Do you use generative AI?

Within limits and with strict guardrails, we use generative AI in our work, both internally and in our shared work with you. Traditional tools multiply effort, but more or less deterministically: Excel always produces the same result given the same numbers and scenarios, for example. LLMs, on the other hand, produce whatever they produce mysteriously, and the outputs may be similar, substantially different, orthogonal to the question at hand, wrong, or brilliant; and they produce output in many formats. So, LLMs should give us pause because of their:

• over-confidence with facts that are wrong or out of date.

• occasional hallucination of answers out of whole cloth rather than saying "I don't know that. Would you like me to look?"

• refusal or inability, until recently and only within certain models within certain LLMs, to share citations for the prior work on which they base their output.

• use of em dashes. We're kidding, but not entirely. At 22 Paths we've all spent years in academia and in technical positions that require clear expository writing. A common and effective piece of punctuation to set off important clarifying information is an em dash. It's more forceful than a comma, less formal than a semi-colon, and easily scanned. But now the cat is out of the bag--have a look at your LinkedIn feed--and it's easier to rephrase a sentence or two than to give our readers pause. Except for the em dashes that I've been using, and will continue to use here. But it's easy enough to tell whether a block of text has been generated by an LLM, even if we give it explicit instructions to avoid the passive voice; avoid stilted constructions or flowery language; be concise; write for an 8th grade audience; and so on.

All of which is fine: there's nothing inherently wrong with AI-generated text; however, LLMs are also sensitive to the risks posed by the "stakes continuum" of information. Some information is low stakes--a user story, for example, that's going to be discussed and revised--while other information is high stakes--the analysis of an ambient recording of an encounter between a patient and practitioner, its summary, the set of facts and codes the LLM (or ancillary AI processes) extracts from it. These must be complete and correct; and while it's possible to use techniques like RAG to minimize hallucinations, and agents to allow human-in-the-loop review of critical information, Gen AI accelerates the process of transcription and fact extraction but does not absolve its users of their responsibility to be experts in their domains. This we take seriously: we're still responsible for the soundness of the counsel we give and the artifacts we produce. We'll add follow-up pieces to explain the limits under which we use generative AI to produce content, and what guardrails we keep in place to ensure that, while accelerating our processes, we don't careen into space.