Why Algospeak Hides Real Conversations From Keyword Monitoring

What is algospeak?

Substituting a coded word for a term you expect a platform to remove or bury is algospeak, and it is the reason a topic can be discussed at volume while your exact-match keyword list reports silence. The mechanism is anticipation: a poster guesses what an automated moderation system will flag, then writes to dodge that guess while keeping the meaning legible to the audience they want. That makes algospeak a coverage problem for monitoring rather than a language curiosity, because the substituted word carries the conversation and the tracked word never appears.

Where does the term algospeak come from?

The word entered wide use through Washington Post reporting by Taylor Lorenz, published 8 April 2022 under the headline Internet 'algospeak' is changing our language in real time, from 'nip nops' to 'le dollar bean'. That piece defines algospeak as "code words or turns of phrase users have adopted in an effort to create a brand-safe lexicon that will avoid getting their posts removed or down-ranked by content moderation systems." The article is paywalled past the opening section, so the definition and the first examples quoted here are what is publicly readable on the Post's page, and the attribution is as reported there rather than from the full text. No source I can reach names a coinage date or a single coiner, so treat April 2022 as when the term spread, not when it was invented.

The first peer-reviewed study of the practice is You Can (Not) Say What You Want: Using Algospeak to Contest and Evade Algorithmic Content Moderation on TikTok by Ella Steen, Kathryn Yurechko and Daniel Klug, published open access in Social Media + Society on 31 August 2023. The authors interviewed 19 TikTok creators and state that theirs is the first article to analyse algospeak as a distinct social media phenomenon. Their finding matters for monitoring: participants anticipated how the algorithm would read a video and chose substitutions that evaded moderation while still letting their target audience find the post.

What does algospeak look like in practice?

The examples below are documented substitutions from 2022 and 2023 sources, not a current dictionary. Steen and colleagues describe algospeak as abbreviating, misspelling or substituting words, and the forms move between platforms and communities.

Documented form	Reported meaning	Documented in	Date
unalive	dead, death	Washington Post	April 2022
SA	sexual assault	Washington Post	April 2022
spicy eggplant	vibrator	Washington Post	April 2022
le$bean	lesbian	Steen, Yurechko and Klug	August 2023
seggs	sex	Steen, Yurechko and Klug	August 2023
clock app	TikTok	Steen, Yurechko and Klug	August 2023

Read that table as illustration, not as a lookup to paste into a query builder. Much of this vocabulary exists because people discuss self-harm, sexual assault, sexuality, sex work and illness under moderation systems that suppress those subjects whether or not the post breaks a rule. Steen and colleagues name non-contextuality, randomness, inaccuracy and bias against marginalised communities as the moderation failures driving the practice.

Why does algospeak break keyword monitoring?

Exact-match monitoring measures the presence of a string, and algospeak removes the string while leaving the conversation intact. The failure is recall, not precision: the posts that matched are still about your topic, so nothing looks broken in the feed, and the missing posts leave no trace anywhere in the system. Measure recall directly instead of assuming it.

Recall = tracked-term matches / all posts about the topic

Suppose you read one community's posts and judge 200 of them to be about your topic, and your tracked term appears in 140. Recall is 140 / 200 = 0.70, so 30 percent of that conversation never reaches your dashboard. Those are illustrative inputs, not a published benchmark, and your own figure will differ by platform, community and subject.

One community's recall check

Tracked-term matches140 of 200

Missed by the keyword list60 of 200 (30%)

How does algospeak distort volume and sentiment?

Missing posts bias both counts and tone, because algospeak is not distributed evenly across the conversation. Substitution concentrates where moderation pressure is strongest, which means the crisis threads, health complaints and adverse-experience posts are the ones most affected, while neutral product chatter keeps using plain words. A sentiment analysis run on that skewed sample reads more positive than the real conversation, and a volume chart understates every spike that coincides with a moderated subject.

This is a different failure from the two most common monitoring complaints. Negative keywords fix false positives, and near-duplicate detection fixes repeated copies of one story. Neither one recovers a post your query never matched.

What can a monitoring team actually do about it?

Nothing restores full coverage, and any vendor promising a complete algospeak dictionary is selling you a list that decays. A preprint by Jan Fillies, Ronald E. Robertson and Jeffrey Hancock, Algospeak, Hiding in the Open, models the practice as coevolution between evaders and detectors, which is the reason a static list ages: the vocabulary moves in response to being detected. The same paper names the ceiling on that movement: a substitution obscure enough to beat every detector stops being understood by the people it was written for.

