Most tone checkers never check your tone

I recently gave three AI research agents the same brief, asking which tone checkers are the best right now. Claude, ChatGPT and Gemini each came back with a long, confident report. They disagreed on plenty, cited some sources that don’t exist, and one of them invented a charming origin story about my own company. So I did what I keep telling people to do with AI research. I kept the structure, traced every claim back to a primary source, and cut whatever didn’t survive contact with reality.

This is the merged, verified version. I sorted the seventeen tools by the job each one does, and numbered the receipts at the bottom.

The bias card face up, because I co-founded one of these tools

Subtext is my company, and it appears in this post. It is an app that reads how a personal message will land before anything rewrites it, which means the sorting principle this whole article runs on, diagnosis before rewriting, conveniently flatters my own product. Weigh everything I say about it accordingly.

To keep this useful anyway, I held my own product to the same standard as everyone else’s. Claims come from product pages, documentation, app store listings or peer-reviewed research, and each one links out with a numbered source. Where a claim only exists in a vendor’s marketing, I say so. And I cut what the AI agents confidently reported but no source supports, including that flattering origin story (a “former therapist” supposedly built Subtext; nobody did, Anton is an engineer and I do the words).

A real tone checker diagnoses before it rewrites

Every tool in this space gets called a tone checker, but there are two different jobs hiding under one label.

Diagnosis tells you how your text currently reads. Rewriting changes it into something else. The difference sounds academic until you look at a real draft, so here’s the kind of read Subtext gives before it touches a word:

Draft “Fine. Do whatever you want.”

Reads as clipped · “whatever you want” reads as withdrawal · lands as passive when you probably meant tired

A rewriter will happily turn that into “Sounds good, go ahead with whichever option works best,” a nicer sentence that never mentions any of that. A diagnostic tool stops you first and shows you the read above. Once you see it, you can decide what to do about it. Maybe you soften it. Maybe you meant the edge and keep it. I’ve written up what an AI message checker can check, and how to do it yourself.

The rewrite-only tools skip that step. They give you a nicer sentence without ever telling you what was wrong with yours, which is how people end up sending polished messages that still say the wrong thing. So I’ve sorted the field into tools that read first, tools that only rewrite, and tools that tune your tone to one specific audience.

In one email study, senders expected their tone to land 78% of the time. It landed 56%.

Before the tools, the reason they exist.

In 2005, Justin Kruger, Nicholas Epley and colleagues published five experiments on tone in email¹ in the Journal of Personality and Social Psychology. In one study, senders predicted their tone (serious or sarcastic) would come through about 78% of the time. Receivers actually got it right 56% of the time², barely better than a coin flip, and the receivers were even more confident than the senders, believing they had read 89% of the messages right. The authors’ diagnosis, in their own words¹:

This overconfidence is born of egocentrism.

You hear your own message with the full soundtrack of your intentions. Your reader gets the words on a screen. It is also why you can’t reliably tell on your own whether a text sounds rude.

Twenty years later that finding is the business model of every tool below.

Four tools that read your tone before touching a word

Subtext

Subtext is the tool I co-founded, built for personal messages rather than documents. You paste a draft, drop in a screenshot of the conversation, or record a voice note, and it reads the whole exchange before saying anything. It marks the exact words that could feel harsh, passive or easy to misread, then names the feeling underneath³. After the read, one-tap adjustments (Warmer, Shorter, More casual, Add confidence)⁴ rewrite the message while keeping your phrasing as the starting point, and it never sends anything for you.

The store listing is blunt about who it’s for, naming people who rewrite the same message five times before sending, overthinkers, and anyone wondering “does this sound okay?”⁴. A lot of the warmest reviews come from neurodivergent users, the people an AI tone checker for neurodivergent adults is written for, especially ADHD users who feed it long, chaotic drafts and get back something clear that still sounds like them⁵. It runs on iOS, Android and a synced web app, in 17+ languages written natively rather than translated. Free to try, then a subscription unlocks the rest. I’ll name two limits. It’s built for messages, so long-form documents belong elsewhere, and its own terms say plainly that it’s a communication assistant, not therapy⁶, and that your texts aren’t used to train AI models.

