Â鶹Éç

Artificial intelligence is changing how news stories are sourced, shaped and distributed. The Â鶹Éç Research & Development natural language processing (NLP) team is building practical tools that help colleagues write faster, more consistently and with fewer mistakes, while keeping editorial judgement firmly in human hands.

We’re also exploring how AI tools can strengthen the Â鶹Éç’s reputation for accuracy and trust. Our research includes fine-tuning large language models to comply more closely with Â鶹Éç editorial and style guidelines.

Â鶹Éç style guide checker

The newsroom has changed how specialist subediting integrates into its workflow, and journalists and editors now check their articles against over 1,200 rules in the Â鶹Éç News style guide.

We’re researching whether natural language processing techniques such as named entity recognition (automatically finding people, places and organisations in text), regex pattern matching and AI can help address this issue. These techniques will enable a tool that rapidly checks for errors before publication. It’s designed to work within existing production tools, and we’re currently working with teams across the Â鶹Éç to ensure it meets editorial needs.

Â鶹Éç style assist

The , a Â鶹Éç-funded initiative, employs around 150 journalists at regional newspapers around the UK. They supply over 250 stories a day, but many go unused, partly because of the into the Â鶹Éç house style. Rewriting typically involves editing for length, structure, flow, and language, and can take around half an hour per article.

The Â鶹Éç style guide tool uses AI to redraft articles so they comply with the core style requirements, allowing journalists to focus on refining and polishing the final version. The tool has been through several trials and is now being integrated into Â鶹Éç News’s production tool. We expect it to be available for our journalists to use this year.

Â鶹Éç News live

Live news text streams are becoming one of the most popular ways for audiences to engage with Â鶹Éç journalism. There are typically more than 25 live events every week across news and the nations and regions, each involving contributions from up to a dozen journalists working in a time-pressured environment. A significant portion of that effort involves manually transcribing TV and radio broadcasts to extract relevant quotes quickly enough for inclusion on the live page.

We’re working with the news AI team to streamline this process. With accurate transcription and AI-based analysis, we can automate much of the quote-finding and context extraction, provided we can reliably align transcript timings with the output of the large language model (LLM), the audio and other elements. Â鶹Éç R&D’s time addressable media store (TAMS) offers a strong foundation for synchronised, time-linked content retrieval of this type. We have begun a three-month prototype exploring how speech-to-text, AI analysis, and time-aware media infrastructure can help journalists focus more on editorial decisions rather than laborious transcription tasks.

Â鶹Éç LLM

Commercial large language models are powerful but can be costly to use at scale and often fall short of Â鶹Éç style requirements. We’re developing a fine-tuned model that aligns more closely with Â鶹Éç editorial values. It’s designed to handle time consuming newsroom tasks such as rewriting, headline generation, tagging, and summarisation. Once trained, it can also evaluate outputs from other AI systems for editorial compliance, style adherence, and tone.

We have not built the model from scratch. Instead, we are fine-tuning open-source and open-weight models using Â鶹Éç data. To ensure alignment and safety, we’re combining techniques such as instruction tuning, constitutional alignment and preference learning so Â鶹Éç editorial guidelines can be used to directly shape the AI output.

Evaluation work and impact

Building trust in AI means putting evaluation at the centre of how we design, explain and deploy our tools. Over the past year, we’ve developed a mix of machine-based and human evaluation methods to assess accuracy, reliability, and stylistic fit.

In the Â鶹Éç style assist project, for example, we worked with journalists in the nations and regions to define clear accuracy measures. We assessed hallucinations, false assertions and misquotations by comparing AI-generated sentences with human-written equivalents. Independent assessors then forensically reviewed the component parts of 2,400 AI-generated sentences to determine whether the source material supported each claim.

Alongside quantitative checks, we developed qualitative style measures around flow, structure, tone, and clarity to judge how well AI outputs match Â鶹Éç house style. This also let us test how well our automated metrics predict editorial fit.

However, these measures only matter if they translate into making our colleagues’ work easier. So, the bigger picture for our team is always impact, expressed in practical terms such as:

• Minutes saved per article (with no loss of quality).
• Average number of style inconsistencies across Â鶹Éç output on a weekly basis.
• Hours saved on transcription, freeing capacity for original journalism.

Collaborate

We collaborate closely with colleagues across the Â鶹Éç and also work with world-leading academic groups on our foundational research challenges. These examples are just a small sample of our current work, and we hope to be able to share more with you soon.

If you’re interested in learning more or exploring a collaboration, please contact us at airesearch.mgmt@bbc.co.uk