Five things do help, in descending order of return:

Monitor communities as well as strings. Watching a subreddit's whole feed catches posts where your tracked term never appears, which is what the subreddit finder and RedReplier's keyword monitoring on Reddit are for when you pair them.

Read samples on a schedule. Pull a random sample of posts from the communities you care about, judge relevance by hand, and compute the recall figure above. That is how you learn which substitutions are live now, and it doubles as a quality check on your brand monitoring setup.

Grade meaning, not spelling. AI relevance grading scores whether a post is about your topic instead of whether it contains your string, and RedReplier explains why each mention was flagged so you can audit the judgement.

Date your keyword list. Put a review date on it, add the substitutions your sampling found, and remove terms that went cold. Your boolean search operators are worth tuning, and the tuning is never finished.

Report the gap honestly. When you present volume or sentiment to a stakeholder, say which subjects you know are undercounted. That is a better outcome than a confident chart built on a query you never tested.

How is algospeak different from other internet slang?

Algospeak is a response to a moderation system, and other netspeak is a response to a community. Steen and colleagues draw this line explicitly: leetspeak, LOLspeak, textspeak and similar variations exist to signal identity or group membership, while algospeak is used as a reaction to experiencing content moderation on a platform. The consequence for anyone doing social listening is that algospeak moves on a platform-policy clock rather than a cultural one, so a term that was coded last quarter can be plain text this quarter.

Posters who believe their account is suppressed change how they write before they change what they say, which is the same instinct behind the folk theories in the entry on the Reddit shadowban.

Frequently Asked Questions

Who coined the term algospeak?

No source I can verify names a single coiner. Washington Post reporting by Taylor Lorenz on 8 April 2022 is the piece credited with popularising the word, and Steen, Yurechko and Klug cite that article when they connect the phenomenon to TikTok. The safe claim is that the term spread in 2022, not that any one person invented it.

Is algospeak only a TikTok thing?

No. Steen and colleagues write that algospeak exists on many social media platforms while being largely connected to TikTok, because of TikTok's audiovisual format and its algorithmic moderation. A 2026 preprint on coded-language detection annotated posts from both TikTok and Bluesky, which is evidence the practice is not confined to one app.

Does algospeak affect brand monitoring or only topic monitoring?

Topic monitoring takes the larger hit. Moderation systems police subjects, not brand names, so your brand survives in plain text while the subject beside it gets substituted. The exception is a post that pairs your brand with a suppressed subject, which is exactly the post you most want to see.

Can I just add algospeak terms to my keyword list?

You can add the ones your own sampling finds, and you should. What you cannot do is stay current, because the vocabulary shifts in response to detection and varies by community. Treat any list you build as a floor that decays and needs a review date.

Is writing about algospeak the same as helping people evade moderation?

No. Describing that substitution happens, and measuring how much of a conversation it hides, is a research and measurement question. Publishing a maintained lookup table of current substitutions is a different act with different consequences, which is why the examples above are dated citations rather than a working list.

How do I know whether algospeak is a real problem for my keywords?

Run the recall check. Take a random sample of posts from one community that discusses your topic, judge each one for relevance by hand, and divide the number your tracked terms caught by the number that were actually on topic. If that ratio is close to 1, algospeak is not your bottleneck. If it is not, you now know the size of the gap instead of guessing at it.

Why can't negative keywords or near-duplicate detection fix the algospeak gap?

Negative keywords remove false positives from matches you already have, and near-duplicate detection collapses copies of one story. Neither tool adds a post your string never matched in the first place, and that is the specific gap algospeak opens.

Why does an algospeak dictionary go stale so fast?

A preprint by Fillies, Robertson and Hancock models algospeak as coevolution between people evading moderation and the systems detecting it, so the vocabulary moves in reaction to being caught. The same paper names the ceiling on that movement: a substitution obscure enough to beat every detector stops being understood by the audience it was written for. Any list you build is a floor that decays and needs a review date.

What subjects get hit hardest by algospeak substitution?

Steen, Yurechko and Klug tie the practice to subjects moderation systems suppress whether or not a post breaks a rule, among them self-harm, sexual assault, sexuality, sex work and illness. Substitution concentrates where that moderation pressure is strongest, so crisis threads, health complaints and posts about adverse experiences take the biggest hit while neutral product chatter stays in plain words.

Is algospeak related to shadowban theories?

Both come from the same instinct: someone who believes a platform is working against them changes their behavior to route past it. The entry on the Reddit shadowban documents that same instinct in the folk theories people build about suspected suppression. Algospeak applies that instinct to word choice rather than to posting frequency or account behavior.