When search engines were invented, they simply scanned the internet and came up with a list of the most popular, relevant sites. Then came ‘generative AI’, which creates new text and images from the search results by accumulating vast quantities of text or images. It uses this to find answers, not as a human would do, by understanding the question, but simply by identifying patterns in the data. Many of us now routinely use this without even being aware of it. It is built into Google and Microsoft Office, and most of the time, it is pretty reliable. It is predicted that worldwide, nearly 400 million people will use AI this year, a 20% rise on 2024.
We are now moving into the world of ‘agentic’ AI, and ‘Deep Research’.
Agentic AI
‘Agentic’ AI does not merely answer questions. It takes text, images, sounds, and other forms of data; manipulates them with a range of tools; collaborates with other AI agents, and learns from past experiences to improve over time. Sixty percent of advertising agencies already use AI to create advertisements, and agentic AI is expected to be able to manage entire marketing campaigns: creating, testing and adjusting creative materials automatically to maximise audience engagement. Meanwhile, such AI agents are already curating news feeds based on user preferences, summarising lengthy articles and generating video highlights for breaking news events. In Italy last month, we saw the first magazine edition produced entirely by AI, based only on prompts from journalists.
Although it has only just begun to appear in business software, experts are predicting that agentic AI will be in a third of such software within three years, enabling automation of workflows and replacing websites and apps and workers within a decade.
Deep research
‘Deep research’ tools – Open AI’s Chat GPT, Google’s Gemini and the latest arrival, Perplexity – apply these techniques to more and more current data, including research papers, databases, and live web data. They break down complex queries, check their understanding of the question, analyse data, and synthesise the results into concise reports, with summaries, citations and visuals. They can identify changing trends and deeper context. In minutes, a journalist covering a new social movement can find historical context, identify changing public attitudes, social media trends and the views of key players. Each has its own strengths, as the table shows.
How does deep research perform in practice?
We decided to test Perplexity on a controversial issue which we have covered in the past in East Anglia Bylines. We asked this question:
“There is controversy about how to connect North Sea windfarms to the national grid. What are the arguments, and what is public opinion?”
It took Perplexity four minutes to generate an 1800-word report, breaking down the task into three stages, and reporting what it was doing at each stage. It consulted and listed 47 sources, all authoritative and relevant. They included government, media, academic and industry sites, and political documents from several parties. A human researching the same question could have found them all, but pressed for time, would probably not have done so.
The report’s subheadings give a sense of its coverage:
Grid connection and policy disputes
Delays and queue issues
Regulatory responses
Competing Infrastructure Proposals: Onshore vs. Offshore Solutions
Onshore infrastructure approach
Offshore grid alternative
Community and environmental concerns
Local opposition and environmental impact
Economic and tourism concerns
Political dimensions
Public opinion on wind farms and grid connections
Support for wind farm development
Information gaps and communication needs
Balancing competing interests: the path forward
Coordinated approach
Regulatory reform and streamlining
Conclusion
References
It was a solid report, covering the main issues at least as well as a human journalist would do. But there were weaknesses. Greenpeace was the only site associated directly with protest groups, probably reflecting the fact that such groups have a smaller online presence, or perhaps were regarded as less authoritative. However, the report records significant public opposition in Suffolk, citing an MP, the chairman of Aldeburgh Museum and ‘a local campaigner’, though curiously, the word ‘pylon’ does not appear anywhere. It was thin on the offshore grid options which have attracted much local debate.
At first glance, the report appears thin on public opinion surveys. The only one cited was about onshore windfarms in Northumberland, which was not strictly relevant to the question, probably because the question did not make it clear that ‘North Sea windfarms’ meant offshore ones. But a Google search suggests that surveys on attitudes to renewable energy never ask specifically about the most controversial issues in East Anglia – pylons or grid connection. There was also a problem with outdated information on the source sites, with two references to a former minister and a retired county councillor, as if they were still in post.
Issues
The test of Perplexity confirms that Deep Research AI can produce very competent reports extremely quickly. But some issues have been widely noted. Where there are gaps in data, such tools can produce ‘hallucinations’, creating plausible, but entirely fictitious, reports. Similarly, the AI can only read what is there, and available online data will tend to reflect the cultural interests and biases of the world when it was created. And for some questions, there will be biases on issues like ethnicity or gender. So, the output continues to need checking by a reasonably informed editor. And as with any research, question-wording is critical.
There are also environmental issues. AI needs vast computer power to process data, so the data farms which provide this require a lot of electricity. A planned new datacentre in Slough will consume twice as much electricity as Heathrow airport at its peak. They also need clean water for cooling, typically two million litres a day, equivalent to the water consumption of 14,000 people. In a dry region like East Anglia, where water is already scarce, this is a significant figure.
And, of course, there are employment issues in any industry dependent on written or visual material. A vast range of jobs will be affected. Some may become richer, but many will just disappear. But this process is probably unstoppable. Only where individual creativity is highly prized may people escape the effects.
How are we using AI at East Anglia Bylines
AI is an invaluable tool, when used with care. At East Anglia Bylines we use AI as a research tool, to identify sources and evidence, or to summarise long or complex documents. We sometimes use it to suggest or refine headlines or tweets.
But everything we publish has been written by a human. We do not publish unedited AI generated material. And in view of the intellectual property issues, we do not use AI generated images.
This article is based in part on a webinar organised by Statista, who provide a global resource of verified statistical data
More from East Anglia Bylines