Grammarly’s tone detector

Grammarly runs the most widely deployed tone diagnosis in the world. It identifies the tone of your writing from word choice, phrasing, punctuation and even capitalisation⁷, using a combination of rules and machine learning⁸. When it launched in 2019 it could name about 40 tones, from friendly and joyful to aggressive and annoyed⁹. It needs a bit of text to work with: around 150 characters in the browser¹⁰, 90 on the mobile keyboard¹¹. The labels are the diagnosis; sentence-level tone rewrites sit in the paid tiers. For companies, Grammarly Business adds brand tones, where admins pick from more than 50 tones and mark them on-brand or off-brand¹² and everyone gets nudged in real time. Grammarly has since added Reader Reactions, an agent in Grammarly docs that simulates how a reader you pick, such as your manager or a client, might interpret what you wrote¹³. If your writing lives in documents and work email, this is the benchmark.

Goblin Tools (the Judge and the Formalizer)

Goblin Tools is a collection of tiny free AI tools by Belgian developer Bram De Buyser, described on its own site as “small, simple, single-task tools”¹⁴ built mostly for neurodivergent people. Two of them matter here. The Judge senses the mood or emotion behind any message you’ve received, and the Formalizer instantly rewords your text for the right tone¹⁵, with a “spiciness” dial for how much it changes. That receive-side decoder is rare; almost everything else in this post only looks at what you’re sending. The website is free and the mobile apps are cheap one-off purchases.

Sapling

Sapling is aimed at customer support and sales teams, and it publishes the most transparent tone taxonomy in the group. Its free tone checker identifies more than 25 emotional tones (joy, confusion, gratitude, fear) at both sentence and document level¹⁶, and the developer API covers 28 tones with per-sentence probabilities¹⁷. If you want tone detection built into your own product, this is the practical successor to the old IBM API we’ll bury in a minute.

Seven tools that rewrite into whatever tone you pick

These are useful and I use some of them, but apart from Copilot in Outlook, none of them will tell you how your original draft reads. They hand you a different draft.

Apple Intelligence Writing Tools sit at the system level and rewrite selected text to be more professional, friendly, or concise, nearly everywhere you can type including third-party apps¹⁸, plus a free-form “describe your change” box¹⁹. Free with iOS 18.1, iPadOS 18.1 or macOS Sequoia 15.1 on supported hardware. Three presets, zero diagnosis. I compared what Apple, Google, Grammarly and ChatGPT each do to your messages feature by feature.

Microsoft’s Copilot does the equivalent inside Office, where you select text in Word, choose Rewrite, and use the Adjust Tone menu (Neutral, Casual and Concise among the options)²⁰. Outlook is the exception. There, Copilot’s email coaching offers suggestions on the tone, clarity and reader sentiment of a draft²¹, which makes it the one tool in this group that does some diagnosis, for work email only. What you get depends on your Microsoft plan and which Outlook you use, and I haven’t seen evidence on how accurate the coaching is.

Wordtune is the quickest of the dedicated rewriters. You highlight a sentence, tap Casual or Formal, plus Shorten and Expand²², and pick from the suggestions. Free tier, with paid plans from about $7 a month billed annually²³, as of August 2026.

Hemingway Editor built its name on readability, colour-coding dense sentences and passive voice, which is adjacent to tone rather than tone itself. The paid Hemingway Editor Plus adds AI rewriting, including adjusting selected text to be more professional or friendly²⁴ and dedicated more casual, more formal and more persuasive tools²⁵. Still the best cure for bloated prose I know.

Three quick ones. LanguageTool is the multilingual pick, a grammar and style checker in around 30 languages with a paraphraser in the paid tier; tone is a side effect here, not the product. ProWritingAid is for long-form manuscripts, a deep stack of style and readability reports, with no live read on how a text message will land. And QuillBot is the students’ paraphraser with preset modes, Formal among them; it rewrites fluently and tells you nothing about your original.

Four tools that tune your tone to one specific reader

This group answers a different question. How should I sound to this specific person, or as this specific company?

Crystal profiles your recipient with the DISC personality framework, then its Writing Assistant checks the words, phrases, style and tone of your email in real time inside Gmail or Outlook and steers you toward that person’s communication style²⁶. I find it slightly unsettling and occasionally brilliant, and it was clearly built for sales.

Lavender is a cold-email coach that scores every draft from 0 to 100 as you write, with tone among the factors, and suggests fixes tied to reply-rate data²⁷. There’s a free tier of 5 emails a month, then plans from $29 a month (as of August 2026)²⁷. If your job is outbound, the score becomes a habit fast.

Textio²⁸ owns the hiring niche, flagging biased or exclusionary language in job posts and performance feedback and scoring how the language will land with candidates. And Writer works at enterprise scale: upload writing samples and it builds a voice profile²⁹ so a whole organisation’s AI output stays in one tone.

