An experiment with online analysers and AI

Last month, a global gathering ‘to connect, imagine and rewire how change happens’ came to a sort of end.  Called the Networks Festival, it ran for 6 weeks and involved about 1500 people in 35 different events, all gathered into a knowledge hub to watch again.

But as so often, the most dynamic stuff lay in the spaces between events when participants talked among themselves. Only, because we weren’t all in one place, the corridors, cafes, and fringe spaces were all actually on WhatsApp.

The organisers (Greaterthan and Fito) realised this and brilliantly created a WhatsApp Community with a range of group spaces for people to gather and generate the buzz.  

If you’re an organiser, keen to spot the patterns and trends, or to capture and good outcomes, then this setup offers an opportunity to find out things you’d have little hope of discovering at a big, sprawling in-person event. 

So, keen to see what insights I could gather, I ran a little test on the ‘Cheerleaders/Knowledge Harvesters’ chat, a minor sub-group with just 78 members.  

This group shared just 5 photos, 19 links and 3 docs compared to the 263 photos/videos, 246 links and 8 docs in the main group. Either would have been manageable research tasks, but I wanted to see what would get served by the web if I didn’t have the time or perhaps lacked the skills to do that work for myself.

So the first thing I did was export the chat from my mobile, which you do by clicking the group icon and scrolling to the bottom of the group info. page (past the members) until you see three green and three red options.  When you choose ‘Export chat’, it asks if you want to include media or not and then produces a zip file which you can send via one of the options on your phone (I emailed mine to myself).

I then looked up ‘WhatsApp analyser’, which immediately produced 9 options on the first page alone.  I tried three at random, and found they all did pretty much the same thing.

Here’s an image from WhatsAnalyze which shows how these apps pick out the message stats – how many, when, who, and how often along with a snap of keywords/emojis.  

It’s useful for showing up group dynamics and behaviours, but to understand why the patterns look the way they do you’d still need to match the stats to the content. If you’re aiming to mine the conversations for value in the content, they’re not much use. 

 So I turned to AI, asking Gemini (google’s version of Chat GPT) ‘what are the trending topics in this chat?’

This was much more interesting – turning up an infographic graphic summary and a 14 page report.  The graphic summary showed the chat lifespan, type and mandate, picking out the ‘content creation workflow’ and topics ranked by engagement. Most usefully, it also identified three ‘chalenge/pain points’ and three ‘Opportunities & Future Growth’ points.  All of this would be helpful for organisers as part of an event review.   

But the full report is where AI comes into its own.  The results themselves aren’t important (I chose a group with limited scope) – it’s more the analytical possibilities I was interested in.  

Bearing in mind I only asked a single question about trending topics, and could have gone on to ask for more – here’s what I found.

I got served a well structured paper in five sections, including 

• a summary up front
• an introduction setting context
• a descriptive answer to my question
• Some insights into the subject matter
• Some emerging challenges and opportunities, and
• A conclusion with strategic implications for the organisers, including some recommended actions

The summary was pretty good, although it did show up a limitation of this kind of analysis.  By highlighting the thread with the highest volume of chat, it inferred (without actually saying it) that it was the most important. In fact, this wasn’t the case, but a reader who wasn’t there wouldn’t know that.  

On the plus side, it offered a nice reflection on the reasons the chat worked well (structured facilitation and organic peer-to-peer support) as well as a pat on the back for taking a thoughtful approach with a view to sustaining engagement for the long-term. 

The intro and context were spot on – drawing directly from group info and messages spelling out the purposes by the group admin.  That said, it did also note the ‘highly deliberate and well-defined community structure’, pointing out how important clearly defined roles and objectives are ‘for effective engagement and output’. So the old adage ‘rubbish in, rubbish out’ applies, and if a group is vague in its setup, it’ll get vague results when it comes to review. 

On trending topics (the question I’d actually asked) the report was accurate. It picked out the central theme (developing skills for creating impactful content), and the focus on the practicalities of using platforms well for effective outreach. Again, the analytical model offered unsolicited insight, noting group behaviour was collaborative and supportive in a peer-to-peer learning environment.

