This post is a reading exercise, not an investigation. Every quote below comes from a page Wispr Flow published itself, or from a public post by a Wispr Flow employee. All of them were fetched and checked on 2026-08-29. Nothing here is reconstructed from a review site, a competitor blog, or a summary.
The method is the same for each of the three findings: put document A next to document B, quote both in full, and stop there. Pages change, especially pages about compliance, so the date each one was read is part of the claim. Anyone rechecking these after 2026-08-29 should expect to find different text and should compare against the archive rather than against this post.
TL;DR
- The marketing page and the Security and Compliance FAQ describe the certification status in different terms.
- The FAQ describes what product analytics exclude. A Wispr Flow employee published a country-split word-frequency analysis of user dictation on LinkedIn on 10 August 2026.
- The co-founders' 1 June 2026 post describes an industry pattern of collection by default. The FAQ states the default for trial and standard accounts.
1. The certification wording
Document A, the privacy overview at wisprflow.ai/privacy, read 2026-08-29, advertises:
SOC 2 Type II Compliance
and
ISO 27001
Document B, the Security and Compliance FAQ, read 2026-08-29, lists:
SOC 2 Type I: completed April 2026 by A-LIGN, clean unqualified opinion, Security scope. Report available under NDA.
and, for ISO 27001:2022:
Stage 1: completed April 2026 by A-LIGN. Stage 2 scheduled June 2026.
Type I and Type II are different audits. Stage 1 and Stage 2 are different stages. Both pages are live, both are Wispr's, and the reader can open them in two tabs.
Worth noting alongside this: of the ten dictation vendors checked on 2026-08-29, Wispr Flow was the only one that named its auditor, the audit scope, and the dates anywhere at all. The FAQ is the more specific document, and it is the one that disagrees with the marketing page.
2. The analytics wording
Document A, the same Security and Compliance FAQ, read 2026-08-29:
Privacy controls are retention-based rather than anonymization-based: under Privacy Mode dictation content is not kept, so there is no dictation dataset to anonymize. Operational telemetry and product analytics exclude dictation content by design, and Wispr maintains no anonymized dictation dataset for internal use.
Document B, a post published on LinkedIn on 10 August 2026 by a Wispr Flow employee, read 2026-08-29:
We looked at which filler words and phrases show up most across Wispr Flow users in India vs the U.S. Not what people are saying, just the filler words.
The post goes on to give per-word ratios between the two countries.
The post had 193 reactions and 32 comments when it was read on 2026-08-29. It does not say what its sample was, and the FAQ's own account is that Privacy Mode content is not retained and so could not be counted. The post is public and professional, and it is linked here because a claim without its source is not worth reading. This post refers to its author by role only. The document is the point, and the disagreement is between two Wispr pages, not with a person.
3. The training default
Document A, "Privacy is the foundation", by co-founders Tanay Kothari and Sahaj Garg, dated 1 June 2026, read 2026-08-29:
Most AI tools train on what you say, and they do so without clearly informing you. ChatGPT uses your conversations to improve its models unless you go and change it, and even after you do, it can hold onto your chats for a while.
The pattern is the same everywhere. Collection happens automatically, the disclosure is technically there but hard to find, and it's left to you to figure out what's happening with your data. I don't think that's right. When a product handles input this sensitive, you should be in control of your data from the start and not have to dig for the private option.
The post carries a two-column graphic. Its "Industry default" column reads: "Opted in by default / Data stored and used for training / Privacy settings buried in menus / Recordings reviewed by humans".
Document B, the Security and Compliance FAQ, read 2026-08-29:
Privacy Mode off (standard mode): audio and transcription data may be used to evaluate, train, and improve Wispr's models. This is the default for trial and standard accounts.
Enterprise and HIPAA BAA customers run with Privacy Mode on by default. If you handle confidential or privileged information, enable Privacy Mode before dictating.
The Data Controls page, read 2026-08-29, states where the switch lives:
If you allow your data to be used to improve models, your data (i.e. audio, transcript, edits) may be used to evaluate, train, or improve AI models, by Wispr. You can control this in the Wispr Flow app: Settings > Data and Privacy.
Enterprise accounts default one way. Trial and standard accounts default the other way. The 1 June post is about the industry, and the FAQ is about the product.
The wider pattern
Nine dictation vendors were checked against their own primary documents on 2026-08-29, Wispr Flow among them. One of them, Otter, publishes a complete, dated, public subprocessor list at otter.ai/subprocessors, last updated 31 March 2026. Wispr Flow publishes a partial list of five vendors on its trust centre and states that the authoritative list is Annex 2 of the DPA, available under NDA. For the rest, the search ended in a 404 or in category names such as "trusted service providers".
Four of the ten name no speech-to-text or LLM vendor in any primary document: Notta, Speechify, Aqua Voice, Willow Voice. That is silence in the documents, and silence is recorded here as silence, not as an inference about what the software does.
The document-versus-document shape is not unique to one vendor either. MacWhisper's privacy policy, dated February 2024, states "No data (audio, text or other) leaves your device", while its support docs name Groq, ElevenLabs, Deepgram, DeepL, OpenAI and Anthropic as cloud paths. VoiceInk's README says "100% offline processing ensures your data never leaves your device", while its privacy policy lists twelve optional cloud providers. In both cases the marketing sentence describes the default configuration and the other document describes the capability.
The full comparison, vendor by vendor and field by field, lives at /dictation-app-privacy.
Where Voice Type stands
Three subprocessors: Apple for App Store distribution and payments, Cloudflare for site hosting and DNS, PostHog for website analytics. None of the three is in the path of a spoken word.
LLM rewrite is off by default, set in the shipping source at WhisperState.swift as llmRewriteEnabled: Bool = false, and so is auto-rewrite.
Every rewrite provider is disclosed by name: sixteen remote ones, a custom OpenAI-compatible endpoint, and three that never leave the machine, listed in the shipping source at LLMProvider.swift:213-232. The capability is the thing worth disclosing, and the default is not the whole story.