The answer is only partly going to come from understanding AI. And given the amount of hype, we do need an answer.

The usual hype cycle?

My colleagues and I have been working on this since GPT-3 was launched, in mid-2021. Here I focus on the text-based interactions because they’re the bulk of the potential business use, and they’re a good representation of other use cases (text to still image, audio, video, etc.).

Like much past technology, including AI-related ones (e.g., enterprise AI, robotics, self-driving cars), the current limits are clear but they’re not well understood well, resulting in lots of inflated expectations and pointless — instead of useful — experimentation.

What works well, and can already revolutionize the lower end of much creative or knowledge work.

• Long-form text that “feels” right — e.g. copy for ads, and casual chat banter (ChatGPT is the poster child of this).
• Imagery and artwork (two-dimensional). Generative AI has made Adobe acquire Figma, and now that Promethean fire is in the hands of billions of people. Experimentation may lead to uncanny results — such as new workflows for media and the arts, new genres, and generally the democratization of visual creation.

What doesn’t, and needs smart solutions:

• Deepfakes and bias. These models are the representation of the widespread (and typically free) web content and as a result, they’re heavily skewed by anything that’s viral. So, sexism, racism, and outlandish stuff are over-represented. The issue here is that the generative AI content is often good enough to elude filters made to cull bots’ spam. I foresee incidents (e.g., crime, probably hate crimes, and potentially worse) coming out of generative AI. Clearly, companies don’t want that happening on their watch, especially as they have branded interactions with their customers or employees.
• IP protection. This is the Wild West, and companies who want to use generative AI at scale, for profit, need their lawyers involved early.
• Content moderation, given that the volumes are already spiking. Because of the above, we need top-notch human-AI moderation in the media that propagate generative AI content. Regulators are lagging, and many companies struggle with self-restraint policies.
• Some of it is “meh”: more than a bit cheesy, unimaginative, or just weird. Much out-of-the-box output doesn’t do a great job of solving nontrivial problems. And when pushing for real creativity, the stochastic nature of generative AI does show.
• Veracity is lacking. In my view, this is the most important aspect of knowledge work. Accuracy and reliability aren’t always there, and for complex topics is hard to tell apart right from wrong. And there’s no guarantee that this will be fully solved anytime soon. So if I can’t trust it, will we use it? Think of what happens with autonomous driving — something that works 99% of the time but doesn’t tell you when it doesn’t, won’t be used for mission-critical work.

Deep learning in language models generates a sort of intelligent response through the emergent properties of correlations between words at hyper-scale. In other words, it works on semantics but doesn’t understand the world through symbolic, abstract, and conceptual reasoning — unless those are embedded into word correlations. That’s different from how human works.

And that’s where the opportunity is: pairing machines with humans, at scale.

What can we really use this for?

I don’t see how technical AI improvements alone will fix many of the problems above for a while. So, my recommendation is to think of solutions as built on machines plus people, in an end-to-end process flow.

Four avenues seem viable at this point.

Hybrid search + generative AI. Build tools that combine traditional search (which surfaces the right content) with generative AI (which summarizes, identifies gaps, and can be queried) and human feedback (both from experts and laypeople).

Take care of your corpora. Work on curated corpora (semi-unstructured fact and databases) where the data set is constrained, and combine that with smaller or less powerful models (e.g. for OpenAI that would be Ada, Babbage, or possibly Curie, instead of DaVinci; or older BERT versions). That’s been used for good old chatbots (think: IT support) and can scale to better knowledge interaction. The opposite, that is big models with small corpora, may or may not give you real answers and they will do so in such a realistic way that you won’t be able to tell.

Industrialize experimentation. Creators (copywriters, bloggers, designers, and artists) can use the raw materials from the machines to mass-customize their production; or try lots of variants and then curate the best and combine them into that one masterpiece; and even possibly create their own data corpora that reflect their unique style. For these uses, a smaller corpus will be enough to provide the right tone, and the bigger language model will make the narrative more engaging — whereby the risk of talking nonsense is more easily addressable by discarding what does sound like baloney.

Get a digital sparring partner. Generally — use them to prompt humans to think, by throwing at them different perspectives, like a normal coach, or a somewhat absent-minded assistant would do — instead of hoping that machines can substitute our boss, or be a synthetic, mistake-proof “slave”.

We do need some boring yet reliable and useful output here. Forget about the magic, and think about this as an engine in larger operational, scalable processes. In other words, leverage AI-augmented collective intelligence, not just artificial intelligence.

This post complements the tech-driven organizational design materials at Supermind.Design and some previous blog posts on designing an AI-augmented collective intelligence. Read them on Medium if you’re interested in using these techniques in your own organization, and get in touch if you want to discuss.