It also referenced resources and real-world examples which we’d used, so even though I’d only asked for ‘trending topics’, what I actually got included knowledge and information to share. Productive exchanges were summarised with a snapshot in table form too. Here’s part of it: 

Table 1: Key Trending Topics & Engagement Overview

Topic	Key Contributors	Message Count (Approx.)	Shared Resources (Links/Docs)	Engagement Type
Content Creation & LinkedIn Strategy	Agnetha,  Bjorn Benny Anni-Frid	20+	Google Docs, LinkedIn posts, LinkedIn article on algorithm 1	Collaborative Drafting, Feedback, Best Practice Sharing, Algorithmic Discussion
AI for Collaboration & Systems Thinking	Dolly Parton, Kenny Rogers	15+	Gregory Bateson’s “Mind and Nature” (quote), LinkedIn post 1	Philosophical Debate, Critical Engagement, Tool Application

 

This is gold dust.  The kind of thing that happens at the fringes and is rarely captured at in-person events. Left in the WhatsApp chat, this information would soon disappear from sight. When it’s summarised and classified like this, it can be fed into online resource bases like the ‘knowledge hub’, a shared google drive, or a platform like Hylo for easy access by the community.   

Not done yet, the AI analysis then looked at the ways the group interacted in order to assess how effective it had been. It homed in on Adrian Robke, our chat host who’d led a session on structured, step-by-step workflow for content creation. AI loved Adrian.  It produced another table which I’ve duplicated in full so you can see the potential of chat analysis to produce something that wasn’t there before.  

Table 2: Content Creation Workflow & Tools/Platforms

Workflow Step	Description	Tools/Platforms Mentioned	Purpose/Benefit
Session Selection	Choosing a festival session to write about, checking existing content.	networkshub.sutra.co, docs.google.com/spreadsheets 1	Content sourcing, topic identification, avoiding duplication.
Drafting	Writing the initial content, incorporating AI summaries as a foundation.	Google Docs (or other open doc), WhatsApp group for sharing 1	Collaborative writing, version control, easy sharing for feedback.
Feedback	Sharing drafts with the group and receiving constructive comments.	WhatsApp group for sharing and comments 1	Peer review, quality improvement, diverse perspectives.
Publication	Incorporating feedback, editing, and sharing the final post.	LinkedIn, Substack, Newsletters, Slack channels 1	Dissemination, audience reach, knowledge sharing.

 

It concluded that spontaneous partnerships, constructive criticism, and practical support in the chat was proof of potential for temporary, task-oriented groups to evolve into self-supporting learning networks. 

Challenges, opportunities and strategic implications came in the last two sections of the report. It pointed out three minor friction points for the event organisers, suggesting how they could be addressed. The AI model then got a bit pleased with itself, claiming its analysis had revealed ‘clear strategic pathways for the Fito Network’, which ‘appears ready for more advanced content and structured peer-to-peer learning’.   

I think this gives the clearest indication of the way the analytical model has been designed as part researcher, part consultant. It’s classic tech firm strategy – free hook (the research) leading to premium service (consulting) with dollar signs all over it. In this trifling example, the report suggested its findings highlighted an approach which could ‘significantly enhance the network’s long-term value proposition and overall impact’.  

In practice – like a fresh faced grad with a handful of post-its in a room full of experienced leaders – it was merely observing the organisers existing understanding. They’re way ahead, so its conclusions were more validating than insightful. 

The report derived six strategic implications from our chat before going on to make four recommendations.

The ‘strategic implications’ were 

• The content was ‘high value’ and worth exploring more. 
• The curated knowledge hub was a good blueprint (with a caveat about monitoring evolving algorithms)
• Good facilitation with clear process was a critical factor 
• Self sustaining peer-to-peer learning needs structured programmes/channels 
• There’s interest in exploring the role of tech in systems change more deeply 
• Address pain points to reduce onboarding friction 

 

And recommendations for strategic development were:

• Develop Advanced Workshops on complex topics for deep dives into the subjects that resonate most with the membership.
• Create Dedicated Collaboration Channels explicitly designed for collaborative content creation. 
• Showcase Member Contributions regularly highlighting “Cheerleaders & Knowledge Harvesters” content across all Fito Network platforms. 
• Conduct Post-Transition Survey to gather feedback process, preferences and unmet needs. 

