Hi, I work in tech and I’m also an author and I feel we need to have a chat. Because I’m seeing a lot of misinformation/conflation happening.

There are many different kinds of AI, and disagreement about what should even be termed “AI” in the industry (it’s the hot new thing, so companies are rebranding all sorts of products they offer as “AI” which is muddying the waters re what is Artificial Intelligence and what’s just a sparkling workflow). That being said, let’s go over some key terminology/how these iterations differ (at least in my part of the data management AI world).

1. Narrow AI: This is a limited, passive, task-specific kind of AI. It’s programmed to respond within certain constraints. Narrow AI is behind the chat bot you talk to when you want to make an Amazon return, or voice assistants you speak to when trying to make an insurance claim. They don’t learn; if you ask them a question outside their purview, they will not provide an answer. Narrow AI can be useful for efficiency and saving costs (it can also be hella annoying if you’re stuck speaking to one that can’t help you with your problem and you don’t know the magic set of words to get to an actual person who can).
2. Agentic AI: This is a more advanced system that attempts to mimic human decision-making within a specific context using LLMs. This is a model that can “learn” and adapt. There are many subsets of Machine Learning within Agentic AI like Supervised learning—where a system can be trained to identify something by certain characteristics (used for cancer detection!) or Unsupervised Learning which is a matching/pattern recognition exercise (like identifying new disease subgroups!). Agentic AI can also help more generally with looking at patient data within the context of their medical history/the most up-to-date medical best practices, and provide insights. Aside from use in the medical industry, Agentic AI might be used within larger corporations for supply chain management—monitoring and automating interactions between suppliers, vendors, freight companies, etc. to make sure the correct number of products are ordered and shipped at the correct time to the correct locations, even if those numbers fluctuate. It can be used for companies who want to enable self-service for their business units to query data and create new data sets based on those queries using natural language (“show me all customers who purchased x product within this time frame in this geographic location,” “now create a new data set with this information”). Like Narrow AI, Agentic AI can improve efficiency, and is helpful in contexts when the breadth of data/moving parts involved is so substantial a human may not be equipped to manage it alone. However, it still needs to be implemented responsibly (more on that after Generative AI).
3. Predictive AI: This uses pattern recognition/machine learning/statistics/algorithms to predict outcomes. Predictive AI taps into an amount of data that was previously not possible for human teams to manage (much like Agentic AI). Predictive AI can anticipate stock market changes, extreme weather, mechanical issues, supply chain impacts, healthcare outcomes, crime surges, and more, saving time, money, and potentially even preventing major problems like vehicle recalls and death due to natural disasters. However, predictive AI is limited by the data it’s trained on, which has resulted in algorithmic biases (like when used in law enforcement/policing contexts, or healthcare contexts). So, as is true for any of the AI models I’ve mentioned so far, while it can be a positive tool, implementors should be cognizant of the fallible, human, foundation it’s built upon and try to mitigate bias.
4. Generative AI: This is what most people think of when they think of AI, thanks to chatbots like ChatGPT. Generative AI is a regurgitative leech. It creates “new” content based on the massive amounts of data it has been trained on. Nothing ChatGPT creates is actually new, though. It’s not thinking for itself. When you ask it to write a story or create a picture, it’s using an amalgamation of the writing and art it has copied from real creators without credit. There is very little useful about Generative AI like ChatGPT. Especially when you consider the environmental ramifications of using it. Generative AI used within a social context (as a therapist/friend/romantic stand-in) is dangerous. And if used as an authoritative search engine (which is disturbingly prevalent, now) it’s equally problematic due to common issues like hallucinations, the spread of fabricated news stories/outlets, dangerous deepfakes, extremist bias, and more. And if you’re thinking, well I just use it to help with outlining papers/re-writing emails/condensing notes, there are already studies that raise concerns about generative AI weakening critical thinking skills. I cannot tell you how relieved I am that I left my job as a professor the year before ChatGPT came out.

Now, to be fair, a lot of Agentic AI depends on LLMs, as I mentioned, which makes it prone to encountering the same sorts of issues with hallucinations/bias as Generative AI chatbots. But in my experience, companies/hospitals/government entities are aware of this and, when implementing responsibly, use Agentic AI as a tool that requires supervision and adjustment, not some holy, infallible, authority (like many public-facing chat bot users). I also think the potential benefits of Agentic AI currently outweigh my concerns about its use of generative AI (though the environmental impacts are still worrisome). So perhaps even further nuance is needed between Generative AI used within Agentic contexts and public-facing Generative AI used purely for entertainment/ “education” in chatbots.

Which only emphasizes my point that lumping all AI together is not beneficial. There’s nuance! Anyway.

TL;DR my personal thoughts on AI:

1. Narrow AI—Can be Good when implemented appropriately!
2. Agentic AI—Can be Good when implemented appropriately!
3. Predictive AI—Can be Good when implemented appropriately!
4. Generative AI (public-facing)—Kill it with fire.

I work in technical sales.

Last week, one of our execs used ChatGPT to create a business value assessment for a prospect (basically, why should this prospective customer choose our tech for their business needs). The data was good—but the language was impersonal, mechanical, and didn’t include the kind of terminology that would resonate with these business folks. The exec’s boss noted these concerns while we were reviewing our two part preso+demo plan (I’m the demo-er) and said “hey, Erica is a writer, get her to look it over and get it adjusted to be more relatable.”

It was quick work for us to retool the language. The director was pleased with the new version, and the presentation was incredibly well-received by the customer. They even asked to get a copy of the BVA to share internally. We’re now in contract negotiations with that prospect.

Our director spent ten minutes reviewing that document on our team meeting this afternoon emphasizing the importance of writing and editing skills and warning the team not to rely entirely on AI because it doesn’t understand customers the same way a human can.

Writing skills are important. Editing skills are important. If you are young, please hear me when I tell you that putting in the work to develop solid communication skills will benefit you in nearly ANY career path.

Me, 2/3 of the way through writing a eulogy: I should take a Tumblr break. I need some whimsy. I need some shitposts.

Me, tears still on my face: immediately confronted with this post

Me: you know what, oddly, this is exactly what I needed. Thank you, Tumblr.