AI writing tools: What Medium's TK app changes
Medium's TK app puts AI inside the writing process. See how its tools work, what writers gain, and the trust, privacy, and authorship questions to weigh.
Medium's TK App Puts AI Inside the Writing Process. What Changes for Writers?
AI writing tools aren't just for fixing a finished draft anymore. Medium's new TK app puts an assistant right inside the editor, where it can answer questions, pull up saved notes, suggest headlines and edits, and help you move from a half-formed idea to a usable draft. That shift matters because it blurs a line writers have always managed - editing versus authorship. A spelling suggestion is one thing. Something that nudges your structure and wording is another.
So TK lands in the middle of an argument that won't go away: does AI support improve writing, or does it quietly replace the writer's judgment? Medium clearly wants the former. It also acknowledges the real worry - low-quality output, unclear provenance. Its answer is to push for disclosure and make early drafts shareable for human feedback.
This article looks at what TK does, where it might fit in an actual writing workflow, and what its choices mean for authorship, privacy, and publishers. The question isn't whether AI belongs in writing at all. It's what you delegate, what you keep owning, and how clearly you tell readers where the help came from.
What is Medium's TK app and how does it work?
TK is a writing app from Medium that places an AI assistant inside its editor. Instead of switching to another service to ask for help, a writer can type "@TK" and then make a request. Medium lists potential uses like quick factual questions, finding material the writer previously saved, and getting feedback on a draft or headline.
The name comes from an old editorial habit. In publishing, "TK" is a placeholder for information that's still missing. It's deliberately odd in English, which makes it easy for editors and proofreaders to spot unfinished details. Medium borrows that convention for an assistant meant to help fill in gaps as you develop a piece.
From blank page to working draft
TK's features cover a few different jobs that people often lump together as "AI writing." Some are more like research or retrieval - asking a quick question or pulling up earlier notes. Others are more editorial: requesting feedback on a headline, asking for a critique, or checking a draft against a publication's style guide.
That split matters. Pulling up your own saved notes is not the same job as generating new text. Both can involve AI, but they raise different questions about accuracy, originality, and responsibility for the final wording.
TK also includes writing prompts and a social layer for sharing early drafts for other people's input. In practice, it sits somewhere between a text editor, an AI assistant, and a collaborative drafting space. Medium also says writers can optionally connect their own AI agent through TK's MCP server or a Claude skill, so the app can use information from the agent's memory about a writer's projects and positions.
What is available at launch?
TK launches on the web and is free to people with a Medium account. Medium says a desktop app is planned. The app can import and export Markdown, though Medium says it isn't itself a Markdown app.
One privacy point is important: writing in TK is not encrypted. The product is built for collaboration, and users can set access permissions, but those controls don't replace encryption. If you're working with confidential drafts or sensitive source material, you'll want to understand that difference before you commit.
How could TK change a writer's workflow?
The practical value of an assistant in the editor isn't just that it can produce text. It's that it might reduce small interruptions that break your focus: opening a new tab to verify a detail, digging for a note from an earlier brainstorming session, or asking a colleague to check a headline before your draft is ready.
Those interruptions have a cost, too. You move away from the draft, change context, find the information, and then come back to the argument you were building. If TK can retrieve useful material right where you're writing, the workflow may feel more continuous. But there's no guarantee. A distracting or inaccurate suggestion can derail you just as effectively as a browser tab.
Assign the assistant bounded tasks
If you adopt TK - or any AI writing tool - the easiest way to keep control is to split the work into clear categories:
- Retrieval: Find notes or material you already saved.
- Review: Flag unclear passages, weak headlines, or mismatches with a style guide.
- Generation: Draft new wording, examples, or sections.
The first two tend to be easier to oversee because you can compare the output to what you already wrote or apply your own judgment to the critique. Generation can still be helpful, but it raises the stakes. A smooth, fluent passage can smuggle in unsupported claims, flatten a distinctive voice, or make a draft sound "done" when the underlying thinking isn't.
For example, a writer might ask TK to pull up saved notes about a product decision, then write the explanation themselves using those notes. Or they might ask whether a headline actually matches the article. That's different from asking TK to invent the argument and then treating its wording as the final work.
Keep the writer in the decision loop
A workable workflow should make the assistant's contribution visible to the person who owns the final draft. In practice, that means checking factual answers against reliable sources, treating editorial suggestions as proposals rather than orders, and revising any generated language until it matches what the writer actually means.
TK's ability to map drafts to publication style guides could help editors apply consistent standards, especially across larger teams. Still, a style guide isn't a replacement for editorial judgment. It can catch tone or formatting mismatches. It can't decide whether a story is fair, whether evidence is strong enough, or whether the writer has actually made a real argument.
The best use case, then, isn't "let AI write faster." It's "remove one specific obstacle that doesn't belong in the way," without handing over responsibility for the piece. That framing also makes it easier to tell whether you're saving time - or simply generating more text you still have to review.
