Grammarly

What Grammarly is and who develops it

Grammarly is a writing assistant that proposes corrections and rewrites as a text is prepared. Max Lytvyn, Alex Shevchenko and Dmytro Lider founded it in 2009. It now belongs to Superhuman: Grammarly remains the writing product’s name, while the company has broadened its identity and portfolio.

Its place in an editorial process depends on the problem being addressed. Correcting a typo, clarifying a message and creating a draft from an instruction are interventions of different scope. Before using it, identify which parts of the content have already been approved and which remain open to revision.

Reference documentation: Grammarly: about the company.

Origins and product development

• 2009: the writing-assistance project begins, with roots in Ukraine.
• 2015: the free browser extension launches, broadening access.
• 2019: tone detection arrives, interpreting language cues to indicate how a message might be received.
• 2021: Windows and Mac applications extend assistance beyond a single editor.
• 2023: optional generative features introduce content creation from instructions.
• 2024: Authorship is announced in August, with a Google Docs beta scheduled for September, focusing on recording how a document develops.
• 2025: following the addition of Coda and the Superhuman email client, the company adopts the Superhuman name on October 29.

This evolution explains why an older description focused exclusively on correcting English is insufficient. It also makes it necessary to distinguish historical features, research and options available in a current account.

Reference documentation: A History of Innovation at Grammarly · Meet Grammarly’s Tone Detector · Grammarly Raises over $200 Million · AI features and detector limitations · Grammarly Launches Grammarly Authorship · What is Superhuman? · Superhuman stands with Ukraine.

Company, funding and acquisitions

Institutional funding began with $110 million in 2017, followed by $90 million in 2019. In November 2021 the company announced over $200 million more, with Baillie Gifford and funds and accounts managed by BlackRock among new investors. These were separate transactions: the cumulative first two rounds should not be mistaken for the second round’s size.

In 2020 Grammarly made its first external investment, in Docugami. Coda’s acquisition completed in January 2025; May brought an announcement of $1 billion in growth financing from General Catalyst. Superhuman Mail broadened the group into email. In February 2026 Superhuman announced an agreement to acquire Rows to strengthen Coda.

The company retains ties to Ukraine and documents support for its team and initiatives there. It also confirms blocking access to its products in Russia in 2022. These events belong to its corporate history; they do not evaluate the service’s linguistic quality.

Reference documentation: Aeris advises Grammarly on its $110 million minority investment · Grammarly Raises $90M in Financing · Grammarly Raises over $200 Million · Grammarly Announces $1 Billion Growth Financing With General Catalyst · Announcing Grammarly’s Investment in Docugami · Grammarly to Acquire Superhuman · Superhuman to Acquire Rows · Superhuman stands with Ukraine.

Technology: correction, generation and research

The assistance combines language processing, machine learning and linguistic expertise. Correction proposes changes to existing text; generation can create new content. Settings allow generation to be disabled without separately removing the AI underpinning linguistic suggestions.

The mEdIT research studied multilingual grammar correction, simplification and paraphrasing with instruction-tuned models. The 2024 article stated that this research was not then incorporated into the product. A team’s research publication therefore does not automatically identify the model powering each commercial feature today.

The engineering team has also documented optimising correction models for devices, reducing memory requirements and latency. This concerns specific components, rather than proving that all Grammarly functions work offline. For users, the practical question is how the output behaves in their application and whether changes can be rejected.

Reference documentation: AI features and detector limitations · Advancing AI-Powered Intelligent Writing Assistance across Multiple Languages · Grammarly’s Journey of On-Device AI at Scale.

Features and editing decisions

The portfolio includes proofreading, paraphrasing and agents proposing reader reactions. These offer different perspectives on a draft: finding errors, restating an idea or anticipating interpretations. A simulated reader’s reaction is not a test with actual recipients.

A useful editorial sequence has three passes. First, fix typos without changing content; next, check whether each sentence conveys a recognisable idea; finally, review tone and length for the audience. This separates the decision to correct an error from the decision to change a message’s purpose.