 

As I said earlier, these results aren’t particularly valuable, but they do show the great potential of analysing WhatsApp chats.  I’d only asked the simple question ‘what are the trending topics in this chat?’ and got a whole lot more.

So in conclusion, I think selective analysis of WhatsApp chats could help to some degree with issues raised in the consultation paper Bridging in a fragmented world. The thinking there is that knowledge should be seen as ‘a dynamic and relational process’ rather than as discrete parcels of inert knowing. If that’s the case, then the real value comes when simple, insightful, practical clarity is shared with usability in mind.  

So it’s definitely worth using tech to help reveal more of the insights we have when we meet/talk online, and it can help distil messiness into simplicity – but the recommendations above DO apply.  

When knowledge is curated in structured ways that promote collaboration and exchange, we’re moving into the realm of collective intelligence. That’s where the real wizardry lies. 

BUT I can’t end here without flagging that my results came at a cost.  

I asked Gemini one last question, and I think it’s right that I should let it have the last word.

_______

 Me: ‘what would you recommend to balance the environmental cost of using you for searches?’ 

Gemini:

It’s great that you’re thinking about the environmental impact of using AI for searches! It’s true that the computational power and data centers required for AI, including large language models (LLMs) like the one you’re interacting with, consume significant amounts of energy and water.

Here’s what I’d recommend to help balance the environmental cost:

1. Be Mindful of Your Search Habits:

• Be Specific: The more precise your queries, the less processing power might be needed to generate a relevant response. Avoid overly broad or vague questions that could lead to more complex computations.
• Consolidate Queries: If you have multiple related questions, try to phrase them in a way that allows for a single, comprehensive answer rather than a series of individual searches.
• Evaluate Necessity: Before performing a search, especially a complex one, consider if the information is truly essential or if you could find it through simpler, less energy-intensive means (e.g., a quick look at a known website, an existing document).
• Bookmark Frequently Used Sites: Instead of searching for the same websites repeatedly, bookmark them for direct access.

2. Optimize Your Devices and Internet Usage:

• Energy-Efficient Settings: Adjust your device’s power settings to conserve energy (e.g., dimming brightness, using sleep/hibernate modes).
• Unplug Devices: Unplug chargers and devices when not in use to avoid “vampire power” draw.
• Extend Device Lifespan: Keep your electronic devices for as long as possible. The manufacturing and transportation of new electronics have a significant carbon footprint. Repairing and recycling are much more eco-friendly.
• Stream Intentionally: If you stream video or music, consider downloading content you’ll consume multiple times, and stream in lower resolutions when high quality isn’t necessary.
• Manage Emails: Delete unnecessary emails and unsubscribe from newsletters you don’t read. Stored data in data centers consumes energy.
• Green Web Hosting: If you manage a website, consider using a green web host powered by renewable energy.

3. Support Sustainable Tech and Initiatives:

• Choose Eco-Friendly Search Engines: Consider using search engines that actively work to reduce their environmental impact.
  • Ecosia: This search engine uses 100% of its profits for climate action, primarily through planting trees in biodiversity hotspots around the world. They also power their operations with renewable energy.
  • Ekoru: Focuses on ocean restoration, allocating a significant portion of its profits to initiatives like ocean cleanups. It also operates on hydroelectric power.
  • GiveWater: Aims to expand access to clean water globally, with profits supporting water-related initiatives.
• Advocate for Green AI: Support research and development into more energy-efficient AI models and data center technologies. Encourage companies to be transparent about their environmental impact and to prioritize renewable energy sources.

Understanding the Impact:

It’s important to note that the environmental impact of AI is complex and an active area of research. While a single AI query might consume more electricity than a standard web search, the overall impact depends on many factors, including:

• Model Size and Complexity: Larger and more complex AI models generally require more energy for training and inference.
• Data Center Efficiency: The energy efficiency of the data centers hosting the AI models plays a crucial role.
• Energy Mix: The source of electricity powering the data centers (renewable vs. fossil fuels) is a major determinant of carbon emissions.
• Water Consumption: Data centers also require significant amounts of water for cooling.

By adopting more conscious digital habits and supporting sustainable tech initiatives, you can contribute to a more environmentally responsible internet.