Subscribe to DSC Newsletter

Which jobs will AI (Artificial Intelligence) kill?

AI was very popular 30 years ago, then disappeared, and is now making a big come back because of  new robotic technologies: driver-less cars, automated diagnostic, IoT (including vacuum cleaning and other household robots), automated companies with zero employee, soldier robots, and much more.

Will AI replace data scientists? I think so, though data scientists will be initially replaced by "low intelligence" yet extremely stable and robust systems. There has been a lot of discussions about the automated statistician. I am myself developing data science techniques such as  Jackknife regression  that are simple, robust, suitable for black-box, machine-to-machine communications or other automated use, and easy to understand and pilot by the layman, just like a Google driver-less car can be "driven" by an 8 years old kid. 

My approach to automating data science and data cleaning / EDA (exploratory data analysis) is not really AI: it's just a starting point, but not a permanent solution. In the long term, it is possible that AI will handle complex regression models, far more complex than my Jackknife regression: after all, all the steps of linear or logistics regression modeling, currently handled by human beings spending several days or weeks on the problem, involve extremely repetitive, boring, predictable tasks, and thus it is a good candidate for an AI implementation entirely managed by robots.  

As machine learning (ML) more and more involves AI, and the blending of ML and AI is referred to as deep learning, I can see data science evolving to deep data science (DDS) or automated data science (ADS), where AI, robots, or automation at large, take a more prominent role. 

True AI systems can even predict travel time in real time based on expected traffic bottlenecks and road closures

Which jobs are threatened by AI?

Just like data science will take years to get a high level automation, where as much as 50% of human tasks are replaced by robots, I believe that these professions are at risk, but the erosion will be modest and slow, taking a lot of time to materialize:

  • Teachers: some topics such as mathematics or computer science can be taught by robots, at least for the 10% of students that are self-learners. Generally speaking, topics that are currently taught by robots include flying a plane, training on an AI-powered simulator. Ironically, planes can be flew without human pilots, but studies have shown that passengers would be very scared to board a pilot-less plane. The biggest threat for teachers is not AI though, it is online training.
  • Grading student papers, detect plagiarism. But students / authors are getting more sophisticated, using article-generating software powered by AI, to avoid detection. This could lead to an interesting war: AI robots designed for fraud detection fighting against AI robots designed to cheat.
  • For publishers, automatically writing high-quality, curated articles in a short amount of time. An article such as this one is a good candidate for automated, AI-powered production. The first step is to identify articles that are good candidates (for curation)  for a specific audience; this is also accomplished using AI. 
  • Can AI writes AI algorithms, or in short, can AI automate AI? I believe so; after all, I was one of the pioneers who wrote programs that write programs (software code compilers or interpreters also fit in this category). I guess this is just an extension of this concept.
  • Automated diagnostic (or automated doctor, but also automated lawyer). I guess this will eliminate a small proportion of these practitioners. But what about a robot performing a brain surgery with higher efficiency than a human surgeon? Or a robot manufacturing an ad-hoc, customized client-specific drug for maximum efficiency? 
  • Automated chefs replacing expensive cooks in a number of restaurants. Or think about a McDonald restaurant where the only human is a security guard - everything else being outsourced to AI-powered robots, including cleaning, preparing food, delivering to customers, processing payments, filing tax returns and accounting, ordering from vendors, and so forth. This would require significant system-to-system communications, but I believe it is feasible.
  • Automated policemen or soldiers is a source of concern, as you would have algorithms that decide who to kill or who to arrest. So this might not happen for a long time, though drones are replacing soldiers in a number of wars, and have the power to kill (based on some algorithm) with no one complaining about, as long as it is not happening in US. Terrorists might be attracted too by this type of technology.
  • AI will be present in many IoT applications such as smart cities, precision farming, transportation, monitoring (detecting when an offshore oil platform is going to collapse), and so on.

AI and automation has already replaced many data science tasks long ago

Many people talk about the threat of AI, but as of today, many jobs have already been automated, some more than 30 years ago. For instance, during my PhD years, a lot of data transited through tapes between big computer systems, and involved trips to the computer center, interacting with a number of people taking care of the data flow. This has entirely disappeared.