For example, a report may contain an observation, a hypothesis and an agreed action. A successful revision keeps those categories visible. Turning them into a fluent but ambiguous paragraph would make the report less useful, even if every underline disappeared.

Reference documentation: Artificial Intelligence at Grammarly.

Language support: Spanish, Catalan and English

The documentation reviewed offers basic correction in over twenty languages. English and Spanish also have advanced features; Catalan is absent from the published list. Translation, rewriting and mobile assistance have separate coverage, so support in one function should not be extrapolated to another.

For work in three languages, prepare equivalent samples containing names, acronyms, negations and sector terminology. The criterion is preserving meaning and natural phrasing in each version, rather than obtaining identical numbers of suggestions. If a function does not support the required language, assign that stage to another resource and retain bilingual review of the result.

Reference documentation: Grammarly multilingual writing assistance.

How to review a document

In Docs, the agent lets users open suggestions, accept or dismiss them and undo edits. Broader review organises proposals into categories and offers paragraph rewrites. Availability depends on plan and environment; the guide also covers use through Superhuman Go.

1. Define the assignment: record the audience, purpose, length and facts that must remain unchanged.
2. Save the original: retain a version for before-and-after comparison.
3. Review in layers: separate correction, clarity, tone and content generation.
4. Decide each change: accept a proposal only when it improves the text without changing the fact conveyed.
5. Check the whole: read headings, lists and references after editing, not just changed sentences.
6. Close the review: record the owner and final version in the document system used.

This sequence is Sabentis’s editorial proposal. The software assists language review; the team retains approval and publication decisions.

Reference documentation: Grammarly agent user guide.

Platforms, plans and access choices

The service offers web access, extensions and desktop and mobile applications. Its plan page distinguishes Free, Pro and Enterprise, with differences in rewriting, administration and data-protection capabilities. Check the specific offering instead of carrying prices or features forward from an older comparison.

For a team trial, identify where writing actually happens: a web editor, email, a shared document or a desktop application. Then check whether required functions appear, whether they interfere with accessibility or shortcuts and whether settings can be managed consistently. Licence cost is only part of adoption cost; review, training and administration also matter.

Reference documentation: Grammarly Prices and Plans.

Authorship, similarity and AI detection

Once enabled, Authorship records the writing process in environments including Word, Google Docs and the Grammarly editor. Its reports help distinguish text origins. This record differs from inferring whether a finished document contains AI-generated text.

Editorial review should separate three questions: where a claim comes from, how it changed and who takes responsibility for the final version. A correctly cited passage can still contain an incorrect fact; manually typed text can repeat another person’s idea without attribution. No single score resolves all three questions.

For training and professional publishing, agreeing how assistance is disclosed and retaining sources and versions provides verifiable context. Decisions about authorship or permitted use should consider that process and the assignment’s rules.

Reference documentation: Grammarly Authorship.

Data, training and permissions

Documentation distinguishes account types: content use for improvement and training initially defaults to on for individuals and teams purchased online, with an opt-out. It defaults to off for Education and business offerings purchased through Sales. Check the actual account rather than assuming identical conditions across paid plans.

The Trust Center distinguishes authorised data access and documents saved in the editor, retained until deleted. Disabling training does not prevent every form of processing needed to provide the service.

Within an organisation, the practical decision starts with document type. Publishable text, an internal investigation and a communication identifying individuals require different decisions. Agreeing authorised applications, configuration owners and version retention helps prevent an extension active across multiple surfaces from receiving content outside its intended use.

Reference documentation: Product Improvement and Training Control · Grammarly User Trust Center.

Documented evaluations and debates

In 2021, Dorothy Mayne of UW–Madison’s Writing Center described useful corrections alongside errors and concerns about linguistic diversity. This is an evaluation from that period, not a measurement of the current version. It distinguishes editing form from supporting learning and the author’s intentions.

