Information Warfare: DeepSeek & The Shady World Of Influencer Marketing
Businesses are building information libraries, and that has many implications. The more complete the information library’s picture of the marketplace is, the better the company can compete. One of the more interesting angles this opens up is businesses competing with each other using information asymmetry and disinformation. It's a framework called Decision Dominance. It originates in the defense domain but has profound implications for businesses.
Decision Dominance aims to get your opponent to react to an inaccurate picture (based on their information library) of the marketplace. Decision Dominance approaches the problem in two ways. First, slow down your competition’s decision-making or information gathering so it's impossible for them to respond to changing conditions before they change again.
In my ‘The Agents Are Coming’ series, I teach systems thinking because businesses must treat information as an asset. AI agent platforms must be designed to optimize data flows, data-to-information transforms, and information deployment. It’s expensive and often infeasible to take the first approach against a competitor with high information maturity.
The second approach is releasing information that paints an inaccurate or deceptive picture of the competitive landscape. It’s a data poisoning attack at scale, but the targets are people as well as models. The more time I spend listening to the DeepSeek saga, the more I wonder if we are seeing Decision Dominance deployed as a competitive tool.
In this post, I’ll explain what DeepSeek did, how it probably did it, and how AI platforms play a growing role in marketing campaigns. I’ll wrap up with a breakdown of the 5 levels of an AI agentic platform and connect the dots to the future of influence and perception-driven marketing.
I Kicked The Beehive
I started some controversy by saying, “Deep Seek is a cheap knockoff of an AI pet rock.” Many of DeepSeek’s claims about the cost of model training are deeply misleading. Still, we saw stock markets move dramatically last month. Some people went as far as to question the long-term viability of AI.
It started with a false equivalence. People spreading the DeepSeek message compared the total spending of companies like Microsoft or OpenAI to the $6 million spent training DeepSeek's latest model. If you know the basics of how OpenAI’s o1 model is trained, it’s much easier to do it a second time and optimize the training process. DeepSeek had the benefit of optimizing prior work.
There are no guarantees if you’re working on the leading edge of AI research, so AI labs train several models at a time. OpenAI was reminded of that lesson the hard way with Orion. It put all its GPU eggs into a single model training run and didn’t have much to show for it at the end. That’s why frontier labs need so much compute. Model training starts with thousands of very small model training test runs.
As the model’s size scales up, multiple flavors, sometimes even multiple versions, train simultaneously. AI labs learn and improve on the fly in later training phases with little guidance from prior work.
Getting the DeepSeek LLM we saw released took at least 5X the compute they advertised because the startup was training more than just the winning version. It also leveraged other frontier LLMs and smaller models to train DeepSeek. Were those compute costs factored into the total?
The startup failed to disclose additional costs beyond compute resources like data curation and talent. In a defining irony moment, OpenAI accused DeepSeek of stealing data from ChatGPT. Is it theft if you steal something that someone else stole first? A lawyer friend said there’s an argument to be made that since DeepSeek is open-source, it had technically returned the stolen property to the owners.
What about the optimizations Deep Seek achieved? The combination of multiple existing approaches and some optimizations to how the GPUs transfer data led to improvements. That’s undeniable, but the impacts and implications have been completely blown out of proportion by the influencer machine.
The hype cycle they started led to a tech stock correction, so it’s impossible to fully cover this story without wading into the backlash against my post. I kicked the beehive. The bots and influencer crowd didn’t like it. Their amplification campaign ran into a disruption wall, and they didn’t have a response. When NVIDIA reports its quarterly earnings on Wednesday, you’ll see the influencer cycle kick into high gear again.
Influencers. A Highly Questionable Field
The surest sign of a questionable industry is when people in it don’t want to be associated with their titles. I am an influencer and thought leader. Those terms used to mean something very different than they do today. Influencers, futurists, and thought leaders were well respected until recently.
Thought leadership is the most effective way to establish a business as an industry leader or innovator. It establishes an individual’s expertise and vision in their field.
Futurists help businesses transform by showing how technology breaks assumptions that used to be rules and the implications of technical disruptions. They unlock creativity and imagination.
Shady behaviors and bad actors crept in through social media. It enabled anyone to establish credibility by assembling a large community. Followers and content engagement are social credit, much like Trustpilot ratings or 5-star reviews on Amazon. The data story collides with people willing to put aside their ethics to make a quick buck.
In 2015, I had a front-row seat to Goodhart’s Law in action.
“When a feature of the economy is picked as an indicator of the economy, then it inexorably ceases to function as that indicator because people start to game it.”
Spilling The Tea About The Influencer Industry: Fake Influence Peddlers
The primary indicators of the influencer economy are followers, views, reshares, and likes. Today, going viral on X can lead to VC funding or small businesses becoming household names overnight. ChatGPT proved the viral business model scales. Influencers can make products go viral by posting about them on social media or writing articles that feature them. Awareness and engagement lead to interest.
