One of the most prevalent stereotypes about generative artificial intelligence is that they cannot be used to search and delve deeper into topics because they tend to hallucinate.

Hallucination, despite its human-associated implications, is a technical term used to describe when text-to-text generative artificial intelligences, such as ChatGPT, Gemini, or Claude, fabricate information. Although many researchers are striving to minimise this phenomenon, it may be nearly impossible to completely eradicate these errors. The propensity to hallucinate appears inherently linked to the nature of these machines – and, if you think about it, to human behaviour, creativity, and intelligence as well.

In some instances, the issue is addressed by essentially censoring the machine. For example, if a large language model like Gemini lacks sufficient information about a topic or a person, it might respond: “I do not have enough information about that person to help with your request. I am a large language model, and I can communicate and generate human-like text in response to a wide range of prompts and questions, but my knowledge about this person is limited. Is there anything else I can do to help you with this request?”

Alternatively, to ensure that a chatbot has accurate information about a topic for content creation assistance, we can personalise and fine-tune the information available to the chatbot. For example, I have provided my ChatGPT with personal details so that they become part of its knowledge base. When asked about myself, my personalised ChatGPT responds appropriately.

My personalised ChatGPT answers to a request about myself. Please note that this information may not be available elsewhere and that the answers vary, since ChatGPT is still a probabilistic machine, not a deterministic one

Ask for search, source quoting and linking

When using these tools to enhance our ability to search the web for content, a robust technique is to compel them to provide answers by scanning the web for very specific content while quoting and linking sources. For instance, I have used my chatbots to conduct specific searches, such as querying ChatGPT for PDF documents related to European cohesion policy.

It is crucial, however, to verify the sources, for example by simply clicking on the links provided: sometimes even the links may not exist at all.

Despite this problem, the primary advantage over a simple advanced search using search engines is that the machine also provides a brief summary of the document’s content, offering a glimpse into the information before a full review. If the response is insufficient, you can always request more detailed information in the same chat.

This method can be employed with other file types, like CSV or XML.

You may also want to compare results across different chatbots, such as Gemini, as I did.

Once you begin, you can delve deeper. After locating the documents you need, you can use them to enhance the chatbot’s knowledge base and analyse specific document sections, data etc., maintaining the conversation in the same chat.

Define the search query

With these advanced tools, you can craft complex search queries that surpass simple keywords. For example, instead of a generic search for “effects of climate change,” you might ask, “What are the recent studies discussing the impact of climate change on Arctic marine biodiversity?” Contextualising your query – stating whether you are writing for a magazine or newspaper, the type of article you are preparing, and any preliminary research you have conducted – enhances the specificity and relevance of the results.

Process the query results and generates more context-aware results

From the initial results, you can pose follow-up questions to the AI to refine the search further, brainstorm ideas, or directly generate a summary of the findings.
For instance, you could ask to summarise recent scientific papers on the specified topic, highlighting key findings and data trends, and obtaining more results.

You may also start working on a first draft and provide it to the chatbot, asking for adding more advanced search ideas in a sort of brainstorming-and-search process with the machine.

It’s possible to use chatbots allowed to search the web to verify claims and identify misinformation: by cross-referencing with multiple sources, the platform helps in identifying biassed or fabricated content, helping you find fact-checking articles or independent sources too. This may speed up the process of fact-checking, and could help reduce human error and bias (please be aware that the machines have biases too: the human ones introduced by design and training).

Using specific AI tools for search

Journalists may find specialised vertical AIs, specifically designed as enhanced search engines. Tools like Perplexity AI are tailored also for journalistic needs.

These tools not only search for textual content but also explore videos and images, automatically quoting and linking sources. Once Perplexity accurately understands the query, it displays all the sources it utilised.

*An example of search with perplexity related to a topic I’m working on. The tool also offers images and videos, on the right*

The response also includes a detailed summary of these sources, with specific links to paragraphs for easy reference.

*How summarisation works in Perplexity. This is just the first part of the answer*

Journalists can even use Perplexity to generate ideas, starting from broad questions. In its Pro version – which is not free – the tool may prompt users to specify their query if too broad initially.

For example, when inquiring about ideas for an article on the so called “smart control rooms” across Europe, maintaining the conversation in the chat I’ve started asking for sources, Perplexity asked me to define a specific angle for my article, offering suggestions based on my input.

It’s important to say that I’m working on the topic and I created a documentary about Venice’s smart control room: so, I know where I want to go, but I still need some hints: this is crucial to understand that if you already know a topic, these tools are perfect to assist you in improving your knowledge, your findings and even your coverage, exploring different ideas.

If you want to start completely from scratch, it may be difficult to obtain good results at the end of the process.

I focused on privacy risks and social control, ignoring the suggestions and working on what I’m interested in. Perplexity then provided me with 19 relevant sources.

It’s also possible to expand the table under your request to see the activity conducted by the tool.

In the end, the tool comes out with a series of possible arguments to explore with my work.

While Perplexity offers suggestions, as a journalist well-versed in your subject area, you have the autonomy to decide whether to follow these suggestions or explore other avenues through traditional journalism methods such as direct interviews or on-site investigations.

Final recommendations

Incorporating these AI-assisted routines into your journalistic practices can significantly enhance efficiency. However, it is crucial to remember that all parts of this process require human verification. While AI tools can expedite information gathering and facilitate brainstorming, they must be used wisely. Understand their strengths and weaknesses and supervise their operation. Consider these AI tools as highly specialised, quick, and capable assistants – they are like juniors who cannot work independently, publish content, or lead journalistic efforts on their own.

Emphasising human oversight is always needed when speaking of AIs: it ensures the integrity of the journalistic process, maintaining accuracy and ethical standards while utilising advanced technological tools.

Source of the cover photo: Nat via Unsplash

Perplexity

Latest in Perplexity

AI course: using AI for research and news gathering

Generative Artificial Intelligence for advanced search online