 is a tort or crime that, in common law jurisdictions, entails harm to persons caused by the publishing of statements that are untrue, leading to reputational damage for the claimant.

 is not a legally defined concept but rather an umbrella term for language that could be perceived as discriminatory or threatening. 

At CaliberAI, we use harmful content to denote either: 

 (language that targets people on the basis of their gender/religious/national/racial/sexual identity) or 

Although harmful content is not a legal construct, the publishing of harmful content can often result in reputational damages - and financial losses - for multiple parties.

CaliberAI's machine learning systems analyse text with respect to its syntactic, lexical and semantic features, before outputting a probability score indicating that the text belongs to one of our relevant categories (Defamatory and/or Harmful or Neutral). 

The systems are almost entirely pre-trained on 

, carefully assembled by domain experts, with extensive experience in the management of public debate, so that they learn to recognise the language patterns of defamatory and harmful content.

how-it-works.webp

Decisions to classify a sentence as 'defamatory' or not are made by our 

, based upon the presence or absence of certain 

 (e.g. named entities) and other criteria.

Our annotators training is heavily informed by the advice of our 

 and others with unparallelled experience in 

, which is to say each sentence is classified on a stand-alone basis. Soon, new iterations of our tools will analyse language at the document level and take into account dependencies between sentences.

There is no technical limit on the overall size of a document CaliberAI's API is capable of processing, but it works best on short snippets such as a 

, which will analyse the entire document and return classifications for every sentence.

One of the core principles of linguistic theory

“know a word by the company it keeps”. CaliberAI’s tools use state-of-the-art “

” models which are numerical representations of “words in context”. 

That context is largely given by the statistical tendency of a given word to “co-occur” with other words in large web-based corpora (e.g. Wikipedia).

CaliberAI does not currently leverage specific jurisdictional-legal information (e.g. results of litigation) but will do so in the future. For the moment, our machine learning tools identify the 

 that could potentially trigger a defamation suit in 

. This partly entails training our algorithm on customer data, and in the process tailoring the solution to the specific publisher. 

Our data annotation team and Advisory Panel reflect a spectrum of identities and professional and cultural backgrounds (read our 

), whereby each person brings a different perspective to the task at hand. 

Our research and data collection activities have leveraged studies of hate speech in many different regions around the world, including Africa and The Middle East.

CaliberAI's technology has been built to be as frictionless, user friendly and easy to integrate as possible, and can be deployed in multiple ways. 

 based system for classifying entire articles, as well as cross-platform integrations (e.g. Twitter).

 to set up an account for any of the following products.

 in your browser and receive instant, intuitive in-browser insights.

online-article-evaluator.webp

Email whole articles to CaliberAI and risk-assess for the presence of defamatory or harmful material. 

whole-article-evaluator.webp

Protect yourself from harmful and defamatory mentions on Twitter.

twitter-moderator.webp

CaliberAI stores all text provided by the user, which is subsequently retained for analysis for 

CaliberAI indefintiely retains insights derived from such analysis for billing, support, and performance purposes. 

CaliberAI uses industry standard encryption where appropriate (e.g. HTTPS, TLS, SMTPS etc.)

The publishing of direct quotations with defamatory content is 

 under defamation law generally. The publication of a defamatory statement constitutes a legal defamation regardless of whether the same statement is attributed to a third party or not. 

Furthermore, certain publishers have historically had occasion to misquote third parties or take quotations out of context, thereby triggering defamation suits.

CaliberAI therefore does not treat the presence of quotation marks as a useful signal to be captured in our classification pipeline.

Algorithmic explanation (or 'XAI', as it's often termed) is an increasingly important part of AI regulation, particularly in the EU, under Articles 13-15, 22 of 

 for example, as well as Recital 38, The Artificial Intelligence Act. 

At CaliberAI we have optimised our AI systems to be as XAI compliant as possbile. We do this through the use of 'attention weights', best understood as 

a mechanism for directing an AI system’s focus

In CaliberAI's case, our use of attention weights forces our classifier to show a user which words have contributed the most to the probability score of the classification. 

You can read more about this in our blog post, 

attention-weights.png

At CaliberAI, we understand that each internet platform's 

, and that striking an appropriate balance between effective editing and freedom to publish isn't easy. 

We therefore offer algorithmic ‘confidence thresholds’, which can be thought of as the degree of confidence with which the algorithm makes a classification. 

For example, as can be seen in the illustration below, CaliberAI's systems are 

"He was holidaying in Spain simply to avoid the corruption affair that is engulfing the enterprise"

Frequently Asked Questions