When I founded V Squared in 2012, our first thesis was built on influencing customers’ perceptions of a product’s value. What’s the difference between a Louis Vuitton bag and one from Coach? Customers perceive one as having higher value, so the company has greater pricing power. The stock market operates in much the same way, with forward-looking perceptions of growth potential driving share prices.
If your influence campaign successfully convinces investors that AI’s future prospects are dimming, you will cause their perceptions of value for NVIDIA, Microsoft, and Google to diminish. Their patience for AI returns will plummet. Buying and trading behaviors will follow their new perception, and every data point will be viewed more negatively. Reach enough investors or the right investors, and share prices will fall, reinforcing the accuracy of their perception.
Influence is a real industry and the most effective form of marketing, leveraging the same decision mechanics as recommendations and referrals do. Trust changes perceptions, and behavior follows. Influencers get paid A LOT of money because their campaign ROI can be massive. With money pouring in, many looked for a shortcut to the front of the influencer marketing line.
A cottage industry sprung up, selling fake followers, views, reshares, and likes. The whole system was too easily gamed by bots and engagement farms. It became a cat-and-mouse game between social media companies and engagement farms. Social media companies added requirements for distinct phone numbers to be associated with accounts, and the farms bought cheap smartphones in bulk.
Most social media companies have settled into an informal relationship with bot farms and fake influence peddlers. They drive user engagement, so social media companies look the other way as long as they aren’t too obvious. Bots have become a critical part of social media platform ecosystems. Some algorithms even enable bots, which I’ll get to in a couple of sections.
I came to understand these influence brokers in 2015 while working with a think tank and other influencers to study the bot ecosystem. We set up fake accounts on social media sites and bought followers from multiple brokers. We tracked who the fake followers followed and engaged with most.
I learned how to manage complex graphs and model communities working on the project for 2 years. It coincided with the 2016 US election, and the bot farms feeding social media influencers had strong ties to political and sovereign influence campaigns. The US, China, Russia, Israel, Saudi Arabia, and the UK were all major participants in the engagement marketplace. All used fake influence peddlers.
Today, most countries have some social media influence apparatus. Social media influencers often have connections to political and sovereign influence campaigns. I’m sure some of them know. However, after meeting and getting to know dozens of influencers, I don’t believe many realize who’s behind their largest paychecks. The tricks from sovereign influence campaigns have filtered into corporate influencer marketing playbooks.
Building Business Models By Gaming Metrics
Engagement and click farms make their livings by exploiting flawed assumptions.
If thousands like a person or their content, it must be high-quality, and they must be an expert.
Sales will rise if many people see and engage with a company’s ads (influencer marketing campaigns).
Goodhart’s Law survives on flawed assumptions and proxy metrics.
It’s amazing how effective a fake follower and engagement strategy is. Social media consulting companies used a formula to turn founders and executives into household names.
Step 1: Buy between 25K and 250K followers.
Step 2: Build a content calendar around popular themes and events. The content is both informative and provides perspective on the theme, trend, or event.
Step 3: Buy 1K likes and 200 reshares for every post. Use 10 internal accounts to write comments and get others to discuss the content.
In about 3 months, 25K fake followers became a 100K+ community, and 250K fake followers crossed the million milestone. Content engagement was increasingly organic, and social media strategists backed off the paid engagement. The recipe is inexpensive and delivers rapid results. Once executed, there is no easy way to trace the account’s growth back to its roots.
Salespeople and marketers quickly caught on to the game. They followed the playbook to transform themselves into industry experts and thought leaders. From the beginning, their goal was to sell their influence to clients, but they never actually built any quantifiable influence within their communities. Goodhart’s Law launched into full swing.
Social media marketers who pay influencers used to discover them from Top 10 Experts and Top 100 Most Influential lists. Those lists scored accounts based on followers and engagement. Between 2014 and 2019, most lists didn’t have checks for spam, bots, and fake followers. Influencers figured out that if they kept paying for engagement, they made these lists, and clients would pay them 6-figures a year for content and social media marketing campaigns.
What works for people works equally well for a new category of AI agent…the social media influencer bot. We all follow accounts that are mostly AI-generated content and don’t know it. If a simple playbook and bots can amplify people and rapidly develop their reach, they can do the same for an AI-run account.
AI To The Rescue, But Not Our Rescue
Once objectives are defined (the perception you want to change or groups you want to activate), the first step in an influence campaign is identifying influencers who hold sway over the campaign’s target groups. Then comes the hard part: messaging.
Models and natural language analysis significantly improve audience composition analysis, influencer targeting, and message engineering. Cambridge Analytica and Meta showed the power of user data and models for targeted influence campaigns. Models support influence campaigns by: