0x18: How Not to Suck at AI

People, at least for now, need to be an important part of the AI adoption strategy. The MIT Project NANDA report dips into this and I am going to bring some anecdotal personal experience of what I believe is and is not working.

A Note From Me!

I’ve been so busy learning and playing that I haven’t been writing/recording much. This essay was drafted back in September and recorded in December, but never edited. I felt it was getting stale, so I’m publishing it as is to focus on something cool. Stay tuned!

Now back to the essay.

Many AI curious companies’ adoption strategy has been to simply to turn the AI on, give employees access, and hope they come back with some magical, cost saving and (potentially) job eliminating innovations. While this strategy is likely better than resisting AI, only to find out employees are using their personal AI accounts (AKA ‘shadow AI’), it isn’t much of a strategy.

Because many employees either won’t use AI at all or will stick to the simplest features that make their own tasks a bit easier, essentially, looking at AI as their productivity cheat code. The results: polished emails at best and AI work slop with grave consequences at worst... Let me explain.

I recently coached a friend in the medical community to try ChatGPT. I warned about the privacy concerns and hallucination and sycophancy issues, as well as what I like to call the “beautiful garbage” problem, now coined “workslop”. After spinning up ChatGPT (free), they certainly had that “Ah-ha” moment we all do, both from a productivity perspective but more so realizing they were the last ones in their office to start using ChatGPT! However, a week later the output was peer reviewed by a doctor and the “perfect” AI generated process was found to be full of inaccuracies and contradictions that, and I quote, “could have resulted in patient harm”.

AND this is what scares me about letting the AI cat out of the bag to the general public. Around the world, every hour, tens, if not hundreds, of thousands of people are using AI as their cheat code, producing beautiful garbage, as their own, not reading it and flooding our world with noise at best and poisoning our collective knowledge at worst.

The integrity of our collective knowledge, essentially what we have created over the past few millennia, is at stake with lazy people using AI to generate content as their own, inadvertently polluting our knowledge and consequentially reality.

This is killing the Internet as we know it today, it poisons our academic knowledge and most scary contaminates our economy’s inputs and outputs (business data) with believable inaccuracies at best, but much more dire, bad data that has and will quietly continue to cause actual harm. I have the receipts.

The next AI adoption strategy is similar to the above, but bolts on some optional training or lunch’n’learns to get all who are interested, going. This sounds like a step in the right direction, but I’d argue it has the potential to be worse.

The previous adoption strategy is the proverbial giving a baby a gun, and generally results in polished emails and not much more. However, this YOLO AI+ strategy, more akin to giving a monkey a gun, provides just enough knowledge to be dangerous, but fails to invest in enough training to enable safe productivity at the individual level and little to no ascending benefit to the enterprise.

What I have observed with this adoption strategy is a broad demographic of the AI curious, bringing their varying technical skillsets, knowledge and experiences to exploit the capabilities of AI, but mostly for their own tasks. The upside is the creation of an internal grassroots community that is experimenting, sharing knowledge and trying to find success. Which is great. The downside is that for every employee who starts creating with AI, the company gains a new, untested permutation of one or more business processes, leaving the business no better, and likely much worse from a sustainability, compliance and continuity perspective, at a minimum.

And let’s not forget the privacy and confidentiality concerns. Enterprises have spent millions to protect their data and comply with privacy laws. However, the introduction to AI, especially beyond the simple chatbots, is a new exposure that companies are not prepared for and this is resulting in several situations where AI enablement has had negative consequences.

While Generative AI has given the masses the ability to use natural language to output some beautiful results, a few Youtube videos and lunch’n’learn sessions does not provide the adequate knowledge or experience to really move the productivity needle forward, beyond polished emails and good looking but highly suss TPS reports.

In my opinion, the majority of users need to be the consumers of AI solutions, not the creators. Companies need to recognize that beyond the natural language processing and parlor tricks we have all seen with generative AI… Knowledge, skillset and experience are needed to truly move the needle forward. I believe in order for companies to see the net benefits of this new technology, at the enterprise-level, they need to centralize AI innovation through a crack team or skunkworks, rather than letting the AI cat out of the bag and... finding out.

A crack team can be in‑house or a trusted partner. They temporarily embed within a business unit, run short discovery sessions to map how work gets done, the pain points, and then they define one or more problem statements and use their expertise to prototype and stress test a solution.

An AI crack team isn’t going to solve a company’s AI adoption challenges overnight, but they can solve “a problem” and support further maturity of the adoption of AI and most importantly the successes.