, Gemini, Grok and Perplexity Together?

    Published June 19, 2026

    Why work with GPT, Claude, Gemini, Grok and Perplexity side by side: what each family is good at, how disagreement between them is the useful signal, and why agreement is not verification.

  • Local Browser Storage for AI Chats: IndexedDB and OPFS Explained

    Published June 18, 2026

    How IndexedDB and OPFS store AI chat history and uploaded files locally in your browser — quotas, eviction, the seven-day rule in Safari, and what local-first storage really protects.

  • How to Use ChatGPT Privately Without Saving Chat History

    Published June 17, 2026

    Use ChatGPT Temporary Chat, turn off model training, and avoid saving sensitive drafts — or let Secret Chat keep these protections on by default.

  • How to Check and Rewrite a Risky Work Message Before You Send It

    Published June 17, 2026

    A simple workflow to check a tense work message before sending: spot what sounds rude or risky, rewrite it calmly, and do it privately with Secret Chat.

  • Does ChatGPT Store Your Conversations? What Users Should Know

    Published June 5, 2026

    Does ChatGPT save your chats? How history, training, Temporary Chat, the rebuilt Memory and deletion actually differ — including the two that surprise people: deleting a conversation does not delete the files it saved to your Library, and on the ad-supported tiers your past chats and memory can feed ad targeting. What each control really does, and how to use AI more privately.

  • Best Private AI Chatbots in 2026: Secure AI Tools for Sensitive Data

    Published May 21, 2026

    Compare the best private AI chatbots in 2026, including local AI, enterprise deployments, and privacy-first cloud tools like Secret Chat.