What are the risks of AI-assisted writing?
The big worry isn't that AI always produces bad writing. It's that help can drift into authorship without a clear moment where you consciously decide, "This part is mine" versus "This part is the tool's default."
A writer may start by asking for a headline suggestion, accept a rewritten paragraph, then request another section. Over time, the final draft can look more like the assistant's default choices than the author's original reasoning.
This risk shows up especially in long-form work, where voice, argument, and evidence grow together. A paragraph can read smoothly while still misrepresenting what the author thinks - or while turning a weak idea into something that looks settled. Editing for fluency doesn't automatically fix accuracy or integrity.
Authorship is a workflow question
Publishing debates often frame this as a binary: either a piece is human-written or AI-written. In real workflows, it's messier. A writer might use an assistant to locate their own notes, check grammar, propose a headline, critique structure, or generate a chunk of prose. Those actions don't carry the same implications, even if they all involve AI at some point.
AI detection doesn't solve the confusion. Detection tools can be unreliable, especially when they're expected to tell the difference between AI-generated text, AI-assisted revisions, and text written without AI. A score can't explain how a draft was made - or who should be accountable for what the draft claims.
That's why process disclosure is more useful than pretending there's a single obvious line. Medium says TK lets authors publish notes explaining how AI was used. The stated goal is to make the provenance of writing clearer and easier to inspect, including prompts and source material. That's a start, not a complete publishing standard, and readers and editors still need enough context to understand what the note really means.
Transparency needs to be specific
A useful disclosure doesn't just say, "AI was involved." It should describe the role the tool played. "I used AI" could mean anything from fixing typos to generating most of the article. A more helpful note explains whether AI supported research, provided feedback, or helped with wording - and clarifies who verified facts and made the final editorial calls.
There's a trade-off, too. Detailed records can improve accountability, but keeping prompts and source material can also expose private ideas or information about other people. Writers and publishers need to balance inspectability with confidentiality, especially when drafts include unpublished reporting, client details, or sensitive material.
Medium's pitch is that AI can amplify human intelligence and that writing still matters. That's a reasonable product philosophy, but it doesn't automatically answer how credit should work for any specific piece. The practical standard has to come from writers, publications, and audiences: explain the process, verify the work, and don't imply more human contribution than actually happened.
What should writers and publishers evaluate before using TK?
A writing app should be judged by what it helps people do, not by whether it includes AI. With TK, that means checking whether its assistance makes drafting and review more effective - without creating unacceptable risks around privacy, accuracy, or editorial accountability.
Start with the use case. If the problem is losing track of research and rough notes, retrieval is the feature to test. If your team needs consistent headline or style-guide feedback, the editorial side may matter more. If you're hoping for a system that generates publishable prose, you'll need a stricter review process and a clear disclosure policy.
A short adoption checklist
Before you move a real project into TK, answer these questions:
- What task are we trying to improve? Define a specific bottleneck, like finding saved notes or checking a draft against a house style.
- What information will the tool handle? Because TK writing is not encrypted, decide whether the draft includes material you shouldn't put into a collaborative, non-encrypted service.
- Who checks the output? Assign responsibility for verifying facts, reviewing edits, and approving the final text.
- How will AI use be disclosed? Agree on wording that tells readers or editors what the assistant actually did.
- How will we judge the result? Measure time saved and review effort, not just the number of words produced.
For teams, these questions shouldn't be one-off discussions. Turn them into operating rules. A publication could allow AI for headline alternatives and copyediting while requiring disclosure for generated passages. Another might restrict AI for certain reporting or creative work. The right policy depends on the work, audience expectations, and what happens if an error slips through.
Make the decision reversible
A limited trial is usually better than a rollout based on a product demo. Pick a low-risk project, test one or two features, and compare the workflow with what you do today. Watch whether the assistant reduces interruptions, whether suggestions get accepted or ignored, and how much extra verification it adds.
Also test the edges - not just the easy cases. What happens when a saved note is incomplete, a prompt is unclear, or feedback sounds confident but doesn't match what the writer intended? Those moments show whether the tool supports judgment or quietly pressures users to accept its output.
TK's Markdown import and export may make it easier to move drafts between systems. But portability doesn't remove the need to understand where drafts are stored, who can access them, and what permissions apply. For sensitive work, those operational details should be decided before a team adopts the tool.
Conclusion
Medium's TK app folds AI assistance into the writing environment rather than keeping it as a separate step. That can help writers retrieve notes, handle small questions, and get faster feedback. It also makes the boundaries between editing, collaboration, and authorship feel more immediate - and harder to ignore.
The sensible approach is to define the task before you adopt the tool. Use AI only where its contribution is bounded and reviewable, verify anything that affects factual claims, disclose meaningful assistance clearly, and treat privacy as something you plan for - not something you hope is fine later. TK is worth evaluating on those terms: not as a verdict on AI and writing, but as a set of workflow choices writers and publishers will have to own.