Paperclip Maximizing: Machine Learning And The Problem Of Instrumental Convergence
What do paperclips have to do with digital marketing and machine learning? Admittedly, pretty much nothing. The ‘paperclip maximizer’ thought experiment comes from Nick Bostrom at Oxford University. In essence, it looks at the idea that if you tell a machine to optimize to a specific goal it will do so at all costs.
If you told a machine to maximize the number of paperclips it produces the machine would eventually start destroying things like computers, refrigerators, or really anything made of metal to make more paper clips once other sources of metal run out. This concept has been coined as instrumental convergence.
Paperclip vs Pay-Per-Click
If you transform the idea of paperclips into the idea of paying per click this becomes very relevant to digital marketing. Nowadays, almost every platform touts some version of AI or machine learning to revolutionize campaign performance, which is a boon to everyone.
By releasing control to machines, media buyers can focus on more strategic tasks such as identifying deeper insights for reports and understanding clients’ goals while campaigns continuously improve themselves. They do this by finding and optimizing for what works while avoiding the things that are not driving performance for the brand.
See this in action on Our Work page.
What exactly is ‘performance?’ The easy answer is whatever your KPI may be. It could be clicks, it could be video views, it could be any trackable metric. However, this is already a simplified goal. If you are running a traffic campaign, the marketing goal should not just be to “get more clicks”.
The goal should be something designed to move the needle for the business – brand affinity growth, sales lift, etc. For instance, driving qualified users from a target audience to an advertiser’s website and increasing brand favorability is a strong goal. This is something bigger than a single metric. A KPI can be a stepping stone and an important indicator of success, but it is not the final objective.
A truly successful campaign will not simply be the campaign that drove the most clicks at the cheapest price point. Success lies in the campaigns that drive true performance towards core business objectives.
Looking for a partner to harness AI technology and drive marketing performance? Reach out to Coegi for a discovery call today.
What Machines Lack
Context. Context is something that a robot has not yet mastered.
As a marketer, I know that increasing clicks, directionally, should push me closer to my goal. The machine knows this too, but this is all the machine knows. It will endlessly optimize to a single goal.
Maximizing the number of clicks given a fixed amount of budget. This can lead to unintended consequences – think back to the paperclip example.
A machine might say only run display banners and forget about high impact formats such as video and CTV. The machine might push 100% of impressions into in-app environments. It would say never buy another out of home ad again. The machine would never know to build brand awareness, because there is no optimization point it can use.
Machines Need Guidance
We understand the role of the media buyer is not going away, but it is morphing. There is a new symbiotic relationship between buyer and machine which empowers them to maximize your brand as a whole. Successful media buyers do not need to spend 80 hours per week finding every winning media combination.
Most of these tasks can be done through harnessing technology and freeing up time to look at the media plan from a higher level. These benefits of using AI go directly to our clients in the form of more time dedicated to listening to client needs, smarter digital media plans, and ultimately higher performing campaigns.
How To Incorporate Smart AI In Your Media Buying
- Release media optimization controls to AI machines and spend time on strategic campaign elements.
- Define performance success beyond the metrics by establishing meaningful campaign goals
- Use context to avoid instrumental convergence and potentially harmful optimizations from unattended machines.