The free single-page checkers are fine, with one caveat

Beyond the big names sits a long tail of free paste-your-text-here pages. Sapling’s free checker (above) is the credible one. LiveChatAI’s tone checker³⁰ identifies a dominant tone and highlights which phrases drove the verdict, detection only, no rewrites. AskAI’s version³¹ wraps a chat model around a 0 to 100 formality score and flags tone drift across long documents, with a paid tier for heavier use. I’d trust them for a blog paragraph or a cover letter.

Anything you paste into a free web page travels to someone’s server, and most of these pages say little about what happens next. I’d paste a marketing paragraph in without a second thought, and I’d keep a breakup text well away from them.

The 17 tools, by what they do first

The table groups all seventeen tools above, Subtext included, by what each one does first.

Subtext is the author’s own product. The rows follow this post’s sections and are not a ranking. Text pasted into any free web checker travels to someone’s server.

The tone API everyone used to cite is dead

For a decade, “tone checking” in the technical sense meant IBM Watson Tone Analyzer, the API behind countless demos and research papers. It’s gone. IBM announced the deprecation on 24 February 2022, and the service shut off on 24 February 2023³². What remains is a small pre-built tone classification model inside Watson’s Natural Language toolkit³³, covering English and French with a short list of categories such as excited, frustrated, impolite, polite, sad, satisfied and sympathetic.

I include the obituary because it explains the market. The standalone tone API era ended, and the energy moved into the assistants above and into large language models that rewrite anything on demand. Which brings us to the uncomfortable part.

What the research says about letting AI touch your words

The rewriting part is now trivial. The consequences are less tidy, and there’s a real literature on them.

Researchers call it AI-mediated communication: when an AI system modifies, augments or generates messages between people³⁴, as Jeff Hancock, Mor Naaman and Karen Levy defined it in 2020. The headline finding came a year earlier, in a CHI study of Airbnb host profiles, where Jakesch’s team found that readers rated text they believed was AI-written as less trustworthy³⁵. A follow-up study of AI smart replies found the suggestions changed the language people actually produced, nudging it toward the positive, and shifted how conversation partners perceived each other³⁶. I went through whether AI message rewriters actually work in more depth separately.

There’s also a bias problem in the models themselves. A 2025 study out of Monash University generated thousands of assistant dialogues from deliberately neutral prompts and found a consistent skew. The models default to sounding, in the researchers’ words, “overly polite, cheerful, or cautious even when neutrality is expected”³⁷, with tone classifiers detecting the skew at macro-F1 scores up to 0.92³⁷. The machines have an accent. If you let one rewrite everything you send, you start speaking with that accent too. The NLP field even has a name for the machinery underneath, text style transfer, with a whole survey literature³⁸, and none of it is trained on what makes you sound like you. That accent is why Subtext keeps your own phrasing as the starting point of any rewrite instead of swapping it for a house style.

My conclusion from all of this is to use these tools as a mirror first and a ghostwriter rarely, and yes, notice whose conclusion that is. A tone read before sending goes after the 78/56 problem from the Kruger study, though it is still a guess made from outside your head. In a 2025 ACL study, human readers and language models both matched the emotions writers said they felt only imperfectly³⁹, so treat any checker’s read, mine included, as a second opinion. A full rewrite of everything you send creates a new problem, the one where your mum can tell a robot answered her.

How I’d choose

For personal messages and hard conversations, my shortlist is Subtext or Goblin Tools. Subtext if you want the diagnosis, the rewrite and the conversation context in one place with your voice preserved; Goblin’s Judge and Formalizer if you want free, single-task and no account. For documents and work email, Grammarly’s detector is the benchmark, with Hemingway as the edit for bloat. If you’re on Apple hardware, Writing Tools cover preset tone shifts at no extra cost, and if your job already pays for Copilot, its Outlook coaching gives work email a basic read. You only pay for a dedicated tool when you need actual diagnosis. For sales outreach take Lavender, for recipient-tuned messages Crystal, for hiring language Textio, and for company-wide brand voice Writer. Broke and curious? Sapling’s free checker plus Goblin Tools will teach you the concept for nothing.

And whichever one you pick, keep the last word. The research above is fairly clear that the goal is a message that sounds like you on a good day, not like a model with impeccable manners.

If the “does this sound okay?” moment is the one you know too well, that’s the exact moment Subtext was built for. Paste the draft or screenshot the thread so it reads both sides, and it names what your message is about to signal and shows the exact words doing it, then offers versions that still sound like you. Try Subtext in your browserGet the appTry Subtext in your browser

---

I merged three AI research reports for this post and cut what didn’t check out, but I won’t pretend the survivors are beyond argument. If you make one of these tools and think I’ve sold yours short, or you know a checker that diagnoses better than the four I found, tell me and I’ll test it and update the post.

Samet Durgun is the co-founder of Subtext, an app that catches the emotional tone of your messages and rewrites them in your own voice. He’s based in Berlin.

Sources

Every link above goes to the primary source where one exists. The citations are numbered below, in order of appearance.

Why tone breaks

1. Kruger, J., Epley, N., Parker, J., and Ng, Z. (2005). Egocentrism over e-mail: can we communicate as well as we think? Journal of Personality and Social Psychology, 89(6), 925-936. DOI: 10.1037/0022-3514.89.6.925. Open PDF:
2. American Psychological Association, Monitor on Psychology (February 2006). E-mails and egos.

The tools

3. Subtext, official site. (feature descriptions)
4. Subtext: AI Writing Assistant, Apple App Store listing (features, audience, in-app subscription prices).
5. Subtext: AI Writing Assistant, Google Play listing (use cases and user reviews).
6. Subtext, Terms and Privacy.
7. Grammarly, Writing Tone Detector and Tone Suggestions.
8. Grammarly blog, Meet Grammarly’s Tone Detector.
9. VentureBeat (24 September 2019). Grammarly uses AI to detect the tone and tenor of your writing.
10. Grammarly Support, Using the tone detector in the browser extension.
11. Grammarly blog, Tone Detector for Mobile.
12. Grammarly Business, Brand Tones.
13. Grammarly Support, Reader Reactions user guide (checked September 2026).
14. goblin.tools About, quoted via Autism Community Training.
15. Goblin Tools, official Google Play listing.
16. Sapling, Tone Checker / Detector.
17. Sapling, Tone Detection API documentation.
18. Apple Support, How to use Writing Tools with Apple Intelligence.
19. Apple Support, Use Writing Tools on Mac.
20. Microsoft Support, Use Copilot in Word with a screen reader (Rewrite section, Adjust Tone menu).
21. Microsoft Support, Get email coaching with Copilot in Outlook (checked September 2026).
22. Wordtune, Free Online Rewriting Tool.
23. Capterra, Wordtune pricing (secondary; prices as of August 2026).
24. Hemingway Editor Help, Rewriting text with AI tools.
25. Hemingway Editor blog (5 March 2024). Announcing powerful AI tools for adjusting tone and style.
26. Crystal, Getting Started with the Crystal Writing Assistant (archived; page removed from crystalknows.com since).
27. Woodpecker (2026 review; secondary, prices as of August 2026). Lavender AI Review.
28. Textio, official site (augmented writing for hiring; bias and tone guidance).
29. Writer, Introducing voice.
30. LiveChatAI, Free AI Tone Checker.
31. AskAI, AI Tone Checker.
32. IBM Watson SDK deprecation notice (Tone Analyzer deprecated 24 February 2022, withdrawn 24 February 2023).
33. IBM, Watson Natural Language Processing tone classification block.

Writing with AI

34. Hancock, J. T., Naaman, M., and Levy, K. (2020). AI-Mediated Communication: Definition, Research Agenda, and Ethical Considerations. Journal of Computer-Mediated Communication, 25(1), 89-100.
35. Jakesch, M., French, M., Ma, X., Hancock, J. T., and Naaman, M. (2019). AI-Mediated Communication: How the Perception that Profile Text was Written by AI Affects Trustworthiness. CHI 2019.
36. Mieczkowski, H., Hancock, J. T., Naaman, M., Jung, M., and Hohenstein, J. (2021). AI-Mediated Communication: Language Use and Interpersonal Effects in a Referential Communication Task. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1).
37. Bodara, H., Mushfiq, M. M., and Siddiqui, I. F. (2025). Bias Beneath the Tone: Empirical Characterisation of Tone Bias in LLM-Driven UX Systems. UISE ’26 workshop (ACM).
38. Jin, D., Jin, Z., Hu, Z., Vechtomova, O., and Mihalcea, R. (2022). Deep Learning for Text Style Transfer: A Survey. Computational Linguistics, 48(1), 155-205.
39. Li, J., Zhou, Y., Narayanan Venkit, P., Islam, H. B., Arya, S., Wilson, S., and Rajtmajer, S. (2025). Can Third Parties Read Our Emotions? Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 21478-21499.