In March 2026, CEO Shishir Mehrotra announced Expert Review’s deactivation following criticism of how it represented experts’ voices without adequate control by those concerned. His Decoder interview also addressed the controversy. This history should not be confused with an actual human review by the named specialists.

To evaluate an editorial tool, Sabentis proposes testing your own texts and recording both helpful suggestions and rejected changes. Check whether it preserves voice, facts and purpose as well as readability. A historical review or commercial demonstration cannot replace that contextual trial.

Reference documentation: Revisiting Grammarly: An Imperfect Tool for Final Editing · Shishir Mehrotra: Expert Review disabled · Decoder interview with Shishir Mehrotra.

Security history and maintenance

A historical NHS alert documented an extension authentication vulnerability and its patch in 2018. The archived alert does not describe the product’s current state. In December that year, HackerOne announced the bounty programme’s public opening after an earlier private phase.

The practical lesson is to maintain ongoing review of extensions, versions and permissions. An old incident does not establish that a problem remains open; an old fix does not establish the absence of new risks. Enterprise selection should examine current documentation and responsibility for updates.

Reference documentation: Grammarly Plugin Authentication Vulnerability · Grammarly’s Bug Bounty Program Goes Public.

Applications in occupational safety and health

It can help make prevention communications clearer and easier to read, especially when several people contribute to a document. Occupational safety messages need direct instructions and consistent terminology. A stylistic improvement should not change the strength of an obligation or remove an important condition, exception or deadline that readers need to understand.

In safety communications, its value may lie in making subjects and actions explicit: who reports an issue, where a document is found and what information must be supplied. In a report, more precise sentences can help separate an observation from an interpretation. In training materials, it can highlight overly dense paragraphs. These are editorial applications proposed by Sabentis, rather than checks that a preventive measure is appropriate for a hazard.

Worked example and checks

For example, a campaign owner reviews a housekeeping email. They use suggestions to shorten long sentences while preserving approved instructions, owners and deadlines, then check that recipients can identify the action they are expected to take and where to ask questions.

1. Take the approved housekeeping email and mark information that must remain: recipients, responsible people, date and expected action.
2. Review writing errors first and clarity second, avoiding blanket acceptance of every suggestion.
3. Compare any rewrite with the original: “must” and “could” do not convey the same commitment.
4. Check that a shortened sentence retains necessary exceptions, locations and references.
5. Read the final version as a recipient and explain in your own words what action it requests.

Retaining a change comparison makes it possible to explain why each material modification was accepted or rejected. If a suggestion replaces a technical term with a more common but less precise one, keep the term and explain its meaning. The expected result is a clearer email with the same approved content, rather than a new safety instruction created by a style recommendation. A second reader can check whether the action, deadline and responsible role remain equally explicit.

Limitations and conditions of use

A tone suggestion may soften a prohibition too far or change technical meaning. The tool reviews language, not preventive measures. Language capabilities can vary by supported language, application and environment settings.

A language-quality score does not measure actual understanding, technical accuracy or compliance. Generative features may introduce facts absent from the draft. For internal documents, review the surfaces on which the assistant is active and the organization’s conditions of use.

Choosing and complementing the tool

Choose Grammarly when the main problem is improving an existing text during writing. DeepL can complement translation, followed by independent review. Microsoft Word helps retain comments and changes during editing. Where operational instructions are involved, agree in advance which parts can be reworded and which must remain verbatim.

A useful selection trial uses a real paragraph without confidential information and compares suggestions for the available language, application and product configuration. The existence of a feature does not guarantee identical behavior across all these contexts.

Frequently asked questions

Are grammar correction and text generation the same?

No. Grammarly distinguishes linguistic suggestions from generative functions that can create or materially change content.

Should I accept a suggestion to achieve a higher score?

Only if it preserves meaning and improves the message for its readers. The score does not replace editorial judgment.

Can it certify who wrote a document?

An automated detector is not conclusive proof of authorship. A drafting history and the sources used provide different contextual evidence.