Toxic Telegram messages: what automated checks miss
Distinguish profanity, personal attacks, quotation and repeated harassment. Build contextual tests and a human review process without assuming a toxicity feature exists.
“Brilliant work” can be sincere praise or a public sneer after someone makes a mistake. A tool that receives only those two words cannot recover the preceding exchange from nowhere. A person may also need nearby messages and an explanation before deciding what happened.
Automation is useful for a specific rule such as a clearly defined vocabulary restriction. Judging behavior takes more care. Defendy’s AI checks target spam; they do not provide a separate toxicity detector, a conflict-severity score or an automatic queue of disputed conversations. You can still build a workable moderation process around the controls that exist.
Replace “be less toxic” with observable rules
“Toxicity” is often shorthand for several different problems. A profanity rule concerns vocabulary. A personal attack concerns how someone addresses another person. Harassment may involve a pattern that is hard to see in one message. Strong disagreement can be uncomfortable without breaking a rule.
Separate at least these four cases:
- Prohibited vocabulary. Using the expression itself violates a published rule. Explain whether quotations and discussion of the term are exceptions.
- A personal attack. The post demeans a participant rather than challenging the argument. Identify the target and the relevant wording.
- Repeated targeting. Individual replies may look mild, but the sequence and refusal to stop change the picture.
- Forceful disagreement. Criticism of a proposal, product or decision is not automatically misconduct.
“Do not insult other members or continue directing jokes at someone after they ask you to stop” gives a reviewer more to work with than “no toxicity.” Concrete threats or disclosure of someone’s private information need timely human attention; do not wait for an imagined model score to cross a threshold.
Ask what the isolated sentence leaves out
The following examples are fictional. They illustrate questions for a reviewer rather than universal verdicts.
| Text | Possible context | What to check |
|---|---|---|
| “Brilliant work” | Thanks for a fix, or mockery after a mistake | The reply target and preceding exchange |
| “They called me an idiot” | A report quoting an insult | Whether the author is reporting an attack or renewing one |
| “Nobody asked you” | A rude reply, or one of many aimed at the same person | Repetition and previous requests to stop |
| “This proposal makes no sense” | Blunt criticism of an idea | Whether there is a personal attack beyond disagreement |
Context does not automatically excuse behavior. “We are friends” does not override a request to stop. A prominent contributor should not have a different standard from a newcomer. Conversely, a moderator’s dislike of a viewpoint is not evidence that its author violated a conduct rule.
What the research actually supports
In an ACL 2020 study, John Pavlopoulos and colleagues examined Wikipedia comments with the preceding post and discussion title. Context could increase or reduce perceived toxicity. The tested context-aware classifiers did not show improved performance. The useful caution is that supplying more conversation does not automatically solve the classification problem. This was not an evaluation of current models or Defendy.
Paul Röttger and colleagues introduced HateCheck at ACL 2021, using functional tests to expose weaknesses hidden by aggregate hate-speech detection metrics. The transferable idea is to test particular distinctions, not just collect one overall score. Hate speech and everyday rudeness are different categories; the paper’s results do not establish the performance of a particular Telegram bot.
Use Defendy’s controls for their actual jobs
Profanity filtering uses a selected dictionary and exceptions. The current interface provides system, custom and combined dictionary modes, media-caption filtering, message deletion and a member action. A vocabulary match does not prove harassment. A lack of vocabulary matches does not establish that an exchange is harmless. See the profanity guide.
Stop words let you define your own lexical conditions. Choose the matching method and consequences around a specific rule, then test ordinary words, quotations and permitted discussion. A broad exception can miss abuse; a short substring can affect innocent language. The stop-word guide explains the supported methods.
AI spam checking answers a different question. Automatic verification receives limited message text and a group title, rather than the complete conversation. Its instructions distinguish ordinary questions and opinions, including rude ones, from spam. A rude post passing this check is therefore not proof that the product attempted and failed a toxicity assessment. The manual /spam path also targets spam and can initiate deletion; it is not a harmless dispute-review command.
Test distinctions with paired examples
Use a private test group and invented examples. For each pair, change one feature that matters to your policy:
- a direct insult versus a report quoting that insult;
- an attack on a person versus criticism of their proposal;
- a single comment versus a sequence after a request to stop;
- a statement versus its negation;
- a neutral reference to a group of people versus hostility toward that group.
Before testing, ask two moderators to label the expected outcome independently and identify the applicable rule. Discuss disagreement. If the rule cannot resolve a case, improve the wording or leave it for human review rather than declaring one label objectively correct.
Check false positives, such as a complaint removed because it quotes an insult. Also check false negatives, such as repeated mockery that uses no prohibited word. Include the languages and ways of speaking your community uses. Results on a few English examples do not establish performance in another language.
If a lexical filter treats both halves of a pair alike, ask whether that is the intended policy. An absolute ban on a particular word may do that by design. A claim about someone’s intent needs different evidence. Do not describe string matching as conversation understanding.
Make the human process specific
A compact case note in the moderators’ private workspace can contain a link to the available message, the minimum necessary context, the relevant rule, a decision and the reviewer. This is an organizational example, not a built-in Defendy review feature. Avoid forwarding complaints and personal details to a public log.
The reviewer checks the sequence, separates reported abuse from a new attack and chooses an intervention consistent with the published policy. Missing context may justify asking a question. An active conflict may require a permitted temporary intervention while review continues, but its scope and reason should be clear to the team.
Give members a contact for appeals. A second moderator, where available, should evaluate the facts and rule rather than simply endorse a colleague. Correct a mistaken restriction separately from the content decision. A removed message may need to be posted again; do not promise automatic restoration.
A configured Defendy moderation log helps with selected events. It is not a complete archive of the conflict or an automatic complaint queue. Keep only the personal information necessary for the review, with limited access, and avoid sending the entire dispute to an external model for convenience. Check the data-processing information before relying on that workflow.
Know when the process is working
Track what you can observe: overturned decisions, repeated errors involving quotations, unresolved appeals and disagreements about the same rule. These are team review notes, not claims about a dedicated product dashboard. An absence of complaints is weak evidence if members do not know where to appeal.
The most useful final test is whether you can tell the member which behavior violated which rule and who can correct a mistake. If the explanation stops at “the system decided you were toxic,” the decision is not yet sufficiently grounded.
Sources and feature boundaries checked on 5 October 2026. All examples are fictional; this article provides no measured Defendy performance or universal toxicity threshold.
