At a conference in Europe in June, the ABC’s Artificial Intelligence specialist Gareth Seneque gave insights into the way the national broadcaster has been doing research and development on applying Generative AI and Large Language Models (LLMs)  at scale.

Seneque, who is the Machine Learning and Artificial Intelligence Engineering Manager at the ABC, told delegates to the ABU-Rai Days Conference in Italy that the ABC has built a new, modular, General AI platform. In a first principles rethink of the way the ABC does AI at scale, the team has developed a method called ABC Align for connecting this platform and the Generative AI tool built on top of it so that the system learns from editorial feedback, as well as the corporation’s content and AI principles.

“Taken together, this means we can solve existing problems like metadata and transcription, as well as do new things such as assisted information retrieval, editorial curation of algorithms and new ways personalisation, all with one system,” said Seneque.

Modularity is critical to the design of the new system. “We can easily swap in/out our own AI models, as well as those from Google, OpenAI and others… we can also add our own quality/safety layer on top. This design supports our commitment to both innovation and independence.”

The ABC has been experimenting with Generative AI for 4 years and started using GPT2 back in 2020, training the Large Language Models (LLM) on ABC content in its web content management system. Keywords are generated by a smaller LLM, fine-tuned on ABC content.

In the first experiments for ABC Assist, the Generative AI tool powered by the new platform, Seneque’s team has collaborated with the ABC’s Corporate Strategy group, taking their input and iterating it on the user-experience platform (UX) to develop a prototype.

“The step-change in industry capability last year meant we could go well beyond simply putting our content into the AI model… Prompts are everything, they are a way for data science and editorial to collaborate directly and tailor AI systems to the ABC’s specific needs,” he said.

In a demo displayed at the conference there was a measurable increase in both ‘reasoning’ (problem-solving), bias detection, and accurate question/answering capabilities.  Michael Collett, a journo-turned-conversational-designer, and one of the team’s data scientists Ari Kuperman worked on the project to develop prompts that keep Assist’s outputs relevant and factual.

A just published technical paper details how the ABC team has developed a way to align AI systems that is agnostic to the underlying open-source ‘frontier’ models from technology providers.  This is achieved by using a selection of data sources: ABC content, AI principles, and editorial feedback gathered from ABC Assist.

Rather than just using the input data directly, the nuance in the ABC’s model is that it uses  these data sources as input to “generate augmented datasets to elicit specific capabilities from AI models that we care about, like general problem-solving, accurate Q&A and the ability to detect bias.”

It is a two-step process to generate the dataset:

1. Generate Q&A pairs based on logical reasoning derived from content inABC Articles
2. Generate ‘good’ and ‘bad’ answers for each, where the measure of good/bad is relative to the ABC’s AI principles & canonical examples from the ABC Assist toolas feedback

AI models trained using this approach exhibit improved problem-solving AND bias detection capabilities – two capabilities that are often traded off over each other (the so-called ‘alignment tax’ where the ‘safer’ the model, the less capable it is).

The alignment of ABC Assist itself is done via the ‘editorial prompts’ that the ABC team has developed.

“All up this gives us a unique, scalable way of applying AI where we can exert a measure of editorial control over system outputs, keep our data safe and leverage it directly for AI alignment,” said Seneque.

ABC Assist is the tip of the iceberg for the national broadcaster.

The methodology outlined in the paper is the key enabler that allows the ABC to connect the cloud platform with ABC Assist in a feedback loop so that the system learns how to align with the ABC’s AI principles.

The technical paper specifies that the technologists make no editorial claims themselves, rather the aim is to enable the ABC to keep its hands on the wheel with AI systems and to facilitate experimentation within the organisation.

Once the LLMs have been trained using the ABC’s content, principles and feedback from the tool, the system can be applied to a range of use-cases by journalists, program makers and others within the ABC.

This solves the key challenge in using LLMs/Generative AI – how to efficiently apply the tech at scale. It makes use of human oversight and editorial expertise, which is the competitive advantage that media organisations have over large tech providers with no responsible editorial processes.

At last week’s Futurecast Conference, the ABC’s Chief Digital and Information Officer Damian Cronan  also spoke about AI, revealing that he has blocked OpenAI systems from feeding on the ABC’s website content data by including a no scanning message for AI Bots in the corporation’s robots.txt file instructions.

“Open AI said it would respect robot orders, so we took them up on it. Our content is an asset that is valuable. There is an incredible amount of work by our journos, the idea that it can be used by Large Language Models devalues that and it’s my responsibility to enforce that… Trusted content is now of incredible value,” said Cronan at Futurecast.

Cronan views is as an issue of power dynamics between broadcasters and tech platform providers: “With the internet we are staring at social power dynamics and other social changes. We need to go in with eyes open, but we can’t define too many limits because we also need a level of ambition to use it to improve or assist our workforce.

“When there are too many choices, if information is too complicated, consumers will revert to known and trusted brands. For media this is the editorial standards that companies hold themselves to and explain to their audiences.”

Related Articles:

The next generation of AI Broadcasting: ABU-Rai Days Conference

Broadcasters name AI as the industry’s number one priority: ABU-Rai Days Conference