We used to have shared secretaries to write research papers (they could write LaTeX documents), I think this has all but disappeared.

One of the applications that I developed in the eighties was a remote sensing software that could perform image segmentation and clustering, for instance to compute the proportion of various crops in a specific area based on satellite images, without human interactions - thus eliminating all the expensive jobs that were previously performed by humans to accomplish this task.

Final note

Those who automate data science are still data scientists. Just like those developing robots to automate brain surgery work in a team, with many members being brain surgeons. it's just shifting the nature of the job rather than eliminating it.

DSC Resources

Additional Reading

Follow us on Twitter: @DataScienceCtrl | @AnalyticBridge

Views: 6068


You need to be a member of Data Science Central to add comments!

Join Data Science Central

Comment by Anurooop Iyengar on November 15, 2016 at 9:20pm

While I agree to elements on this list, I think there is a more serious threat posed by Cognitive RPA (Which is essentially Data Science + Deep Learning). I will add : -

1. Call centre agents

2. KPO employees

(It is already happening - a MNC company laid of an entire group of paper processing folks)

Disclosure: My company is engaging the market for doing this to Coaches & L&D departments in enterprises

Comment by Kimberly Dee on February 18, 2016 at 5:19pm
While I agree that at higher levels of education ie some MS, PhD programs, AI may replace teachers, I seriously doubt this will happen at the grade school level. I'm no child psychologist (I do have teaching experience from grade school to college), however I believe human interactions with the developing mind is as critical to learning as is the quality of the material itself. A pat on the back for a job well done just won't have the same affect coming from a computer/robot.
Comment by Gary Baggett on February 18, 2016 at 1:40pm

I agree with Jeb Stone.  I think that basically what he is trying to express is that AI isn't doing "creative problem solving" at this stage even though the process of building software from the bottom tasks up to the higher levels of processing are continuing.  This may eventually come to be but not in my lifetime (btw: I'm 60+).  The AI will make automation easier but until the AI becomes a truely creative thinking machine, humans will still have jobs in the problem solving arena.

Comment by Vincent Granville on February 18, 2016 at 10:10am

I think the idea is not so much automating the selection of an ML algorithm (it can be automated if you use some yield metrics to compare performance), but rather designing extremely stable algorithms that won't over-fit even if used by a layman, and yielding nearly as much accuracy and predictive power as more traditional algorithms. And maybe faster algorithms. Then blending multiple algorithms is another promising area.

Comment by Dermot Cochran on February 17, 2016 at 10:24pm

The process of selecting the best machine learning algorithm for a specific data set, is still very much a manual process, and somewhat subjective at present.

Is anyone in working on the problem of how to automatically select the best machine learning algorithm for an unknown dataset? (not just self-tuning of an algorithm or feature learning, but automated creation, composition and selection of the machine learning algorithms).

As a starting point, can we decompose existing algorithms into standard building blocks?

Comment by Jeb Stone on February 17, 2016 at 1:57pm

While you're correct that many types of models "involve extremely repetitive, boring, predictable tasks," this is only true after an experienced human has identified the necessary data assets, sourced them, and defined an implementation appropriate for the desired outcome. It's often the case that important data aren't yet captured, dependent variables aren't defined, clever experimental designs are required, a specific implementation is required to mesh with internal technology, business problems are tricky, etc. etc. It's not like we run regression models just to look at the fit statistics; these models are custom-made parts that plug into other things.

To the extent that data scientists aren't just niche developers, you have to be pretty good at knowing what you need that's not yet defined. By the time AI can figure out on its own what data it needs to solve a problem and how to get it, aren't we talking about the end of *all* jobs? Once AI can do that, humans can no longer lie about observable events and get away with it.

I probably couldn't count how many high-dollar consultants I've had say "Let's try clustering your dataset" when clustering doesn't solve any current business problems; I might expect the same from my robotic replacement. It'll be interesting to see what classes of problems become significantly more automated, though I expect fewer changes over the next 10 years than perhaps you do.

Follow Us


  • Add Videos
  • View All


© 2017   Data Science Central   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service