Can synthetic data help to keep people safe online?

Artificially created images, text and videos can be used to teach AI systems to recognise harmful content

Wednesday 21 Feb 2024

Keeping people safe online means putting in place systems to protect them from harmful content, such as hate speech, violent imagery, or material showing child sexual exploitation and abuse. Online platforms can now use automated technologies to help with reviewing of harmful content. However, there’s a catch: these technologies must first be trained, by humans, to spot the very content they are designed to protect us from. During this training process, human ‘trainers’ may be exposed to vast quantities of harmful content. One potential solution to this ethical dilemma is to replace real examples of harmful content with artificially created ones – a form of ‘synthetic data’.

At The Alan Turing Institute, we have been assessing to what extent synthetic data can help to provide safer online spaces, as part of a project commissioned by the UK government. Here, we explain more about what we mean by synthetic data and how recommendations from our new report could help those aiming to create safer online spaces.  

What is synthetic data and how is it generated?

Synthetic data is artificially created information that is designed to replace, complement or fill a gap in real-world data. It shows great promise in critical sectors like healthcare, where detailed patient data is needed for modelling and to improve treatment outcomes but real-world data risks compromising privacy.

Synthetic data can refer to simple, tabulated data that has been artificially created, but more complex types of synthetic data, like text, audio and visuals, are often created using generative algorithms – in a similar way to conversations created by chatbots or artwork created by AI image generators. Generative AI creates new content by learning the underlying patterns in real-world data. So, it can learn to create artificial examples of, for instance, violent imagery, which can be used for training AI models how to recognise harmful content.

How can synthetic data be used to improve online safety?

Human trainers or ‘annotators’ must manually mark examples of various categories of content to train AI content moderation systems and can therefore be exposed to thousands of pieces of potentially harmful content, putting them at risk of psychological harm. By contrast, synthetic data that is used to replace or augment real harmful content can be generated ready-labelled, making it possible to limit the vicarious trauma experienced by humans.

Used responsibly, synthetic data and the technologies that generate it can also help expedite research and development in online safety by providing more accurate simulations of scenarios where online harms might occur. For example, synthetic audio can be used to train AI models to spot abusive language in video game environments, so that regulators and government can provide better guidance to online gaming platforms. Synthetic data can also be used for testing of changes to how content moderation systems work, reducing the need for testing by real users and potential exposure to harmful content.

Why does this matter right now?

The use of synthetic data for training AI models has increased rapidly in the last three years. According to some estimates, we are likely nearing a situation where most training data is artificially generated.

Meanwhile, countries all over the world are bringing in legislation to help keep people safe online. The EU’s Digital Services Act came into force in 2022, whilst the UK’s Online Safety Act passed into law last year. Under the new regulations, social media platforms must undertake comprehensive risk assessments and make significant improvements to the AI models they use to protect users on their platforms. Synthetic data will likely be key in helping platforms to meet these obligations and in meeting ethical standards for online safety.

Are there any potential drawbacks of using synthetic data?

Whilst synthetic data can be useful for ‘de-biasing’ real-world training data that under-represents harms to certain demographic groups, the generative models used for creating synthetic data can themselves be biased. For example, when prompted to generate images of a nurse and doctor, AI image generator DALL-E produced synthetic photos of a white female and white male, respectively. This issue needs clear attention from researchers and practitioners before the promise of synthetic data for enhancing online safety can be fully realised.

Since synthetic data is not linked to specific real-world individuals, it should carry a lower risk of compromising privacy or revealing sensitive information. However, risk levels need to be clarified, as it is often not clear to what extent real-world data ‘leaks’ into synthetic datasets, particularly for synthetic images, audio and video.

How can we make responsible use of synthetic data for tackling online harms?

Our new report – made possible thanks to expert guidance from the Department for Science Innovation and Technology, the Centre for Data Ethics and Innovation, and Ofcom – provided an important opportunity to bring together minds from across the Turing and the online safety ecosystem more broadly. We explored how synthetic data can be responsibly utilised in practice and made some recommendations for responsible innovation (read the full list here), including:

  • Tracking the origins of synthetic data (in real content) to ensure its trustworthiness.
  • Measuring potential data leakage into synthetic datasets to preserve privacy.
  • Conducting bias assessments of synthetic datasets to ensure fair representation.
  • Controlled release of synthetic data and the models used to generate it.

More generally, the field needs to reach a consensus on how to evaluate the quality of synthetic data. We hope to see these positive steps being taken, so that synthetic data can realise its promise to help digital platforms tackle online harms.

Read the report:
Exploring responsible applications of synthetic data to advance online safety research and development

 

Top image: v_sot