Dear Vsevolod, you have been working in the field of Natural Language Processing (NLP) for a good 15 years. What has changed with the new LLM models?

Almost everything! Five years ago, it was almost impossible to identify sarcasm or irony in texts. Now, however, the models are getting better with each version.

Could you explain that in more detail?

Sure. Recently, the linguistic models – I’m thinking in particular of Gemini and the last versions of ChatGPT – have been able to analyze content that combines different media or data types as a whole. For example, the text and mood conveyed by the associated image in a meme. Music and sounds can also be included. This is very important information for analyzing and comparing different formats. 

That’s interesting because multimodal formats work so well because they appeal to different senses at the same time and can thus reach more users.

Exactly. I think it’s exciting to see the options that this development opens up for content analysis: We can develop a deeper understanding of user behavior and brand perception if we can understand posts in their entirety as soon as they are captured by machines. It is also interesting to see whether the humor of a post is reflected in the comments, i.e. the fundamental question of whether the joke was understood by the audience and met with a response. And, in general, whether or not the tonality of the sender and the audience is similar. 

What does SemanticForce use this kind of know-how for?

For example, when we help our customers select the right influencers. In certain industries, marketing without influencers is no longer conceivable. At the same time, choosing a brand ambassador is a strategic decision, and it would be quite irresponsible to base it only on superficial KPIs such as reach, engagement rate or conversion. 

How do you proceed?

First, we create a long list of candidates. In addition to the criteria mentioned above, we take other aspects into account. Does the quality and aesthetics fit the client’s brand guidelines? Does the audience fit the brand? Is the product mix coherent or too broad? How high is the proportion of bots among the followers? 
After the first customer feedback, we examine the candidates on the short list for possible reputation risks that could reflect badly on the client. Are there any bought followers? Or political risks, directly with the person or in their environment? Are there any pending legal proceedings? These people want to advertise a product or service with their personality. That is precisely why their authenticity must be clearly established in advance. Today this is much easier than it used to be. 

How do you recognize bots?

That’s relatively easy. How often do they post? If they make several thousand posts a week or more than 50 a day, it’s unlikely that they’re real people. Simultaneous activity in many channels is equally conspicuous. A recurring, high number of likes is rather suspicious if you only make a few posts yourself. Many bots use language that is too simple and don’t stand out in terms of creativity.

On a scale of 1 to 10: How high do you think the risk from bots is?

Unfortunately, it’s generally high, even if the risks vary greatly depending on the application. In consumer goods marketing, we will see many AI influencers. 
I would rate the risk in this area as a 4: We will see more quantity and – probably – less quality. 
In politics, I would rate the danger posed by bots as a 9, and it is a real concern for me. Bots are completely unregulated and are undermining our democracies. So far, I have mainly observed that the problem is completely underestimated by politicians.

So you are rather skeptical about the development that AI has triggered?

I wouldn’t say that. But we still have a lot to learn about how to use the technology properly. I also see a lot of opportunities, especially for niche players in the so-called “long tail”: it has never been so easy to find people online who are passionate about the same topic. And companies are now much better off at listening and finding out what motivates their customers.