Automated Keyword Blacklists Penalize Exact Tokens: How Algospeak Uses Substitution to Evade AI

Social media executives frequently assure advertisers that their artificial intelligence systems can instantly detect and remove harmful content before it spreads. But a quick scroll through TikTok or Instagram reveals a very different reality, where teenagers routinely bypass these billion-dollar filters using nothing more than a few creative typos. The evidence shows that automated moderation is fundamentally broken by a phenomenon known as "algospeak."[1][3]

When a platform blacklists the word "lesbian," users simply type "le$bian" or "le dollar bean." When discussions of suicide are banned, creators talk about "becoming unalive." These phonetic and orthographic substitutions allow human readers to understand the intended meaning perfectly, while the machine-learning classifiers remain completely blind to the conversation happening right in front of them.[1][3]

The core issue lies in how these moderation systems are built. Despite the tech industry's claims of cutting-edge semantic analysis, many platforms still rely heavily on rudimentary keyword blacklists. These systems scan text for exact matches of banned words, operating on the assumption that language is static and predictable.[3][6]

The mechanics of evasion

But language is anything but static, and internet users are highly motivated to keep their communities intact. When marginalized groups find their discussions unfairly targeted by algorithmic filters, they adapt by inventing new vocabulary. This cat-and-mouse game has transformed the internet into a massive engine for linguistic evolution, where new dialects are born out of necessity.[1][3][5]

The scale of this evasion is staggering. A 2022 survey by TELUS International found that 51 percent of internet users had encountered algospeak across social media, forums, and gaming communities. Among Generation Z, that number jumped to 72 percent, with more than four in ten young adults actively using coded language themselves to dodge automated filters.[4]

The mechanics of algospeak rely heavily on phonetic substitution, which alters the surface form of a word while preserving its phonological similarity. A human reader sounds out the word "seggs" and immediately understands it means "sex," drawing on phonological recoding and lexical access. The algorithm, however, only sees a string of letters that does not match its blacklist.[3][5][6]

Orthographic substitution works on a similar principle, replacing specific characters with visually similar symbols or numbers. Writing "k*ll" instead of "kill" or using the corn emoji to represent pornography are classic examples of this strategy. These visual tricks exploit the fact that machine-learning classifiers are trained on standard spelling, leaving them vulnerable to deliberate noise.[1][3]

A global linguistic shift

This phenomenon is not limited to English or Western platforms. A 2026 study published by the Association for the Advancement of Artificial Intelligence analyzed over 200,000 movie-recap videos on Chinese social media. The researchers found that content creators systematically altered their hard-coded captions to evade perceived moderation on platforms like Douyin.[2]

The AAAI study demonstrated that Chinese creators use complex character-level perturbations and homophonic substitutions to encode meaning implicitly. These cross-lingual strategies require shared cultural knowledge to decode, making them even harder for automated systems to catch. This behavior was not isolated to a few rebellious users; it was a collectively adopted strategy to navigate restrictive online norms.[2]

The failure of automated moderation has profound consequences for both platforms and their users. While algospeak empowers marginalized communities to discuss sensitive topics without being silenced, it also provides a shield for malicious actors. Cyberbullies and hate groups use the exact same phonetic substitutions to harass targets while evading detection.[1][4]

This creates a paradox where the very tools designed to protect users end up punishing the innocent and protecting the guilty. A creator discussing mental health might have their video suppressed for using clinical terminology, while a troll using coded slurs slips right through the filter. The algorithm's lack of contextual awareness makes it a blunt instrument in a highly nuanced environment.[1][6]

The limits of machine learning

Platforms are caught in a perpetual game of catch-up, constantly updating their blacklists with the latest algospeak terms. But the moment a new code word is banned, the community simply invents another one. "One, it doesn't actually work," said Evan Greer, director of the digital rights nonprofit Fight for the Future, regarding platforms' attempts to stomp out specific words.[1][3]

The rise of algospeak represents a fundamental shift in how language evolves in the digital age. Historically, slang developed organically within geographic communities or subcultures. Today, linguistic innovation is driven directly by the constraints of platform architecture, with algorithms acting as the primary catalyst for change.[3][5]

Linguists note that this is a form of Aesopian language, a historical technique used to bypass political censorship. But instead of dodging government officials, today's internet users are dodging lines of code. The speed at which these new terms propagate is unprecedented, with a single viral video capable of introducing a new euphemism to millions of users overnight.[1][3][5]

The psychological toll of censorship

The impact on marginalized communities is particularly stark. During the height of the pandemic, LGBTQ creators on platforms like YouTube and TikTok reported having their videos demonetized simply for using the word "gay." To protect their livelihoods, these creators were forced to adopt algospeak, substituting words or using emojis to keep their content visible and monetized.[1]

This forced self-censorship takes a psychological toll on creators, who must constantly monitor their own speech to appease an opaque algorithm. The rules of content moderation are rarely transparent, leaving users to guess which words triggered a shadowban. This guesswork leads to everyday algorithm auditing, where communities collectively test boundaries to map out the invisible tripwires.[3][6]

Even seemingly innocuous words can become casualties of this algorithmic crossfire. When platforms attempted to curb misinformation during the COVID-19 pandemic, users quickly pivoted to calling the virus the "Backstreet Boys reunion tour" or the "panini." The algorithm's inability to distinguish between harmful misinformation and genuine public health discussions forced everyone into the same coded vernacular.[1]

The future of online speech

Furthermore, the reliance on algospeak creates accessibility barriers for users outside the cultural loop. A teenager might instantly recognize that "leg booty" refers to the LGBTQ community, but an older user or a non-native speaker might be completely baffled. This linguistic fragmentation splinters the internet into highly specialized enclaves, where understanding the local dialect is a prerequisite for participation.[3][5][6]

For advertisers and brand managers, this linguistic arms race presents a significant challenge. Consumers expect brand environments to be free of offensive content, and 38 percent of respondents in the TELUS survey said brands should be able to identify and remove nuanced phrases immediately. But only a third believed brands were actually succeeding at this task.[4]

The reality is that building an artificial intelligence capable of understanding context, irony, and phonetic wordplay is an incredibly difficult computer science problem. Current large language models still struggle to parse the intent behind a deliberately misspelled word, especially when that misspelling is constantly shifting. Until machines can truly comprehend language the way humans do, the filters will continue to fail.[6]

Ultimately, the persistence of algospeak proves that human communication cannot be easily contained by rigid rules. As long as people have a desire to connect and share ideas, they will find ways to route around the digital roadblocks placed in their path. The algorithm may control the platform, but the users still control the language.[1][6]